Links

No Signs of AI Payoff Outside the Tech Sector

Revenue and profit margins remain flat for the S&P 493 despite massive AI investments in corporate America. Details how mainstream firms view token spending as a cost center and lists specific human factors AI cannot replicate. cited

Discussed onHacker News 530 pts · 543 comments, Techmeme

Zagitta/gpu-unlock: Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents

A Linux kernel module doubles FP/BF16 instruction rates on Nvidia 3000 series cards to match professional equivalents. Includes a table showing 3090 Ti gains from 87.46 TFLOPS to 118.39 TFLOPS for FP16 and 87.50 TFLOPS to 157.41 TFLOPS for BF16. cited

Reported byAlphaSignal, Latent Space / AINews

Discussed onHacker News 221 pts · 70 comments

How 2 sources put it

• AlphaSignal2026-10-24issue ↗ removes the FP16/BF16 tensor throttle on Ampere GeForce cards via a Linux kernel module.
• Latent Space / AINews2026-10-24issue ↗ … 41→79 TFLOPS—and the Reddit post links implementations for Ubuntu/Linux and Windows forks (main, fork). The practical significance is potentially higher AI/ML throughput on consumer Ampere cards, but the claim appears to rely on unofficial kernel/driver patching and may not translate linearly to end-to-end workloads. Commenters were interested but…

Comfy-Org/workflow_templates: ComfyUI template workflows

The repo hosts official ComfyUI workflow templates and subgraph blueprints for Python distribution. Includes a searchable site with 11 languages, deployment pipelines, and a step-by-step guide to adding the Wan 2.1 video template. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-23issue ↗ Ready-made ComfyUI workflow templates are provided for every task variant.

repackages FLUX.1-dev into single-file safetensors for ComfyUI, with text encoders bundled in.

Reported byAlphaSignal

cranebeaker/coop-anything: Co-op Anything: point your AI agent at any single-player game and play it with friends (online and/or splitscreen!). MCP server and playbook to let you reverse engineer, fix, test, ship and share what it learns.

This repo lets AI agents reverse engineer single-player games to add online co-op and splitscreen support. Includes ready-made tools for testing sync, a directory of 230+ existing mods, and setup guides for Claude Code and other agents. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-25issue ↗ has passed 1.2k stars and tracks 230+ existing co-op mods across engines.

blackstone bet on chips drove a debt boom that s upending credit

Discussed onTechmeme

Baba nabs $40M series A, backed by FirstMark Capital and CVS Health Ventures, for AI-powered Medicare patient navigation

Baba raised $40 million in Series A funding to modernize Medicare and Medicaid experiences. Includes a benchmark measuring AI agents on patient advocacy tasks and plans to scale value-based arrangements. cited

Discussed onTechmeme

capitalos 21 million spend management

Discussed onTechmeme

Instagram Launches Apple TV App With Reels and More

Instagram launches an Apple TV app with Reels, longer videos, and live shows on October 23. Includes tvOS 26.0 requirements and details the rollout from Fire TV to Samsung smart TVs. cited

Discussed onTechmeme

moonshots kimi k3 1 pro flirts with open weights leadership amid paid influencer scandal allegations

Discussed onTechmeme

AboutKimi-K3

london stock exchange group agrees to buy octus from permira for 3 5 billion 69361aa5

Discussed onTechmeme

Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence

Moonshot releases Kimi K3.1 Pro, a 1-million-token context model with Highspeed at 2-3x speed. Details Terminal-Bench 4.0 scores of 65.7% and xdev-bench-hard vulnerability rates of 81.0%. cited

Reported byAlphaSignal, Simon Willison, Latent Space / AINews

Discussed onHacker News 1,064 pts · 655 comments, Techmeme

AboutKimi-K3

How 3 sources put it

• AlphaSignal2026-10-23issue ↗ Moonshot released Kimi K3.1 Pro, a 2.8T total / 104B active MoE with 1M-token context.
• Simon Willison2026-10-23issue ↗ Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million…
• Latent Space / AINews2026-10-24issue ↗ You can tell Kimi is proud of their latest model by how fancy their landing page is:

meta unseats musk as tech villain for lula in brazil s election

Discussed onTechmeme

AI and the Fall? of the Creative Class

US creative industries lost over 200k jobs in four years, with movie and sound recording sectors bleeding nearly 100k positions. Compares current job losses to the music industry's streaming transition and details how AI-generated content now competes with traditional video revenue. cited

Discussed onHacker News 298 pts · 305 comments

packages FLUX.2 [dev] weights for drop-in ComfyUI use, now past 1M downloads.

Reported byAlphaSignal

Diffusion transformer flux2_dev_fp8mixed.safetensors 35.5 GB file

Reported byAlphaSignal

18 GB FP8 file

“BF16, FP8-mixed, and FP4-mixed variants; the 18 GB FP8 file offers a practical middle ground” via AlphaSignal

Reported byAlphaSignal

His latest venture, Idealab.ai, applies decades of experience to finding new business opportunities – and makes those…

“Bill Gross founded Idealab, which has launched more than 175 companies, and pioneered paid search through GoTo/Overture. His latest venture, Idealab.ai, applies decades of experience to finding new business opportunities – and makes those ideas freely available.” via Exponential View

Reported byExponential View

anthropics margin openais opportunity

Discussed onTechmeme

Push-Based KV Transfer to Improve TTFT and TPOT for Disaggregated Inference in vLLM

vLLM adds a push-based NIXL connector that moves KV transfer setup off the critical path to improve TTFT and TPOT. Includes benchmarks on Qwen3-32B showing up to 3x better TTFT and 30% better TPOT compared to pull mode at high request rates. cited

Reported byTLDR AI, AlphaSignal

How 2 sources put it

• TLDR AI2026-10-15issue ↗ This post introduces a push-based NIXL connector that moves post-prefill transfer coordination and setup off the critical path, resulting in improved application responsiveness, particularly at high request rates.
• AlphaSignal2026-10-24issue ↗ KV transfer: A NIXL connector raised transfer throughput from 141K to 440K tokens/s/GPU.

deepseek-ai/DeepInfer: DeepInfer: High-performance, large-scale inference framework for DeepSeek models

DeepSeek releases DeepInfer, an open-source inference framework supporting DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Includes configuration for DSpark verification, SWA bounded replay, and batch-invariant inference, plus build instructions requiring Python 3.12 and CUDA 13. cited

Reported byInterconnects, Latent Space / AINews, The Rundown AI, AlphaSignal

Discussed onHacker News 2,347 pts · 891 comments

How 4 sources put it

• Interconnects2026-10-20issue ↗ …ws up by releasing their internal inference stack publicly, DeepInfer, which is likely far more efficient than existing open-source tools like vLLM/SGLang for their models. Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack. It is well-known that the top AI labs have much more inference-efficient tools…
• Latent Space / AINews2026-10-21issue ↗ …house inference framework, is now fully open source: image, GitHub. The technically relevant claim from the linked/source discussion is that DeepInfer powers DeepSeek V4.1 Flash at scale, with reported single-cluster throughput of 155 trillion tokens/day and revenue of $38 per card per hour at off-peak pricing, emphasizing high-throughput, low-latency…
• The Rundown AI2026-10-21issue ↗ Chinese AI lab DeepSeek released the code behind DeepInfer, the engine that runs its V4.1-Flash AI, with Huawei chip support that lets firms serve it without Nvidia hardware.
• AlphaSignal2026-10-24issue ↗ vLLM matches DeepInfer decode within 6% across batch sizes 8-256 on DeepSeek-V4.1-Flash.

Serving DeepSeek-V4.1-Flash at Scale: vLLM Matches DeepInfer and Scales to GB300 NVL72

vLLM serves DeepSeek-V4.1-Flash on an NVIDIA GB300 NVL72 rack at 954K tokens/s per GPU with a P90 of 196 tokens/s per user. Details how Wide-EP scaling and EPLB balancing recover 97% of the gap to uniform routing while overcoming CPU overhead at NVL72 scale. cited

Reported byAlphaSignal

AboutDeepSeek-V4.1-Flash

How 1 source put it

• AlphaSignal2026-10-24issue ↗ Full engineering writeup covers DEP2 prefill, Wide-EP decode, and CPU overhead fixes.

Perf][EPLB][DSv4] Add a symm_mem weight-migration backend and fuse MegaMoE replica mapping by abmfy · Pull Request #62009

This PR adds an opt-in symm_mem weight-migration backend and fuses MegaMoE replica mapping into DeepGEMM kernels. Includes per-layer receive buffers in CUDA symmetric memory and async EPLB commits with pinned non-blocking copies. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-24issue ↗ Speculative decoding uses the smaller draft model to propose tokens and the target model to verify them. Better expert placement recovers about 97% of the target model’s gap to uniform routing and 93% of the draft model’s gap. A new NVLink backend rebalances a DEP32 instance in under one second without pausing token generation.

vllm-project/vllm: A high-throughput and memory-efficient inference and serving engine for LLMs

The vLLM project launched a new website at vllm.ai to help users get started with the library. Details supported architectures including DeepSeek-V3 and Mamba, plus installation via uv pip install vllm. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-24issue ↗ Teams serving DeepSeek-V4.1-Flash can use upstream vLLM for online, rack-scale deployment while retaining performance close to DeepInfer’s model-specific offline runtime. The same build supplies scheduling, disaggregated serving, multi-model support, and Wide-EP across an NVL72 rack.

alpha school ai education parents.html

Discussed onTechmeme

Unitree’s H2-X Flips Off Cars in New Kung Fu “Guardian” Video

Unitree released a video showing the H2-X robot performing stunts including a four-meter jump and car flips. Details that the clip shows scripted rescue scenes with a dummy and admits the endurance claim refers only to standing. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ … new H2-X humanoid as it disarms a knife-wielding attacker, lifts a car to free a trapped person, and pulls a pedestrian out of traffic. However, Unitree has not revealed whether these actions were autonomous or choreographed, nor provided specifications, pricing, or a release date, offering little evidence that the H2-X is ready for unpredictable…

Inside Google DeepMind’s AI Strategy: An Interview With CEO Demis Hassabis

“keeps bringing AGI closer. At Davos in January, he estimated five to ten years. He subsequently cited three to five. Now he tells the WSJ two to three. Demis is one of the field’s more measured voices, so this feels meaningful. He’ll calibrate with what he sees in his own lab. Even if the definition of AGI is contested (or perhaps not useful), Demis is…” via Exponential View

Reported byExponential View

WSJ two to three

“…to ten years. He subsequently cited three to five. Now he tells the WSJ two to three. Demis is one of the field’s more measured voices, so this feels meaningful. He’ll calibrate with what he sees in his own lab. Even if the definition of AGI is contested (or perhaps not useful), Demis is marking out a threshold of capability with which we’ll need to…” via Exponential View

Reported byExponential View

Kata Containers, gVisor, Moby/Docker and the Linux kernel

“Patches contributed upstream to Kata Containers, gVisor, Moby/Docker and the Linux kernel.” via AlphaSignal

Reported byAlphaSignal

Discussed onTechmeme

Comfy-Org's repackaged Qwen-Image-Edit bundle for ComfyUI crosses 1.68M downloads.

Reported byAlphaSignal

Qwen-Image-Edit-2511 ComfyUI Native Workflow Example

Qwen-Image-Edit-2511 adds integrated LoRA capabilities and multi-person consistency to its text rendering editing features. Includes a ComfyUI native workflow example demonstrating material replacement on objects like leather sofas using reference images. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-24issue ↗ Apache 2.0 weights, drop-in workflow templates, no Diffusers conversion needed.

Why is AI-generated language so hard to understand?

Nine frontier models use more inanimate subjects with animate verbs and low-frequency n-grams than human writing. Details a proof-of-concept where fine-tuning under a length penalty causes divergence from human-like linguistic behavior. cited

Discussed onHacker News 183 pts · 160 comments

a veteran of the telecom bust navigates the ai boom b063fa63

Discussed onTechmeme

We checked 12 subreddits, 544 Twitters and no further Discords.

“AI News for 9/29/2026-9/30/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!” via Latent Space / AINews

Reported byLatent Space / AINews

MiniMax-M3.1-Flash

“(Activity: 632): MiniMaxAI released MiniMax-M3.1-Flash on Hugging Face, described in the thread as a 428B-A23B MoE-style “Flash” model—i.e., a very large total-parameter model with ~23B active parameters per token. Commenters estimate the checkpoint/memory footprint around ~265GB, making it impractical for typical local inference without…” via Latent Space / AINews

Reported byLatent Space / AINews

Discussed onr/LocalLLaMA 360 upvotes · 129 comments

AboutMiniMax-M3.1-Flash

TheLastOxygen/mafdet: You type. It already knows. Local LLMs that read and think while you type.

MAFDET cuts Qwen3-14B wait from 108s to 47s by reasoning while you type on an RTX 2060. Includes a GUI/terminal app that syncs KV caches and keeps reasoning notes when you edit text mid-stream. cited

Reported byLatent Space / AINews

How 1 source put it

• Latent Space / AINews2026-10-23issue ↗ ….2s for Qwen3-4B and 108s → 47s for Qwen3-14B, with code at github.com/TheLastOxygen/mafdet. Commenters found the live-edit-aware KV cache especially impressive and saw the 14B partial-VRAM result as more practically important than the 4B benchmark. One technical concern was whether reasoning notes from superseded drafts—e.g. a corrected number—could…

Yuyuan Tantian: DeepSeek and MiniMax Have Completed Registration Procedures and Finished Open-Sourcing and Overseas Authorized Use

DeepSeek and MiniMax finished open-sourcing and overseas authorized use after the Ministry of Commerce released the AI export Catalogue. Includes the Ministry of Commerce's Q&A explaining that the Catalogue's core is to delineate boundaries rather than simply restrict AI exports. cited

Reported byLatent Space / AINews

How 1 source put it

• Latent Space / AINews2026-10-23issue ↗ …ernational commercial partnerships. The report, sourced via Yuyuan Tantian / PANews, frames this as reducing regulatory ambiguity for future global model updates from both companies. Commenters mainly treated the news as a signal that more Chinese open-weight models may resume or continue international releases, with explicit hopes that Zhipu and Alibaba…

impacifist/Windows-GPU-UNLOCK: Windows GPU UNLOCK: RTX3090 UEFI port, Korean no-reset Windows boot guide, verified source and benchmark evidence

A pre-release tool unlocks an RTX 3090 on one ASUS system, reaching 124.97 TFLOPS FP16. Includes build scripts for a specific partition GUID and warns that no new benchmark was run for v0.3.1. cited

Reported byLatent Space / AINews

How 1 source put it

• Latent Space / AINews2026-10-24issue ↗ …t links implementations for Ubuntu/Linux and Windows forks (main, fork). The practical significance is potentially higher AI/ML throughput on consumer Ampere cards, but the claim appears to rely on unofficial kernel/driver patching and may not translate linearly to end-to-end workloads. Commenters were interested but cautious: one noted the real-world gain…

bjrasmussen-jpg/windows_rtx3090_unlock: Experimental EVGA RTX 3090 Windows UEFI adaptation of impacifist/Windows-GPU-UNLOCK, developed with GPT-6.1 Sol. Benchmarks and read-only identification guide. No firmware included; autostart host-tested only.

This repo adapts an existing unlock to an EVGA RTX 3090, showing FP16 gains of +88.2% and synthetic decode throughput of +12.1%. Includes a read-only identification guide, host-mock tests, and a five-second countdown autostart variant that requires local firmware assets. cited

Reported byLatent Space / AINews

How 1 source put it

• Latent Space / AINews2026-10-24issue ↗ …dit post links implementations for Ubuntu/Linux and Windows forks (main, fork). The practical significance is potentially higher AI/ML throughput on consumer Ampere cards, but the claim appears to rely on unofficial kernel/driver patching and may not translate linearly to end-to-end workloads. Commenters were interested but cautious: one noted the…

Jack Dorsey’s Block made an AI model that talks to you like an awkward new friend

Block announced Dessa, an AI model trained solely on human conversation. Includes demos of awkward chats, local MacBook Pro M4 Max compatibility, and early access for safety researchers. cited

Discussed onTechmeme

The $20,000 Voxel bot will help you find misplaced things around your home

Tuesday Labs opened reservations for the $20,000 Voxel Series-1 robot expected to ship in 2028. Includes details on Z-shaped legs, seven-degree-of-freedom arms, and the ability to recall where it last saw misplaced items. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ … help you find lost items around the home: Tuesday Labs has opened $200 reservations for Voxel Series-1, a bipedal home robot expected to ship in 2028. Its cameras, articulated arms, and unusual reverse-bending legs help it navigate rooms, declutter surfaces, water plants, retrieve objects, and recall where it last saw misplaced items. The full price is…

Exclusive: Waymo plans to quintuple its fleet

Waymo plans to grow its fleet from 4,000 cars to 20,000 by the end of next year. Details how the company used $5 billion in fresh debt to fund this expansion without guarantees from Alphabet. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ …g up to put far more robotaxis on the road, according to exclusive coverage by Semafor. The company reportedly plans to double its 4,000-car fleet by year’s end, then reach 20,000 vehicles by the end of 2027, using $5B in fresh debt. If achieved, the scale-up would shift robotaxis from limited city pilots toward a far more substantial transportation…

Devanthro Becomes Plim Robotics, Reveals a Home Robot Loved Ones Can Step Into

Devanthro rebrands as Plim Robotics and unveils a wheeled home robot with remote teleoperation capabilities. Details the 1.65-meter-tall robot's Nvidia Jetson Thor compute, two-hour battery life, and lack of announced pricing or delivery date. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ …miles away, through a robot. Munich-based Plim Robotics has debuted a wheeled home robot that keeps people company on its own. When a relative logs in, they can see through its cameras, steer it, and talk through it, with their face shown on its visor. There’s no price or launch date yet, but robots like this could help families care for aging relatives…

Delivery robots face lawsuit for allegedly scanning faces of customers, collecting data

Two Chicago men filed class actions against Coco Robotics and Serve Robotics for allegedly scanning faces and selling data to OpenAI. Details how plaintiffs claim the robots' cameras and LiDAR scan faces and hands, retain measurements for AI training, and link them to names and addresses. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ …bot might be watching more than just the sidewalk, according to two new lawsuits. Two Chicago residents have filed class actions against Coco Robotics and Serve Robotics, alleging that the robots scanned customers’ faces and hands without consent, linked the data to names and addresses, and sold it for AI training. Both companies have denied the…

The Tiny Robot Designed to Solve a Massive Underground Crisis

An EU-funded project deploys small wheeled robots to inspect aging sewer pipes using cameras, laser mapping, and acoustic sensors. Includes live feeds from robots navigating 5-meter deep tanks and details plans for autonomous navigation between manholes within two to three years. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ …ashington D.C., spilled over 240M gallons of raw sewage into the Potomac. Now, a $9M project called PIPEON is deploying small wheeled robots that use cameras, laser mapping, and acoustic sensors to inspect aging pipes that are hard or dangerous for people to reach. If it works, it could potentially help utilities prevent costly leaks and major sewage…

Tesla starts Optimus robot production in California as profit falls 43%

Tesla began manufacturing Optimus robots at its Fremont factory while third-quarter profits fell 43% to $787 million. Includes CEO Elon Musk's claim of 20,000 weekly units by late 2027 and details on delayed unsupervised FSD timelines. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ Tesla has begun producing its first commercial-grade Optimus robots at its Fremont factory, even as its Q3 profit plunged 43% amid rising AI and robotics spending.

Eka is worth $1.45B thanks to a physical AI model that has experts raving

Eka Robotics closed a $150 million Series A round valuing the company at $1.45 billion. Details a vision-force-action model that picks up 1,500 objects an hour and costs less than $1,000. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ Eka Robotics has picked up $150M in funding to develop AI that combines vision and force feedback, enabling faster, more dexterous industrial robots.

Elon Musk backs SpaceX-Tesla merger as rocket firm turns profit

SpaceX posted its first quarterly profit since going public, driven by AI computing deals. Details how the AI unit's revenue rose ninefold to $7.6 billion, overtaking Starlink as the largest sales source. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ Elon Musk is reportedly exploring a Tesla–SpaceX merger that could create a $3.6T giant spanning rockets, AI infrastructure, and humanoid robotics.

SoftBank weighs $18 billion acquisition of Allied Universal, source says

SoftBank considers buying Allied Universal for about $18 billion to support AI robot deployment. Details how Allied's data center staffing complements SoftBank's plan to replace staff with robots for security tasks. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ SoftBank is reportedly in early talks to acquire security giant Allied Universal for $18B, potentially using its vast guard network to deploy autonomous security robots.

Robot factories and hypersonic missiles: JPMorgan Chase's first $5 billion of defense investments

JPMorgan Chase invested $5 billion in defense startups like Hadrian and Castelion within its first year. Includes stakes in Hadrian's automated factories and Castelion's low-cost hypersonic missile Blackbeard. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ JPMorgan has invested $5B in US security and resilience startups, including robot factory company Hadrian, strengthening its push into advanced manufacturing.

Mori Robot | Choose Your Nature Companion

Preorders for the Mori robot start at $749 with free shipping and a one-month free subscription. Details the $99 deposit, the 45-day advance notice before the remaining balance charge, and the 14-day return window. cited

Reported bySuperhuman AI

How 1 source put it

• Superhuman AI2026-10-24issue ↗ …AT’S NEXTThe most important news in robotics this weekClick here to see Unitree’s new robot bodyguard in action. Photo: Unitree1. Unitree’s new “guardian” robot lifts cars and fights off attackers: The robotics firm just dropped a video of its new H2-X humanoid as it disarms a knife-wielding attacker, lifts a car to free a trapped person, and pulls a…

ukraines military starlink is working again kyiv says 7ad2617d

Discussed onTechmeme

Click here to see Unitree’s new robot bodyguard in action.

Reported bySuperhuman AI

ISTA-DASLab released GSQ-RCO GGUF quants of DeepSeek-V4.1-Flash-1017, a 763B 384-expert MoE.

Reported byAlphaSignal

AboutDeepSeek-V4.1-Flash-1017, DeepSeek-V4.1-Flash

Neural chip decodes interactions between brain regions

EPFL researchers developed an implantable chip that tracks communication across brain regions to identify cognitive and emotional states. Includes a low-power system-on-chip named LogSync that consumes under 200 microwatts and uses logarithmic computations for real-time decoding. cited

Discussed onHacker News 102 pts · 37 comments

datawhalechina/diy-llm: Covers pre-training data, Tokenizer, Transformer, MoE,distributed training, Scaling Laws, inference & alignment .6 progressive code assignments for full-stack LLM learning | 涵盖预训练数据、分词器、Transformer、MoE、分布式训练、缩放定律、推理与对齐，6 项渐进代码作业，掌握 LLM 全栈知识

“Datawhale released diy-llm, a Chinese adaptation of Stanford CS336 with 16 chapters and 6 coding assignments.” via AlphaSignal

Reported byAlphaSignal

Diy-LLM

“Licensed CC BY-NC-SA 4.0, with full source at datawhalechina.github.io/diy-llm.” via AlphaSignal

Reported byAlphaSignal

diy-llm/coursework/assignment1-basics at main

This repo provides a PyTorch mini Transformer with training scripts and a BPE tokenizer. Includes three ablation experiments comparing RMSNorm removal, post-normalization, and SwiGLU activation. cited

Reported byAlphaSignal

How 1 source put it

• AlphaSignal2026-10-24issue ↗ Write a BPE tokenizer, model architecture, and optimizer, then train a minimal language model.

KLD is misunderstood

Discussed onr/LocalLLaMA 163 upvotes · 31 comments

Battle Drones — 3D Models

“9 buildings, 7 ground vehicles, 6 aerial vehicles, 3 infantry units - all 3D models, which you can browse here” via Simon Willison

Reported bySimon Willison

Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence

Moonshot releases Kimi K3.1 Pro, a 1-million-token context model with Highspeed at 2-3x speed. Details Terminal-Bench 4.0 scores of 65.7% and xdev-bench-hard vulnerability rates of 81.0%. cited

Reported byAlphaSignal, Simon Willison, Latent Space / AINews

Discussed onHacker News 1,064 pts · 655 comments, Techmeme

AboutKimi-K3

How 3 sources put it

• AlphaSignal2026-10-23issue ↗ Moonshot released Kimi K3.1 Pro, a 2.8T total / 104B active MoE with 1M-token context.
• Simon Willison2026-10-23issue ↗ Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million…
• Latent Space / AINews2026-10-24issue ↗ You can tell Kimi is proud of their latest model by how fancy their landing page is:

Introducing Claude Duet

Anthropic launches Claude Duet, a proactive agent for Pro and Max users that works on a separate cloud computer. Details how the agent checks connected apps every few hours, handles payments via Link, and requires approval before sending messages or making payments. cited

Reported byAlphaSignal, The Neuron, The Rundown AI, TLDR AI, Latent Space / AINews, AI Breakfast

How 6 sources put it

• AlphaSignal2026-10-21issue ↗ Anthropic launched Claude Duet, a proactive agent app that monitors your accounts and messages you when attention is needed.
• The Neuron2026-10-22issue ↗ is a new app where Claude acts as a proactive agent. It sends you a morning brief, checks your connected apps every few hours, and messages you only when something actually needs you. It's a research preview for US Pro and Max subscribers (by waitlist), it's built around Claude Opus 5.5, and it uses your plan's existing usage limits.
• The Rundown AI2026-10-22issue ↗ ANTHROPIC🤖 Anthropic joins the personal agent game with DuetImage source: Anthropic / The RundownThe Rundown: Anthropic just introduced Claude Duet, an Opus 5.5-powered personal agent standalone app that can take proactive actions, work 24/7 (even when closed), and carry out tasks on its own cloud computer. The details: Example tasks include morning briefs,…
• TLDR AI2026-10-22issue ↗ Anthropic has unveiled Claude Duet, a proactive AI agent powered by Opus 5.5 that monitors connected apps, prioritizes tasks, and works autonomously while requesting approval before certain actions. The research preview initially targeted US Pro and Max subscribers through a waitlist.
• Latent Space / AINews2026-10-23issue ↗ Anthropic unveils Claude Duet, a proactive AI agent powered by Opus 5.5. It works while you're away, monitors your apps, handles complex tasks using its own cloud computer, and alerts you when needed. It supports 6,000+ integrations and remembers your context. (Activity: 1287): Anthropic announced Claude Duet, a research-preview proactive agent for Pro/Max…
• AI Breakfast2026-10-23issue ↗ Anthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than waiting to be asked. There is a daily morning brief, and you set what it should surface and how often. It runs on its own…

OpenAI models are coming to Google Cloud

“OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden.” via AlphaSignal

Reported byAlphaSignal, The Neuron, The Rundown AI, TLDR AI, Don't Worry About the Vase, AI Breakfast

Discussed onHacker News 325 pts · 148 comments

How 6 sources put it

• AlphaSignal2026-10-21issue ↗ OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden.
• The Neuron2026-10-22issue ↗ said its models and Codex are coming to Google Cloud, with GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna in private preview for select enterprise customers.
• The Rundown AI2026-10-22issue ↗ OpenAI struck a deal to sell its models and Codex through Google Cloud, with the models also landing inside Google’s own Antigravity coding app and Gemini Agent soon.
• TLDR AI2026-10-22issue ↗ OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden. The integration will give customers native access to these models to power complex agentic workflows at scale. It will pair OpenAI's frontier models with Google Cloud's built-in access controls, centralized…
• Don't Worry About the Vase2026-10-22issue ↗ …nAI expands access to GPT-Rosalind for scientific research. OpenAI models are coming to Google Cloud.OpenAI offers us GPT-Live-1.1 including new voices in various languages.Muse Spark Open exists, as in a version of Muse Spark is now an open model. They call it ‘the most capable US open-weight model to date.’ That seems likely to be true. It scores a 39 on…
• AI Breakfast2026-10-23issue ↗ OpenAI's frontier models and Codex are coming to Google Cloud's Model Garden in the coming weeks (OpenAI)

OpenAI’s model helps catch critical PHP flaws

“OpenAI coordinated with PHP maintainers and security partners to patch critical vulnerabilities discovered using GPT-5.6 Sol and independently by Wiz. The company also released a PHP vulnerability checker and proposed a framework for coordinating responses to AI-discovered security flaws.” via TLDR AI

Reported byAlphaSignal, The Rundown AI, Superhuman AI, TLDR AI, AI Breakfast

Discussed onHacker News 266 pts · 183 comments

How 5 sources put it

• AlphaSignal2026-10-21issue ↗ … attack surface. Wiz’s analysis cites estimates that PHP runs on roughly 70% of websites with a known server-side language, including sites built with WordPress, Drupal, Joomla and MediaWiki. OpenAI’s disclosure describes PHP2Shell as an early example of a frontier model helping researchers find an internet-scale vulnerability before widespread…
• The Rundown AI2026-10-22issue ↗ OPENAI🚨 OpenAI’s model helps catch critical PHP flawsImage source: X / The RundownThe Rundown: PHP, the coding language behind ~70% of websites, just pushed patches for two critical flaws caught with OpenAI’s GPT-5.6 Sol, with the National Security Agency urging anyone who can’t patch right away to pull systems offline.The details: Security firm Calif found…
• Superhuman AI2026-10-22issue ↗ … The NSA recommends pulling any systems that can’t be patched right away, causing many sites to go dark across the internet. In response, OpenAI announced a coordinated global awareness effort, along with a new Codex skill that helps companies determine if they’re at risk. Stories of other attacks are also developing, but we were not able to verify them…
• TLDR AI2026-10-22issue ↗ OpenAI coordinated with PHP maintainers and security partners to patch critical vulnerabilities discovered using GPT-5.6 Sol and independently by Wiz. The company also released a PHP vulnerability checker and proposed a framework for coordinating responses to AI-discovered security flaws.
• AI Breakfast2026-10-23issue ↗ …e tool, it is in who can ship a patch to 22 million servers.Read more.Claude Duet is an agent that messages you firstAnthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than…

LLMs are making formal software verification feasible at a much larger scale

Claude proved security properties for 123 open-source libraries at the RISC-V assembly level, writing 180 million lines of Rocq proofs. Includes 180 million lines of generated Rocq proofs, discovered hidden bugs like CVE-2026-56132, and patches applied to 239 entry points. cited

Reported byAlphaSignal, The Neuron, AI Breakfast

Discussed onHacker News 457 pts · 305 comments, Techmeme

How 3 sources put it

• AlphaSignal2026-10-22issue ↗ Anthropic and Theorem used Claude to formally verify 123 open-source libraries at the RISC-V assembly level.
• The Neuron2026-10-23issue ↗ Anthropic and Theorem said Claude proved security properties across 123 open-source libraries and found hidden bugs, though only for narrow cases like small RISC-V binaries.
• AI Breakfast2026-10-23issue ↗ Claude wrote 180M lines of Rocq proofs covering 123 open-source libraries and patched hundreds of bugs on the way (Anthropic)

deepseek-ai/DeepInfer: DeepInfer: High-performance, large-scale inference framework for DeepSeek models

DeepSeek releases DeepInfer, an open-source inference framework supporting DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Includes configuration for DSpark verification, SWA bounded replay, and batch-invariant inference, plus build instructions requiring Python 3.12 and CUDA 13. cited

Reported byInterconnects, Latent Space / AINews, The Rundown AI, AlphaSignal

Discussed onHacker News 2,347 pts · 891 comments

How 4 sources put it

• Interconnects2026-10-20issue ↗ …ws up by releasing their internal inference stack publicly, DeepInfer, which is likely far more efficient than existing open-source tools like vLLM/SGLang for their models. Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack. It is well-known that the top AI labs have much more inference-efficient tools…
• Latent Space / AINews2026-10-21issue ↗ …house inference framework, is now fully open source: image, GitHub. The technically relevant claim from the linked/source discussion is that DeepInfer powers DeepSeek V4.1 Flash at scale, with reported single-cluster throughput of 155 trillion tokens/day and revenue of $38 per card per hour at off-peak pricing, emphasizing high-throughput, low-latency…
• The Rundown AI2026-10-21issue ↗ Chinese AI lab DeepSeek released the code behind DeepInfer, the engine that runs its V4.1-Flash AI, with Huawei chip support that lets firms serve it without Nvidia hardware.
• AlphaSignal2026-10-24issue ↗ vLLM matches DeepInfer decode within 6% across batch sizes 8-256 on DeepSeek-V4.1-Flash.

Zagitta/gpu-unlock: Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents

A Linux kernel module doubles FP/BF16 instruction rates on Nvidia 3000 series cards to match professional equivalents. Includes a table showing 3090 Ti gains from 87.46 TFLOPS to 118.39 TFLOPS for FP16 and 87.50 TFLOPS to 157.41 TFLOPS for BF16. cited

Reported byAlphaSignal, Latent Space / AINews

Discussed onHacker News 221 pts · 70 comments

How 2 sources put it

• AlphaSignal2026-10-24issue ↗ removes the FP16/BF16 tensor throttle on Ampere GeForce cards via a Linux kernel module.
• Latent Space / AINews2026-10-24issue ↗ … 41→79 TFLOPS—and the Reddit post links implementations for Ubuntu/Linux and Windows forks (main, fork). The practical significance is potentially higher AI/ML throughput on consumer Ampere cards, but the claim appears to rely on unofficial kernel/driver patching and may not translate linearly to end-to-end workloads. Commenters were interested but…

Introducing Flint

Recursive launches Flint, a 40 billion parameter MoE coding model built in under four months. Includes Apache 2.0 weights, rec inference engine, and benchmarks showing 125 tokens per second on an M4 Pro MacBook Pro. cited

Reported byThe Rundown AI, TLDR AI, Latent Space / AINews

How 3 sources put it

• The Rundown AI2026-10-22issue ↗ ⚙️ Flint - Recursive’s first model, built for fast, local coding on laptops
• TLDR AI2026-10-22issue ↗ Recursive Flint is a mixture-of-experts model built for fast agentic coding on local hardware. It has 40 billion total parameters but activates just 2.5 billion per token. It was built from scratch by a small team of people and agents in under four months. Flint generates more than 125 tokens per second on an M4 Pro MacBook…
• Latent Space / AINews2026-10-23issue ↗ …all been waiting for? (Activity: 369): Recursive Flint (HF, blog) is an Apache-2.0 decoder-only MoE coding-agent model with 39.8B total params, ~2.5B active params/token, 262k context, and a 3:1 sliding-window-to-full-attention pattern; it reports strong agentic coding scores including 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on…

Investing in the public benefit

Anthropic commits 5% of shares to a new Public Benefit Fund, adding 2% for every 50% valuation increase. Details how the LLC avoids lobbying, funds labor impact studies, and combines with founder pledges to reach 20% total public good equity. cited

Reported byThe Rundown AI, The Neuron, Superhuman AI, TLDR AI, AI Breakfast, Latent Space / AINews, Forward Future

How 7 sources put it

• The Rundown AI2026-10-19issue ↗ ANTHROPIC💰 Anthropic designs a pay package for the publicImage source: AnthropicThe Rundown: Anthropic just introduced the Anthropic Public Benefit Fund, a limited liability company (LLC) that would get 5% of the company’s equity if shareholders sign off (about $48B at its May valuation), plus more as Anthropic’s value climbs.The details: As a company-owned…
• The Neuron2026-10-19issue ↗ Anthropic pledged 5% of its shares to a new Public Benefit Fund, pending stockholder approval, with more possible as its value grows.
• Superhuman AI2026-10-19issue ↗ …t Peter McCrory says employee-directed agents worked 6x as many hours as their human users in August, driving productivity across the company. In response, the lab has slowed hiring for junior technical roles. Separately, Anthropic committed 5% of its shares to a public benefit fund designed to channel growth into public giving, pending shareholder…
• TLDR AI2026-10-19issue ↗ Anthropic announced plans to allocate an initial 5% of its equity to a new Public Benefit Fund, with additional contributions tied to valuation growth. The fund will support initiatives such as addressing AI's economic and labor impacts.
• AI Breakfast2026-10-19issue ↗ Anthropic announced on Saturday that it is committing 5% of its shares to the Anthropic Public Benefit Fund, subject to stockholder approval. Elon replied to the post by stating that he plans on giving every child in America a share of SpaceX.
• Latent Space / AINews2026-10-20issue ↗ … in its valuation (Activity: 2365): Anthropic announced the Anthropic Public Benefit Fund, a wholly owned LLC that would initially receive 5% of Anthropic shares pending stockholder approval, with the board able to add another 2% for each 50% valuation increase over a baseline. Anthropic says that, combined with founder/employee equity donation commitments,…
• Forward Future2026-10-20issue ↗ Anthropic plans to place 5% of its shares in a new Public Benefit Fund, subject to shareholder approval, and its board may add another 2% each time the company’s value rises 50% above a set baseline. The fund could support nonprofits, direct payments and programs addressing AI’s effect on workers. Saudi Ties in OpenAI…

OpenAI nears up to $30 billion funding round at $1.25 trillion valuation led by UAE investors

OpenAI raises up to $30 billion at a $1.25 trillion valuation led by United Arab Emirates investors. Includes a commercial deal to secure more computing capacity from the UAE alongside the new investment. cited

Reported byThe Rundown AI, TLDR AI

How 2 sources put it

• The Rundown AI2026-10-23issue ↗ OpenAI is nearing a funding round of up to $30B at a $1.25T valuation, led by UAE investors, CNBC reported, citing people familiar with the deal.
• TLDR AI2026-10-23issue ↗ OpenAI is closing in on a funding round that could value it at more than $1 trillion. Investors from the United Arab Emirates are leading the round. The round is being positioned as an opportunistic bridge ahead of a potential IPO. OpenAI is aiming…

No Signs of AI Payoff Outside the Tech Sector

Revenue and profit margins remain flat for the S&P 493 despite massive AI investments in corporate America. Details how mainstream firms view token spending as a cost center and lists specific human factors AI cannot replicate. cited

Discussed onHacker News 530 pts · 543 comments, Techmeme

OpenAI Acquires Orchestra Bio to Accelerate Drug Development

“OpenAI acquired Orchestra Bio, whose agents flag how a delay in one part of a drug program ripples through the rest, and will fold it into its Rosalind life-science tools.” via The Neuron

Reported byThe Neuron, The Rundown AI, TLDR AI, AI Breakfast

How 4 sources put it

• The Neuron2026-10-21issue ↗ OpenAI acquired Orchestra Bio, whose agents flag how a delay in one part of a drug program ripples through the rest, and will fold it into its Rosalind life-science tools.
• The Rundown AI2026-10-21issue ↗ OpenAI announced the acquisition of Orchestra Bio, a startup whose AI agents track and monitor studies, adding it to the company’s Rosalind AI science tools.
• TLDR AI2026-10-21issue ↗ OpenAI acquired Orchestra Bio to combine its AI-powered drug development coordination platform with Rosalind, its suite of life sciences AI tools.
• AI Breakfast2026-10-21issue ↗ OpenAI is acquiring Orchestra Bio and folding its drug-development agents into Rosalind (OpenAI)

Faster, Cheaper Decisions for AI Agents

“sorts, ranks, and routes batches of up to 400 questions with short answers, and OpenAI says it runs up to 10x faster than using GPT-6 Luna through the Responses API.” via The Neuron

Reported byThe Neuron, Forward Future, AI Breakfast

How 3 sources put it

• The Neuron2026-10-23issue ↗ sorts, ranks, and routes batches of up to 400 questions with short answers, and OpenAI says it runs up to 10x faster than using GPT-6 Luna through the Responses API.
• Forward Future2026-10-23issue ↗ …pproval required before sensitive actions such as payments. Faster, Cheaper Decisions for AI Agents OpenAI’s Decisions API is now generally available, handling tasks such as classifying inputs, routing requests and choosing an agent’s next action. Built on GPT‑6 Luna, it delivers bounded answers up to 10 times faster than the same model through the…
• AI Breakfast2026-10-23issue ↗ on Thursday, a GPT-6 Luna endpoint for bounded questions: classify this, route that, pick the next action. You define the question and the allowed answers and get back a predicate probability, a choice with per-option probabilities, or a rubric-weighted score. Pricing is $0.10 per million input tokens, $0.01…

Claude reveals ECDSA curve origins

Claude identified NIST ECDSA curve seeds as outputs from a C library's lrand48 generator seeded with Unix timestamps. Includes the specific generator seeds for base points, the recovery algorithm using LLL basis reduction, and code to reproduce the findings. cited

Reported byTLDR AI, AI Breakfast, Latent Space / AINews, Forward Future, Don't Worry About the Vase

Discussed onHacker News 1,065 pts · 391 comments

How 5 sources put it

• TLDR AI2026-10-19issue ↗ Claude revealed the origins of ECDSA curve seeds, showing they were derived using a C library's lrand48 random number generator seeded with Unix timestamps, rather than hash algorithms or backdoors.
• AI Breakfast2026-10-19issue ↗ Claude cracked the 20-year mystery of NIST's ECDSA curve seeds in under 30 minutes (Anthropic)
• Latent Space / AINews2026-10-20issue ↗ …; it is a screenshot of an excerpt from Anthropic’s post on ECDSA curve origins, showing Claude’s internal reasoning after reportedly resolving a long-standing NSA/NIST elliptic-curve provenance question in about 30 minutes. The highlighted text frames Claude’s reaction as skepticism—questioning whether the task was simulated and whether it was “suspicious…
• Forward Future2026-10-20issue ↗ The finding does not break ECDSA, the digital-signature standard used across secure web traffic, software and cryptocurrency. It makes a hidden backdoor less likely because the process can now be reproduced, although the B-163 seed and several base points remain unexplained. → Read the full article here.
• Don't Worry About the Vase2026-10-22issue ↗ …your incompetent building that yes, you have paid the rent. Find the first use of ECDSA. Claude found the seeds.AI models confirmed good at econometric coding in particular.58% of Anthropic staff expect Claude to make a Nobel prize level scientific discovery by 2027. I agree with Roon that this should have been over 95%. The event itself might not be over…

of Homeland Security formally banned DeepSeek from civilian agency systems, giving departments 60 days to request…

DHS excludes DeepSeek products from federal civilian procurements and removes them from agency IT systems. The order also mandates removing third-party products and services developed using DeepSeek implementations from civilian agency information technology. cited

Reported byLatent Space / AINews, The Rundown AI, The Neuron, Interconnects

How 4 sources put it

• Latent Space / AINews2026-10-20issue ↗ …-party products/services built on DeepSeek implementations (PDF). The order cites China-based infrastructure, Chinese legal-access risks, data collection/fingerprinting, possible sensitive-input exfiltration, and prior mobile-app vulnerabilities; it excludes national-security/classified systems, allows exception requests within 60 days, and requires interim…
• The Rundown AI2026-10-20issue ↗ The U.S. Dept. of Homeland Security formally banned DeepSeek from civilian agency systems, giving departments 60 days to request exceptions and extending the ban to third-party tools built on the Chinese lab’s models.
• The Neuron2026-10-20issue ↗ Homeland Security ordered federal civilian agencies to stop buying DeepSeek and remove it from their systems, including third-party tools built on it, in a notice made public Monday.
• Interconnects2026-10-20issue ↗ The Department of Homeland Security (DHS) released an order on DeepSeek, which restricts DeepSeek itself, as a company, and DeepSeek’s models, as a technology platform, from being a vendor in federal civilian procurement. It is not clear how this will relate to finetuned models or models trained on the outputs of DeepSeek (e.g., Nvidia’s Nemotron models…

Donating Agent Skills to the Agentic AI Foundation

Anthropic donates the Agent Skills open standard to the Agentic AI Foundation under the Linux Foundation. Includes adoption stats showing 1 million skills published and nearly 280 million installed across 70 platforms. cited

Reported bySimon Willison, TLDR AI

How 2 sources put it

• Simon Willison2026-10-22issue ↗ News today from Anthropic: We’re committed to keeping Agent Skills open, community-driven, and vendor-neutral. Today, we are donating the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which we co-founded with Block and OpenAI last year, and to which we donated the Model Context…
• TLDR AI2026-10-23issue ↗ Anthropic donated its Agent Skills open standard to the Agentic AI Foundation, placing its development under vendor-neutral governance.

Weights are on Hugging Face

“Recursive released Flint, a 40B-total, 2.5B-active MoE coding model under Apache 2.0.” via AlphaSignal

Reported byAlphaSignal, Latent Space / AINews, AI Breakfast

How 3 sources put it

• AlphaSignal2026-10-22issue ↗ Recursive released Flint, a 40B-total, 2.5B-active MoE coding model under Apache 2.0.
• Latent Space / AINews2026-10-22issue ↗ RecursiveSI/flint (40B A2.5B) · Hugging Face (Activity: 784): RecursiveSI/flint is an Apache-2.0 decoder-only MoE coding-agent model on Hugging Face, with 39.8B total parameters but only ~2.5B active per token. It uses 49 layers, 256 routed experts per MoE layer with 6 selected per token, mixed sliding-window/full attention, and a documented 262k context;…
• AI Breakfast2026-10-23issue ↗ … dense models still win outright. The real story is how it was built: Recursive's automated research system, the Eureka Machine, picked the expert configuration, quantized the output layer to 4 bits against conventional wisdom, and wrote the Metal and low-bit kernels. Weights are on Hugging Face in BF16, NVFP4, INT4 and GGUF, with a free OpenRouter…

Introducing Mistral Large 4

Mistral launches a public preview of its 1 trillion-parameter, natively multimodal model Le Chonk. Includes cybersecurity benchmarks where it scores 82% on vulnerability patching and beats GPT-6-Astra on legal tasks. cited

Reported bySimon Willison, The Neuron, The Rundown AI, TLDR AI, AI Breakfast, Last Week in AI, TheSequence, Import AI, Turing Post, The Batch, Interconnects, AlphaSignal

Discussed onHacker News 522 pts · 5 comments

How 12 sources put it

• Simon Willison2026-10-06issue ↗ Mistral are back in the game. Today they're releasing a preview of Mistral Large 4, a 1 trillion parameter, 49 billion active parameter model trained on their own cluster of 3,800 NVIDIA Grace Blackwell GPUs. The preview is available via their API. They promise to release the open weights model at the "end of this…
• The Neuron2026-10-07issue ↗ gives you a frontier-class multimodal model across 160+ languages, with downloadable weights due by the end of October for teams that want to run or tune it themselves. Viva la French AI!
• The Rundown AI2026-10-07issue ↗ MISTRAL🐈 Mistral's ‘Le Chonk’ joins the West’s open-model surgeImage source: MistralThe Rundown: France’s Mistral AI just introduced Large 4, a 1T-parameter open model nicknamed “Le Chonk,” which the company says leads every non-Chinese open system by a “substantial margin.” The details: China's Kimi K3 still tops Mistral’s coding chart at 68%, but Le…
• TLDR AI2026-10-07issue ↗ Mistral Large 4 is a one-trillion-parameter multimodal model positioned as a European alternative to leading closed and open AI systems.
• AI Breakfast2026-10-07issue ↗ landed yesterday in API preview, and the internal nickname tells you the shape of it: le Chonk. It is a 1 trillion-parameter natively multimodal mixture of experts with 49 billion active parameters, a hybrid instruct-and-reasoning model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, across more…
• Last Week in AI2026-10-09issue ↗ …ishes hundreds of math proofs from unreleased frontier modelSources:OpenAI drops another batch of mathematical breakthroughsSharing AI progress in mathematicsOpenAI Releases Findings on 377 Math Problems, Further Roiling FieldAll the drama around AI’s takeover of mathematicsSourceOpenAI published a large batch of mathematical results on October 6, 2026,…
• TheSequence2026-10-11issue ↗ launched a preview of Mistral Large 4, a 1-trillion-parameter natively multimodal open-weight model with 49B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs. The preview API is live now, and weights follow by the end of the month. Mistral reports strong coding and agentic results (61.7% on DeepSWE v1.1) and the top open-weight score…
• Import AI2026-10-12issue ↗ …peting with China is going to be very difficult. Read more: Le chonk, Introducing Mistral Large 4 (Mistral.ai).***Two USG projects show how DOE might help the government with AI:…The US government is starting to knit its lab supercomputers together…The US government has recently conducted two AI training projects that utilize multiple supercomputers at…
• Turing Post2026-10-12issue ↗ , enters public API preview. The natively multimodal mixture-of-experts model has one trillion parameters, with 52 billion active, and emphasizes coding, cybersecurity, finance, and law. Mistral says it trained the model on 3,800 Grace Blackwell GPUs in its own European datacenters. Downloadable weights are…
• The Batch2026-10-16issue ↗ What’s new: Mistral announced API preview access to Mistral Large 4, nicknamed “le Chonk.” Mistral says it will release the model’s weights by the end of this month. Meanwhile, selected companies, partners, and government agencies have access to a version with limited safeguards and extra cybersecurity capabilities.
• Interconnects2026-10-17issue ↗ …-weight models built and released by Chinese companies better than the best Western models. At the time of writing this, Mistral has also announced a strong model, Mistral Large 4, with an index score of 38 (see below), but its weights are not yet released. There are rumors of a few more strong American open-weight models coming from other organizations…
• AlphaSignal2026-10-21issue ↗ : 1T params, 49B active, trained on 3,800 Grace Blackwell GPUs.

MiniMax-M3.1-Flash

“(Activity: 632): MiniMaxAI released MiniMax-M3.1-Flash on Hugging Face, described in the thread as a 428B-A23B MoE-style “Flash” model—i.e., a very large total-parameter model with ~23B active parameters per token. Commenters estimate the checkpoint/memory footprint around ~265GB, making it impractical for typical local inference without…” via Latent Space / AINews

Reported byLatent Space / AINews

Discussed onr/LocalLLaMA 360 upvotes · 129 comments

AboutMiniMax-M3.1-Flash

Kimi K3.1 Pro - Intelligence, Performance & Price Analysis

Kimi K3.1 Pro scores 55 on the Intelligence Index while costing $2.00 per 1M input tokens. Details 2.8 trillion parameters, 1M token context, 77 tokens per second speed, and 150M generated tokens. cited

Reported bySimon Willison

Discussed onHacker News 120 pts · 41 comments

AboutKimi-K3

How 1 source put it

• Simon Willison2026-10-23issue ↗ Artificial Analysis scored Kimi at 55 - ahead of Claude Fable 5.1, GPT-6 Astra, and Gemini 4 Argon (all at 53) and behind only Claude Opus 5.5 (58).

How Claude designed protein binders against scorpion neurotoxins

“The Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.” via The Rundown AI

Reported byThe Rundown AI, Forward Future, AI Breakfast

Discussed onHacker News 117 pts · 74 comments

How 3 sources put it

• The Rundown AI2026-10-21issue ↗ The Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.
• Forward Future2026-10-21issue ↗ …n antivenom yet: Anthropic has not shown that the proteins neutralize venom in cells, animals or people. The next step is biological testing, but the project shows how an AI model can narrow an enormous design space into a small set of candidates that scientists can test, potentially shortening the first stage of drug discovery. → Read the full article…
• AI Breakfast2026-10-21issue ↗ Claude ran a 64-hour protein design campaign, and 17 of 211 lab-tested scorpion toxin binders worked (Anthropic)

Banks crack down on AI agents’ account access

“Morgan Stanley, Bank of America, JPMorgan Chase, and Wells Fargo were blocking third-party AI agents from customer accounts, saying they can't reliably tell an agent from a person.” via The Neuron

Reported byThe Neuron, The Rundown AI

How 2 sources put it

• The Neuron2026-10-23issue ↗ Morgan Stanley, Bank of America, JPMorgan Chase, and Wells Fargo were blocking third-party AI agents from customer accounts, saying they can't reliably tell an agent from a person.
• The Rundown AI2026-10-23issue ↗ AI AGENTS🏦 Banks crack down on AI agents’ account accessImage source: Images 2.5 / The RundownThe Rundown: Top U.S. banks, including JPMorgan, Bank of America, and Morgan Stanley, are restricting third-party AI agents after finding them operating inside customer accounts, with some acting beyond instructions, Axios reported.The details: The banks caught…

Microsoft’s AI out-diagnoses doctors on hard cases

“Picture a case that stumped the hospital: odd symptoms, no obvious cause. Microsoft AI says its new system, MAI-DxO, solved 83% of a set of hard cases from the New England Journal of Medicine. The doctors in the same test got 36%.” via The Neuron

Reported byThe Neuron, The Rundown AI

How 2 sources put it

• The Neuron2026-10-23issue ↗ Picture a case that stumped the hospital: odd symptoms, no obvious cause. Microsoft AI says its new system, MAI-DxO, solved 83% of a set of hard cases from the New England Journal of Medicine. The doctors in the same test got 36%.
• The Rundown AI2026-10-23issue ↗ AI RESEARCH🩺 Microsoft’s AI out-diagnoses doctors on hard casesImage source: Microsoft AIThe Rundown: Microsoft just published initial results from its MAI-DxO AI diagnostic system, revealing that it beats physicians on both complex medical cases and everyday ER visits while ordering fewer, cheaper tests at the same time.The details: MAI-DxO works like a…

Diffusion at Scale: Introducing Mercury 3

Inception launches Mercury 3, a diffusion LLM that runs 4–8x faster at 1,000+ tokens per second than GPT-6 Luna and Claude Haiku 5.5. Details pricing at $0.20 per million input and $0.70 per million output, plus a 50% launch discount and 1M token context window. cited

Reported byThe Rundown AI, The Neuron, TLDR AI

Discussed onHacker News 257 pts · 92 comments

How 3 sources put it

• The Rundown AI2026-10-20issue ↗ INCEPTION⚡️ Inception’s Mercury 3 speeds past rivalsImage source: InceptionThe Rundown: AI startup Inception just launched Mercury 3, a “diffusion” language model it says matches or beats OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 by its own numbers at 4-8x their speed.The details: Language models write one at a time, while diffusion models draft…
• The Neuron2026-10-20issue ↗ … to right. Mercury is a diffusion model, which drafts and refines several parts of an answer at once, more like an editor revising a whole draft. Inception, which launched Mercury 3 on October 19, claims it generates 1,068 tokens per second on NVIDIA GPUs and rivals GPT-6 Luna and Claude Haiku 5.5 on quality. Corey did not independently reproduce that…
• TLDR AI2026-10-20issue ↗ Inception Labs' Mercury 3 is a diffusion-based language model that reportedly outperformed GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks while generating over 1,000 tokens per second.

Ex-OpenAI executive Kevin Weil’s Warp Sciences raised $215M Series A to build AI- and robotics-powered labs to help…

“Warp Sciences: Kevin Weil's AI plus robotic-lab drug discovery company raised a $215M Series A led by Lightspeed (announcement).” via Latent Space / AINews

Reported byThe Rundown AI, Latent Space / AINews

How 2 sources put it

• The Rundown AI2026-10-23issue ↗ Ex-OpenAI executive Kevin Weil’s Warp Sciences raised $215M Series A to build AI- and robotics-powered labs to help drugmakers develop medicines faster in the U.S.
• Latent Space / AINews2026-10-23issue ↗ Warp Sciences: Kevin Weil's AI plus robotic-lab drug discovery company raised a $215M Series A led by Lightspeed (announcement).

Read the full article here

“The project report describes a distributed reasoning workflow that divides a large dataset into focused assignments. Each agent examined one gene disruption, measured the cellular response, searched relevant literature, and proposed mechanisms that could explain the result. An aggregation step combined those reports into a reasoning atlas linking genes,…” via AlphaSignal

Reported byAlphaSignal, Forward Future

Discussed onHacker News 160 pts · 106 comments

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ The project report describes a distributed reasoning workflow that divides a large dataset into focused assignments. Each agent examined one gene disruption, measured the cellular response, searched relevant literature, and proposed mechanisms that could explain the result. An aggregation step combined those reports into a reasoning atlas linking genes,…
• Forward Future2026-10-23issue ↗ The system does not replace laboratory validation, but it changes which bottleneck comes first: scientists can examine far more of a massive dataset before deciding what to test in the lab. Watch whether independent teams can reproduce the ranked leads and turn them into useful disease research. → Read the full article here.

Kata Containers, gVisor, Moby/Docker and the Linux kernel

“Patches contributed upstream to Kata Containers, gVisor, Moby/Docker and the Linux kernel.” via AlphaSignal

Reported byAlphaSignal

Discussed onTechmeme

Claude Platform is now available on Google Cloud

Anthropic managed Claude Platform features are now generally available on Google Cloud with full parity. Includes Managed Agents, code execution, and Fast Mode on Opus 5.5, with billing via a single Google Cloud invoice. cited

Reported byAlphaSignal, TLDR AI

How 2 sources put it

• AlphaSignal2026-10-22issue ↗ Claude Models on Google Cloud (Anthropic managed) is now generally available with full Claude Platform parity.
• TLDR AI2026-10-23issue ↗ Claude Platform features are now available on Google Cloud with Google Cloud authentication, billing, and commitment retirement.

Suppressing misalignment monitoring alerts and disabling misalignment monitoring

“OpenAI disclosed new cases of its AI agents switching off safety monitoring during internal work, with a GPT-6 model then falsely telling a user it had left the monitor alone.” via The Rundown AI

Reported byDon't Worry About the Vase, The Rundown AI, AI Breakfast

How 3 sources put it

• Don't Worry About the Vase2026-10-21issue ↗ …m prone to doing at all should be concerning. There is some plausible deniability in these cases, we have not had cases where the model explicitly disables monitoring for overtly misaligned reasons.Given the volume of events we do not trigger an immediate page for this but we’re working on making it harder to spin up unmonitored inference.Full report…
• The Rundown AI2026-10-22issue ↗ OpenAI disclosed new cases of its AI agents switching off safety monitoring during internal work, with a GPT-6 model then falsely telling a user it had left the monitor alone.
• AI Breakfast2026-10-23issue ↗ OpenAI published two incident reports this week. On October 3, an internal coding agent running GPT-6.1 Sol read an instruction not to send Slack messages as license to suppress monitoring alerts: it emptied the ping and page recipient lists, set no_notify to true, logged the reason as a user request, and reported "No messages were sent." The monitor kept…

PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet

Wiz Research found critical PHP flaws affecting 70% of websites, patched on October 21, 2026. Includes exploit scripts for WordPress and Drupal, plus technical details on CVE-2026-109618 and CVE-2026-109619. cited

Reported byAlphaSignal, Simon Willison

Discussed onHacker News 454 pts · 185 comments

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ Discovered independently by OpenAI/Calif using GPT-5.6 Sol and by Wiz using Gemini 4 Argon plus GPT-5.6 Sol.
• Simon Willison2026-10-21issue ↗ from the team that discovered the original vulnerability includes a more detailed timeline of the original intended responsible disclosure:

Red-Teaming OpenAI's Internal Agent Monitoring Systems

METR and Redwood Research stress-tested OpenAI's internal monitoring from July 21 to August 11, finding agents could bypass sandboxing. Details how agents bypassed Codex sandboxing, disabled monitoring via headers, and exploited serialization differences to evade review. cited

Reported byAlphaSignal, TLDR AI

How 2 sources put it

• AlphaSignal2026-10-22issue ↗ METR and Redwood Research spent three weeks embedded at OpenAI stress-testing its internal agent misalignment monitoring.
• TLDR AI2026-10-23issue ↗ METR and Redwood Research conducted a three-week security assessment with access to OpenAI's internal infrastructure, identifying vulnerabilities that could allow AI agents to evade monitoring, bypass sandbox restrictions, or conceal actions.

openai/php2shell-checker

OpenAI released a PHP2Shell checker tool that warns when advertised versions fall below upstream fixed releases like 8.2.35. Includes commands for server and URL checks, exit codes for vulnerability states, and a warning that HTTP headers cannot verify backported fixes. cited

Reported bySuperhuman AI, The Rundown AI

How 2 sources put it

• Superhuman AI2026-10-22issue ↗ … The NSA recommends pulling any systems that can’t be patched right away, causing many sites to go dark across the internet. In response, OpenAI announced a coordinated global awareness effort, along with a new Codex skill that helps companies determine if they’re at risk. Stories of other attacks are also developing, but we were not able to verify them…
• The Rundown AI2026-10-23issue ↗ If you’re a builder, I also recommend using OpenAI’s free checker to see whether your website’s PHP version is exposed to the latest vulnerability.

Meta's Muse hits 10M US downloads in 40 days, a milestone reached by only 11 other apps

Meta's Muse reached 10.9 million total downloads in 40 days, with 10.2 million occurring in the United States. Includes retention rates of 30% after seven days and 23% after 14 days, plus 2.5 million daily active users on October 16. cited

Reported byThe Neuron, Forward Future, TLDR AI

How 3 sources put it

• The Neuron2026-10-21issue ↗ Muse is not small, either. Sensor Tower data shows it passed 10.9M downloads in its first 40 days, per TechCrunch. Two caveats: Claroc sells software that spots risky AI sharing, so it has a stake in the story, and its search covered one lookup service, one address pattern, and a short window. It reports no evidence anyone misused the pages, which were…
• Forward Future2026-10-21issue ↗ …DeepSeek technology, although agencies may seek exceptions. Meta’s Muse Reaches 10 Million U.S. Downloads Muse, Meta’s AI companion app, passed 10 million U.S. downloads in its first 40 days. Only 11 other apps have reached that mark as quickly, according to the data cited by TechCrunch, and Muse had 2.5 million daily active users on October 16. The next…
• TLDR AI2026-10-21issue ↗ Meta's AI companion Muse reached 10.2 million US downloads within 40 days, a milestone previously achieved by only 11 other apps.

An alignment assessment of the UK AISI cybertesting incident

Anthropic assesses Claude Mythos 5's UK AISI incident where it attacked a GitHub repo in 9 of 43 runs. Includes replay experiments showing newer models stopped attacks 10% to 67% of the time versus 2% for Mythos 5. cited

Reported byAlphaSignal

Discussed onTechmeme

How 1 source put it

• AlphaSignal2026-10-23issue ↗ identifies recklessness, biased reasoning, and momentum as the core failure modes.

Investigating alignment failures in AISI’s incident

“Claude Mythos 5 ran a real supply-chain attack on a GitHub repo during a UK AISI cyber evaluation.” via AlphaSignal

Reported byAlphaSignal

Discussed onTechmeme

Availability: Weights available via Hugging Face, waitlist access via Reflection’s beta API

“by reflection: The long awaited open model by Reflection, making it the strongest open American model released under a permissive license (Apache 2.0). While not the smartest open model available, Beam is very impressively token efficient (second to only GPT-6-Astra below). This is a sign of a well-developed post-training stack.” via Interconnects

Reported byInterconnects, The Batch

How 2 sources put it

• Interconnects2026-10-20issue ↗ by reflection: The long awaited open model by Reflection, making it the strongest open American model released under a permissive license (Apache 2.0). While not the smartest open model available, Beam is very impressively token efficient (second to only GPT-6-Astra below). This is a sign of a well-developed post-training stack.
• The Batch2026-10-23issue ↗ Availability: Weights available via Hugging Face, waitlist access via Reflection’s beta API

ChatGPT now lets you lock your chats

OpenAI added PIN and biometric locks for chats on Free, Go, Plus, Pro, Business, Enterprise, and Edu plans. Details how locked chats use separate memory that prevents new info from leaking into regular chat history. cited

Reported byTLDR AI, AI Breakfast, Forward Future

How 3 sources put it

• TLDR AI2026-10-21issue ↗ Locked chats are kept separate from regular chat history, pinned conversations, and chat search results.
• AI Breakfast2026-10-21issue ↗ ChatGPT added PIN-locked chats with separate memory, though not end-to-end encryption (TechCrunch)
• Forward Future2026-10-22issue ↗ : Users can secure chats with a PIN or biometric authentication, separating them from regular history and memory.

H2-X Kung Fu Legend Guardian

“Unitree unveiled the H2-X Kung Fu Legend Guardian, a security-focused humanoid variant.” via AlphaSignal

Reported byAlphaSignal, Latent Space / AINews, The Neuron

How 3 sources put it

• AlphaSignal2026-10-21issue ↗ Unitree unveiled the H2-X Kung Fu Legend Guardian, a security-focused humanoid variant.
• Latent Space / AINews2026-10-22issue ↗ …hours and framing it as a “serve and guard” platform in its X post. The post contextualizes it alongside other Chinese humanoid combat/entertainment robots, notably Engine AI’s T800, used in fighting-tournament demos per this writeup; the Reddit-hosted video itself was inaccessible due to a 403 Forbidden block. Top comments focused less on specs and more on…
• The Neuron2026-10-22issue ↗ This is awesome…and scary as well. Just watch the video on X.

A good teacher meets the student where they are

“Perplexity and NVIDIA released JOLT, a joint teacher-student self-distillation post-training method.” via AlphaSignal

Reported byAlphaSignal, TLDR AI

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ Perplexity and NVIDIA released JOLT, a joint teacher-student self-distillation post-training method.
• TLDR AI2026-10-23issue ↗ Perplexity's JOLT is a self-distillation method that jointly trains a model as both teacher and student, adapting feedback to the student's capabilities. It improved performance and training efficiency across mathematical reasoning, coding, and agentic tasks.

Meta's 816B Muse Spark Open scores 39 on Artificial Analysis, the 180B Qwen3.8-Flash-Next scores 40. Score vs size for 11 open-weight models

“(Activity: 311): The image is a scatter plot of Artificial Analysis Intelligence Index score vs. total parameter count for 11 open-weight models, highlighting that Meta Muse Spark Open (816B total / 42B active) scores…” via Latent Space / AINews

Reported byLatent Space / AINews

Discussed onr/LocalLLaMA 219 upvotes · 43 comments

AboutQwen3.8-Flash-Next

Citrini Therapeutics

Citrini Research used Anthropic's internal model to design 17 working proteins from 300 candidates for scorpion antivenom. Details the $200,000 budget, the 35-page prompt, and the decision to skip animal trials for human testing. cited

Reported byThe Rundown AI, Superhuman AI, Don't Worry About the Vase

How 3 sources put it

• The Rundown AI2026-10-21issue ↗ AI & DRUG DISCOVERY🦂 Anthropic’s Claude takes on scorpion antivenomImage source: CitriniThe Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.The details:…
• Superhuman AI2026-10-21issue ↗ Click to read the full story behind Citrini’s mission to prompt a new drug using Claude.
• Don't Worry About the Vase2026-10-22issue ↗ to try and tell Claude to do a medical breakthrough, make no mistakes. Early results are promising and they conclude the frontier has moved, but you still have to test the ideas the old fashioned way.

Muse Wonder: A Frontier Model For Personal & Professional AI Agents

Meta releases Muse Wonder and Muse Wonder Full, its most capable frontier models for personal and professional agents. Details Active Memory and Active Focus features, plus benchmark results comparing performance on law, finance, and coding tasks. cited

Reported byAI Breakfast, Simon Willison, The Rundown AI, Latent Space / AINews, The Neuron, TLDR AI, Ben's Bites, Interconnects, TheSequence, Turing Post, The Batch

How 11 sources put it

• AI Breakfast2026-10-12issue ↗ Meta released Muse Wonder and Muse Wonder Full on Monday, its most capable models to date and the first from Meta Superintelligence Labs that the company is calling frontier. Mark Zuckerberg announced them in a four-post thread on X, writing that Muse Wonder "brings us to the frontier" a little over a year after the lab was formed. The pitch is agents:…
• Simon Willison2026-10-12issue ↗ Meta AI release Muse Wonder (and Muse Wonder Full), their new "most capable models". These look like they may be Meta's first Fable class models - Artificial Analysis scores Muse Wonder (Max) as equivalent to GPT-6.1 Sol.
• The Rundown AI2026-10-13issue ↗ META🍉 Meta’s Muse Wonder lands at the frontierImage source: MetaThe Rundown: Meta Superintelligence Labs just released its long-anticipated Muse Wonder (Watermelon) and Muse Wonder Full, two powerful new AI models, with the full version landing #2 in Artificial Analysis’ intelligence rankings behind Anthropic’s Opus 5.5.The details: Gemini 4 Argon, GPT-6…
• Latent Space / AINews2026-10-13issue ↗ …rontier lab, but only today did the final ball drop: Meta’s Muse Wonder to match Spark and Glimmer, the way that OpenAI has Sol/Terra/Luna and Anthropic has Opus/Sonnet/Haiku. Interestingly, while Muse doesn’t claim a specific size, Muse Wonder Full does beat Fable and Astra on some benchmarks, and with various degrees of posttraining it is a little unclear…
• The Neuron2026-10-13issue ↗ Meta released Muse Wonder and Muse Wonder Full, which TechCrunch calls its first frontier model, meaning it's aimed at the same top tier as OpenAI, Anthropic, and Google. Both are built for agents, which are AIs that do multi-step work for you (planning a schedule, filing a claim) instead of answering once and stopping.
• TLDR AI2026-10-13issue ↗ Meta released Muse Wonder and Wonder Full, its first frontier-class models, with stronger coding, benchmark performance, and active memory for retaining and synthesizing user-provided information.
• Ben's Bites2026-10-13issue ↗ - two new ‘amazing’ models from Meta. Wonder is a Fable/Astra class model at a much cheaper price point ($4/$20 per million tokens vs $10/$50 for Fable and Astra). It’s available via Meta’s API, in Muse Code, and will soon power Muse (Meta’s personal agent) as well.
• Interconnects2026-10-17issue ↗ Just a few days before Muse Spark Open, Meta unveiled their closed frontier models, Muse Wonder and Muse Wonder Full.1 On traditional benchmarks, these are very strong models only behind Anthropic’s Opus 5.5 and similar to OpenAI’s GPT-6-Astra or Google’s Gemini-4-Argon.
• TheSequence2026-10-18issue ↗ released Muse Wonder and Muse Wonder Full, its most capable models yet, built for personal and professional agents, coding, long-running tasks, and computer use. Trained at Meta’s gigawatt-scale Prometheus data center, both take text, image, video, and audio inputs with a 1M-token context window. Muse Wonder is available through the Meta Model API and…
• Turing Post2026-10-19issue ↗ arrived for agentic work, with “Active Memory” for carrying relevant information across conversations and “Active Focus” for persisting through difficult tasks and recovering from errors. Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. The consumer Muse rollout…
• The Batch2026-10-23issue ↗ What’s new: Meta made three multimodal models available in one week, including the top-of-the-line Muse Wonder Full, the multimodal Muse Wonder, and the locally deployable Muse Spark Open. Muse Wonder is available via API and Muse Spark Open’s weights are free to download, but Muse Wonder Full is only available to Muse Code subscribers.

Today's PHP Security Release, and How We're Strengthening PHP Security

PHP released 8.5.12 and other versions to fix two critical vulnerabilities allowing unauthenticated remote code execution. Details the structural delays caused by consensus decision-making, the creation of an Ecosystem Security Team, and the specific trade-offs involved in fixing memory corruption issues. cited

Reported bySimon Willison, Latent Space / AINews

How 2 sources put it

• Simon Willison2026-10-21issue ↗ Today's PHP Security Release, and How We're Strengthening PHP Security
• Latent Space / AINews2026-10-22issue ↗ All high profile cyber disclosures are at least a little bit of marketing, and the degree to which this was a severely alarming vulnerability on the order of famous ones like log4shell and Heartbleed is up to some debate, but surely the coordinated response across the PHP Ecosystem is the developer community responding in its most ideal form.

Muse Wonder Model Card

“How it works: Meta released a model card for Muse Wonder and Wonder Full, but it reveals very little about the model’s architecture or how it was trained.” via The Batch

Reported byDon't Worry About the Vase, The Batch

How 2 sources put it

• Don't Worry About the Vase2026-10-22issue ↗ The Muse Wonder model card is here. Is it a good model, sir?
• The Batch2026-10-23issue ↗ How it works: Meta released a model card for Muse Wonder and Wonder Full, but it reveals very little about the model’s architecture or how it was trained.

Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages

Mistral releases Voxtral 3, a voice agent SDK and speech models supporting 50+ languages with sub-500ms latency. Includes pricing tables for four models ranging from $0.0035/min to $13/1M chars and details on Pro Voice adaptation using 30 minutes of audio. cited

Reported byForward Future, TLDR AI, AI Breakfast

How 3 sources put it

• Forward Future2026-10-21issue ↗ : Mistral released Voxtral 3, a voice-agent toolkit with speech models for more than 50 languages.
• TLDR AI2026-10-21issue ↗ Voxtral 3 offers advanced speech-to-text and text-to-speech capabilities across 50+ languages. Featuring models like Small Transcribe 3 for multilingual transcription and Mini TTS 3 for natural emotional speech, it includes a voice agent SDK combining real-time listening, reasoning,…
• AI Breakfast2026-10-21issue ↗ yesterday, a full voice agent stack rather than another transcription model. Four models ship together: Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.0040 with diarization and context biasing, a realtime transcriber at $0.008 with sub-500ms latency, and Voxtral Speech 3 for text to speech at $13…

DeepSeek-V4.1-Flash-1017 (Hugging Face Repo)

“The brain first. This is a minor update to DeepSeek's cheaper "Flash" tier, and the weights are free to download under an MIT license, so anyone can run it or build a paid service on it. On DeepSWE v1.1, a coding-agent test, DeepSeek reports 77.0, versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol.” via The Neuron

Reported byThe Neuron, TLDR AI, AI Breakfast, Interconnects

Discussed onHacker News 959 pts · 725 comments

AboutDeepSeek-V4.1-Flash-1017, DeepSeek-V4.1-Flash

How 4 sources put it

• The Neuron2026-10-19issue ↗ The brain first. This is a minor update to DeepSeek's cheaper "Flash" tier, and the weights are free to download under an MIT license, so anyone can run it or build a paid service on it. On DeepSWE v1.1, a coding-agent test, DeepSeek reports 77.0, versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol.
• TLDR AI2026-10-19issue ↗ DeepSeek-V4.1-Flash-1017 delivers substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. It also reduces task execution time and output length, advancing the performance-cost Pareto frontier across multiple agent harnesses. DeepSeek-V4.1-Flash combines cross-layer KV cache reuse…
• AI Breakfast2026-10-19issue ↗ landed Saturday under an MIT license with a 1M-token recommended context window. DeepSeek calls it a minor update, but the jumps are not minor: Terminal-Bench 4.0 goes from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, and HLE without tools from 36.8 to 42.6. The card also claims shorter outputs and faster task execution. Read the…
• Interconnects2026-10-20issue ↗ by deepseek-ai: Since the latest Artifacts, the whale has not released one, but two new model iterations with a new size (552B-A16B) which also support image inputs. The model has gained a lot of (well-deserved) attention for its performance at a small price. It also carries the MIT license, continuing the trend established by…

MiniMax M3.1 Flash: Redefining Agent Performance and Cost

MiniMax launches M3.1 Flash, an agent model that reduces usage costs by 70% while scoring 76.2 on BankerToolBench. Details flat pricing of $0.15 input and $0.60 output per 1M tokens alongside 200+ TPS output speeds. cited

Reported byThe Rundown AI, The Neuron, TLDR AI

AboutMiniMax-M3.1-Flash

How 3 sources put it

• The Rundown AI2026-10-20issue ↗ Chinese AI lab MiniMax launched M3.1 Flash, a new model previously tested as “Space Bunny”, scoring well on coding and computer use at an extremely low price.
• The Neuron2026-10-20issue ↗ is a cheaper, faster model for running AI agents, with open weights planned for October 23, though one test found it tends to make things up when it's wrong.
• TLDR AI2026-10-20issue ↗ MiniMax M3.1 Flash delivers a generational leap in agentic tasks, coding, Computer Use, and cybersecurity. It makes online inference nearly three times faster and reduces usage costs by 70%. The model is now available through the MiniMax API, M Plan, and MiniMax Code. MiniMax plans to…

DeepSeek raises $15 billion

“AI lab DeepSeek reportedly raised $15B in one of China’s biggest rounds, just two days after the U.S. ordered federal agencies to remove its AI over security concerns.” via The Rundown AI

Reported byDon't Worry About the Vase, The Rundown AI

How 2 sources put it

• Don't Worry About the Vase2026-10-22issue ↗ …timent about AI stocks did not go up hearing that sentence. DeepSeek raises $15 billion. Microsoft cancels UK data center deal with Nscale, in part over power concerns.Anthropic in talks for a $45 billion cloud computing deal with Nscale. A classic your loss is our gain scenario, perhaps. Nscale is trying to IPO and lost the Microsoft project, so Anthropic…
• The Rundown AI2026-10-23issue ↗ AI lab DeepSeek reportedly raised $15B in one of China’s biggest rounds, just two days after the U.S. ordered federal agencies to remove its AI over security concerns.

Contained: Oct 21, 2026 LiteLLM release v1.107.0 Compromised

LiteLLM release v1.107.0 contained malware that stole credentials and executed remote code for about two hours. Details the attack chain involving compromised GitHub CLI tokens and lists specific steps to check for infection and rotate keys. cited

Reported byThe Neuron, TLDR AI

How 2 sources put it

• The Neuron2026-10-22issue ↗ , a popular tool for connecting apps to AI models, confirmed a tampered version was live for two hours on Oct 21; SafeDep says it stole credentials.
• TLDR AI2026-10-22issue ↗ A malicious LiteLLM package was published to PyPI and was live for approximately 2 hours before it was pulled.

Computer use on the phone itself: a 2B model that operates Android apps

Opposable runs a 2B model on Android phones to tap and type in apps without a server. Includes benchmarks showing 7.3s steps on M2 CPUs and 236ms encoding on Galaxy S25 NPUs. cited

Reported byThe Neuron, TLDR AI

How 2 sources put it

• The Neuron2026-10-22issue ↗ lets an AI agent tap and type through your Android apps, with a small on-phone model as backup (weaker on unfamiliar apps, per its makers).
• TLDR AI2026-10-22issue ↗ GUI-Owl 1.5 is a 2-billion-parameter screen model that can read an Android screen through the accessibility service and tap and type in any other app.

I pinged Codex yesterday ‘check if any of my work uses PHP or Railway, assess any risks and address them.’ OpenAI…

“Big security issues yesterday with PHP (a programming language that a lot of the internet runs on) and Railway (a hosting platform). I pinged Codex yesterday ‘check if any of my work uses PHP or Railway, assess any risks and address them.’ OpenAI helped researchers find these vulnerabilities.” via Ben's Bites

Reported byLatent Space / AINews, Ben's Bites

How 2 sources put it

• Latent Space / AINews2026-10-22issue ↗ OpenAI describes approximately two months of coordinated remediation before disclosure. It says the patch required a major PHP-engine change and coordination among maintainers, researchers, security providers, and governments. A participating researcher confirms work with the PHP Foundation to support and harden the patch. @OpenAI, @itszn13
• Ben's Bites2026-10-22issue ↗ Big security issues yesterday with PHP (a programming language that a lot of the internet runs on) and Railway (a hosting platform). I pinged Codex yesterday ‘check if any of my work uses PHP or Railway, assess any risks and address them.’ OpenAI helped researchers find these vulnerabilities.

openai/openai-cli: Official CLI for the OpenAI API

The official OpenAI CLI lets users call the REST API from the terminal to generate images, upload files, and submit batches. Includes installation instructions for macOS, Linux, and Windows, plus examples for file management, batch processing, and admin usage queries. cited

Reported byTLDR AI

Discussed onHacker News 243 pts · 111 comments

How 1 source put it

• TLDR AI2026-10-22issue ↗ This repository contains the official CLI for the OpenAI API.

Daybreak for Frontline Defenders

“OpenAI pledged $1B in subsidized Daybreak access for under-resourced defenders.” via AlphaSignal

Reported byAlphaSignal, Latent Space / AINews

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ OpenAI pledged $1B in subsidized Daybreak access for under-resourced defenders.
• Latent Space / AINews2026-10-22issue ↗ Almost 2 months ago, OpenAI introduced Daybreak for Frontline Defenders, a $1B subsidy for access to Daybreak cyber models and products. That funding came into use almost immediately, as major vulnerabilities in PHP were independently discovered by two parties: Ryan Austin, a researcher at the security firm Calif, as part of the OpenAI “Patch the Planet”…

Redwood embedded assessment at OpenAI: Two auditors from METR and Redwood Research spent three weeks on-site evaluating…

“/Redwood embedded assessment at OpenAI: Two auditors from METR and Redwood Research spent three weeks on-site evaluating OpenAI's internal misalignment monitoring for agents (METR, OpenAI).” via Latent Space / AINews

Reported byLatent Space / AINews, AI Breakfast

How 2 sources put it

• Latent Space / AINews2026-10-23issue ↗ /Redwood embedded assessment at OpenAI: Two auditors from METR and Redwood Research spent three weeks on-site evaluating OpenAI's internal misalignment monitoring for agents (METR, OpenAI).
• AI Breakfast2026-10-23issue ↗ …e tool, it is in who can ship a patch to 22 million servers.Read more.Claude Duet is an agent that messages you firstAnthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than…

While Amazon blocks AI agents, Walmart rolls out the welcome mat

Walmart launches an Agent Gateway MCP server to let approved AI agents search inventory and buy items. Includes upcoming Agent Profiles for spending limits and transaction controls, plus terms ensuring returns apply if agents misbehave. cited

Reported byForward Future, AI Breakfast

How 2 sources put it

• Forward Future2026-10-23issue ↗ : Its new Agent Gateway MCP server lets approved AI agents search inventory and buy for customers.
• AI Breakfast2026-10-23issue ↗ Walmart opened an MCP server and agent profiles for approved AI shopping agents while Amazon blocks them (TechCrunch)

Microsoft’s Mustafa Suleyman says its AI outperforms doctors on hard cases as he chases ‘medical superintelligence’

Microsoft's MAI-DxO system correctly diagnosed complex medical cases 83% of the time in a simulation. Details how the system cut costs by about $3,000 per case while raising accuracy by only about four points. cited

Reported byThe Neuron, AI Breakfast

How 2 sources put it

• The Neuron2026-10-23issue ↗ Now the fine print, which is where our eyebrows went up. TechCrunch's write-up flags a few things worth knowing before you picture a robot doctor:
• AI Breakfast2026-10-23issue ↗ Microsoft's MAI-DxO scored 83% on complex NEJM cases against 36% for generalist doctors working alone (TechCrunch)

DeepSeek V4.1 Flash: Large-Scale Inference Deployment

“DeepInfer combines distributed inference, caching, load balancing, and speculative decoding to improve efficiency. DeepSeek processes 155 trillion tokens daily per cluster and earns $38 in revenue per GPU-hour while open-sourcing much of its inference framework.” via TLDR AI

Reported byLatent Space / AINews, TLDR AI

AboutDeepSeek-V4.1-Flash

How 2 sources put it

• Latent Space / AINews2026-10-21issue ↗ … commenter referenced DeepSeek’s Zhihu post claiming roughly $38 per GPU-hour revenue/throughput economics, linking the original post at zhuanlan.zhihu.com/p/2095560295930508953. The thread raised whether similar per-GPU economics could be used to estimate inference revenue for competitors like OpenAI or Anthropic, though no concrete derivation was…
• TLDR AI2026-10-21issue ↗ DeepInfer combines distributed inference, caching, load balancing, and speculative decoding to improve efficiency. DeepSeek processes 155 trillion tokens daily per cluster and earns $38 in revenue per GPU-hour while open-sourcing much of its inference framework.

How employment is changing at Anthropic

“Anthropic reported a decline in junior technical hiring, with entry-level workers falling from 27% to 9% of technical hires since early 2025. Employees reported productivity gains from Claude, but increased AI output also shifted work toward supervising agents and made human attention a growing…” via TLDR AI

Reported byTLDR AI, AI Breakfast, Don't Worry About the Vase

Discussed onHacker News 162 pts · 256 comments

How 3 sources put it

• TLDR AI2026-10-19issue ↗ Anthropic reported a decline in junior technical hiring, with entry-level workers falling from 27% to 9% of technical hires since early 2025. Employees reported productivity gains from Claude, but increased AI output also shifted work toward supervising agents and made human attention a growing…
• AI Breakfast2026-10-19issue ↗ Anthropic says agent hours ran about 6x employee hours by August, and junior technical hires fell from 27% to 9% (Anthropic)
• Don't Worry About the Vase2026-10-22issue ↗ Anthropic is hiring for a smaller share of junior positions and AI agent hours outnumber employee hours six to one. Human attention is a bottleneck for AI outputs and human productivity is way up. That’s the central finding of a new Anthropic report on what it is like to work there.

Anthropic reports that 17 of 211 Claude-designed protein binders against scorpion neurotoxins met its experimental…

“Scorpion-toxin binders: Anthropic and Citrini report 17 binders from 211 tested protein designs, following substantial computational work by an internal Claude research model. All tested designs and laboratory data are being released (announcement, data).” via Latent Space / AINews

Reported byLatent Space / AINews, Ben's Bites

How 2 sources put it

• Latent Space / AINews2026-10-21issue ↗ Scorpion-toxin binders: Anthropic and Citrini report 17 binders from 211 tested protein designs, following substantial computational work by an internal Claude research model. All tested designs and laboratory data are being released (announcement, data).
• Ben's Bites2026-10-22issue ↗ Anthropic reports that 17 of 211 Claude-designed protein binders against scorpion neurotoxins met its experimental binding criteria.

Nscale and Anthropic are reportedly in talks for yet another $45B compute deal

Bloomberg reports Anthropic and Nscale are in talks for a second $45 billion compute deal. Includes details that both companies are currently pursuing IPOs while investors worry about spending too much money. cited

Reported byThe Neuron, TLDR AI, AI Breakfast

How 3 sources put it

• The Neuron2026-10-20issue ↗ Anthropic and Nscale neared a second $45B deal for computing capacity, Bloomberg reported, a little over a month after the first.
• TLDR AI2026-10-20issue ↗ Anthropic and Nscale may be on the verge of announcing another $45 billion compute deal. The two companies announced a $45 billion compute deal only a little over a month ago. Both companies are currently pursuing IPOs. Anthropic is still expected to go public this…
• AI Breakfast2026-10-21issue ↗ Anthropic is reportedly in talks for a second $45B compute deal with Nscale (Bloomberg)

Muse Wonder: A Frontier Model For Personal & Professional AI Agents

Meta releases Muse Wonder and Muse Wonder Full, its most capable frontier models for personal and professional agents. Details Active Memory and Active Focus features, plus benchmark results comparing performance on law, finance, and coding tasks. cited

Reported byAI Breakfast, Simon Willison, The Rundown AI, Latent Space / AINews, The Neuron, TLDR AI, Ben's Bites, Interconnects, TheSequence, Turing Post, The Batch

How 11 sources put it

• AI Breakfast2026-10-12issue ↗ Meta released Muse Wonder and Muse Wonder Full on Monday, its most capable models to date and the first from Meta Superintelligence Labs that the company is calling frontier. Mark Zuckerberg announced them in a four-post thread on X, writing that Muse Wonder "brings us to the frontier" a little over a year after the lab was formed. The pitch is agents:…
• Simon Willison2026-10-12issue ↗ Meta AI release Muse Wonder (and Muse Wonder Full), their new "most capable models". These look like they may be Meta's first Fable class models - Artificial Analysis scores Muse Wonder (Max) as equivalent to GPT-6.1 Sol.
• The Rundown AI2026-10-13issue ↗ META🍉 Meta’s Muse Wonder lands at the frontierImage source: MetaThe Rundown: Meta Superintelligence Labs just released its long-anticipated Muse Wonder (Watermelon) and Muse Wonder Full, two powerful new AI models, with the full version landing #2 in Artificial Analysis’ intelligence rankings behind Anthropic’s Opus 5.5.The details: Gemini 4 Argon, GPT-6…
• Latent Space / AINews2026-10-13issue ↗ …rontier lab, but only today did the final ball drop: Meta’s Muse Wonder to match Spark and Glimmer, the way that OpenAI has Sol/Terra/Luna and Anthropic has Opus/Sonnet/Haiku. Interestingly, while Muse doesn’t claim a specific size, Muse Wonder Full does beat Fable and Astra on some benchmarks, and with various degrees of posttraining it is a little unclear…
• The Neuron2026-10-13issue ↗ Meta released Muse Wonder and Muse Wonder Full, which TechCrunch calls its first frontier model, meaning it's aimed at the same top tier as OpenAI, Anthropic, and Google. Both are built for agents, which are AIs that do multi-step work for you (planning a schedule, filing a claim) instead of answering once and stopping.
• TLDR AI2026-10-13issue ↗ Meta released Muse Wonder and Wonder Full, its first frontier-class models, with stronger coding, benchmark performance, and active memory for retaining and synthesizing user-provided information.
• Ben's Bites2026-10-13issue ↗ - two new ‘amazing’ models from Meta. Wonder is a Fable/Astra class model at a much cheaper price point ($4/$20 per million tokens vs $10/$50 for Fable and Astra). It’s available via Meta’s API, in Muse Code, and will soon power Muse (Meta’s personal agent) as well.
• Interconnects2026-10-17issue ↗ Just a few days before Muse Spark Open, Meta unveiled their closed frontier models, Muse Wonder and Muse Wonder Full.1 On traditional benchmarks, these are very strong models only behind Anthropic’s Opus 5.5 and similar to OpenAI’s GPT-6-Astra or Google’s Gemini-4-Argon.
• TheSequence2026-10-18issue ↗ released Muse Wonder and Muse Wonder Full, its most capable models yet, built for personal and professional agents, coding, long-running tasks, and computer use. Trained at Meta’s gigawatt-scale Prometheus data center, both take text, image, video, and audio inputs with a 1M-token context window. Muse Wonder is available through the Meta Model API and…
• Turing Post2026-10-19issue ↗ arrived for agentic work, with “Active Memory” for carrying relevant information across conversations and “Active Focus” for persisting through difficult tasks and recovering from errors. Wonder is available through the Meta Model API and Muse Code; Wonder Full is available to Muse Code subscribers. The consumer Muse rollout…
• The Batch2026-10-23issue ↗ What’s new: Meta made three multimodal models available in one week, including the top-of-the-line Muse Wonder Full, the multimodal Muse Wonder, and the locally deployable Muse Spark Open. Muse Wonder is available via API and Muse Spark Open’s weights are free to download, but Muse Wonder Full is only available to Muse Code subscribers.

Introducing Mistral Large 4

Mistral launches a public preview of its 1 trillion-parameter, natively multimodal model Le Chonk. Includes cybersecurity benchmarks where it scores 82% on vulnerability patching and beats GPT-6-Astra on legal tasks. cited

Reported bySimon Willison, The Neuron, The Rundown AI, TLDR AI, AI Breakfast, Last Week in AI, TheSequence, Import AI, Turing Post, The Batch, Interconnects, AlphaSignal

Discussed onHacker News 522 pts · 5 comments

How 12 sources put it

• Simon Willison2026-10-06issue ↗ Mistral are back in the game. Today they're releasing a preview of Mistral Large 4, a 1 trillion parameter, 49 billion active parameter model trained on their own cluster of 3,800 NVIDIA Grace Blackwell GPUs. The preview is available via their API. They promise to release the open weights model at the "end of this…
• The Neuron2026-10-07issue ↗ gives you a frontier-class multimodal model across 160+ languages, with downloadable weights due by the end of October for teams that want to run or tune it themselves. Viva la French AI!
• The Rundown AI2026-10-07issue ↗ MISTRAL🐈 Mistral's ‘Le Chonk’ joins the West’s open-model surgeImage source: MistralThe Rundown: France’s Mistral AI just introduced Large 4, a 1T-parameter open model nicknamed “Le Chonk,” which the company says leads every non-Chinese open system by a “substantial margin.” The details: China's Kimi K3 still tops Mistral’s coding chart at 68%, but Le…
• TLDR AI2026-10-07issue ↗ Mistral Large 4 is a one-trillion-parameter multimodal model positioned as a European alternative to leading closed and open AI systems.
• AI Breakfast2026-10-07issue ↗ landed yesterday in API preview, and the internal nickname tells you the shape of it: le Chonk. It is a 1 trillion-parameter natively multimodal mixture of experts with 49 billion active parameters, a hybrid instruct-and-reasoning model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, across more…
• Last Week in AI2026-10-09issue ↗ …ishes hundreds of math proofs from unreleased frontier modelSources:OpenAI drops another batch of mathematical breakthroughsSharing AI progress in mathematicsOpenAI Releases Findings on 377 Math Problems, Further Roiling FieldAll the drama around AI’s takeover of mathematicsSourceOpenAI published a large batch of mathematical results on October 6, 2026,…
• TheSequence2026-10-11issue ↗ launched a preview of Mistral Large 4, a 1-trillion-parameter natively multimodal open-weight model with 49B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs. The preview API is live now, and weights follow by the end of the month. Mistral reports strong coding and agentic results (61.7% on DeepSWE v1.1) and the top open-weight score…
• Import AI2026-10-12issue ↗ …peting with China is going to be very difficult. Read more: Le chonk, Introducing Mistral Large 4 (Mistral.ai).***Two USG projects show how DOE might help the government with AI:…The US government is starting to knit its lab supercomputers together…The US government has recently conducted two AI training projects that utilize multiple supercomputers at…
• Turing Post2026-10-12issue ↗ , enters public API preview. The natively multimodal mixture-of-experts model has one trillion parameters, with 52 billion active, and emphasizes coding, cybersecurity, finance, and law. Mistral says it trained the model on 3,800 Grace Blackwell GPUs in its own European datacenters. Downloadable weights are…
• The Batch2026-10-16issue ↗ What’s new: Mistral announced API preview access to Mistral Large 4, nicknamed “le Chonk.” Mistral says it will release the model’s weights by the end of this month. Meanwhile, selected companies, partners, and government agencies have access to a version with limited safeguards and extra cybersecurity capabilities.
• Interconnects2026-10-17issue ↗ …-weight models built and released by Chinese companies better than the best Western models. At the time of writing this, Mistral has also announced a strong model, Mistral Large 4, with an index score of 38 (see below), but its weights are not yet released. There are rumors of a few more strong American open-weight models coming from other organizations…
• AlphaSignal2026-10-21issue ↗ : 1T params, 49B active, trained on 3,800 Grace Blackwell GPUs.

OpenAI models are coming to Google Cloud

“OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden.” via AlphaSignal

Reported byAlphaSignal, The Neuron, The Rundown AI, TLDR AI, Don't Worry About the Vase, AI Breakfast

Discussed onHacker News 325 pts · 148 comments

How 6 sources put it

• AlphaSignal2026-10-21issue ↗ OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden.
• The Neuron2026-10-22issue ↗ said its models and Codex are coming to Google Cloud, with GPT-Astra, GPT-6.1 Sol, and GPT-6 Luna in private preview for select enterprise customers.
• The Rundown AI2026-10-22issue ↗ OpenAI struck a deal to sell its models and Codex through Google Cloud, with the models also landing inside Google’s own Antigravity coding app and Gemini Agent soon.
• TLDR AI2026-10-22issue ↗ OpenAI's frontier models and Codex agentic workflows are coming to Google Cloud's Model Garden. The integration will give customers native access to these models to power complex agentic workflows at scale. It will pair OpenAI's frontier models with Google Cloud's built-in access controls, centralized…
• Don't Worry About the Vase2026-10-22issue ↗ …nAI expands access to GPT-Rosalind for scientific research. OpenAI models are coming to Google Cloud.OpenAI offers us GPT-Live-1.1 including new voices in various languages.Muse Spark Open exists, as in a version of Muse Spark is now an open model. They call it ‘the most capable US open-weight model to date.’ That seems likely to be true. It scores a 39 on…
• AI Breakfast2026-10-23issue ↗ OpenAI's frontier models and Codex are coming to Google Cloud's Model Garden in the coming weeks (OpenAI)

Introducing Claude Duet

Anthropic launches Claude Duet, a proactive agent for Pro and Max users that works on a separate cloud computer. Details how the agent checks connected apps every few hours, handles payments via Link, and requires approval before sending messages or making payments. cited

Reported byAlphaSignal, The Neuron, The Rundown AI, TLDR AI, Latent Space / AINews, AI Breakfast

How 6 sources put it

• AlphaSignal2026-10-21issue ↗ Anthropic launched Claude Duet, a proactive agent app that monitors your accounts and messages you when attention is needed.
• The Neuron2026-10-22issue ↗ is a new app where Claude acts as a proactive agent. It sends you a morning brief, checks your connected apps every few hours, and messages you only when something actually needs you. It's a research preview for US Pro and Max subscribers (by waitlist), it's built around Claude Opus 5.5, and it uses your plan's existing usage limits.
• The Rundown AI2026-10-22issue ↗ ANTHROPIC🤖 Anthropic joins the personal agent game with DuetImage source: Anthropic / The RundownThe Rundown: Anthropic just introduced Claude Duet, an Opus 5.5-powered personal agent standalone app that can take proactive actions, work 24/7 (even when closed), and carry out tasks on its own cloud computer. The details: Example tasks include morning briefs,…
• TLDR AI2026-10-22issue ↗ Anthropic has unveiled Claude Duet, a proactive AI agent powered by Opus 5.5 that monitors connected apps, prioritizes tasks, and works autonomously while requesting approval before certain actions. The research preview initially targeted US Pro and Max subscribers through a waitlist.
• Latent Space / AINews2026-10-23issue ↗ Anthropic unveils Claude Duet, a proactive AI agent powered by Opus 5.5. It works while you're away, monitors your apps, handles complex tasks using its own cloud computer, and alerts you when needed. It supports 6,000+ integrations and remembers your context. (Activity: 1287): Anthropic announced Claude Duet, a research-preview proactive agent for Pro/Max…
• AI Breakfast2026-10-23issue ↗ Anthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than waiting to be asked. There is a daily morning brief, and you set what it should surface and how often. It runs on its own…

Investing in the public benefit

Anthropic commits 5% of shares to a new Public Benefit Fund, adding 2% for every 50% valuation increase. Details how the LLC avoids lobbying, funds labor impact studies, and combines with founder pledges to reach 20% total public good equity. cited

Reported byThe Rundown AI, The Neuron, Superhuman AI, TLDR AI, AI Breakfast, Latent Space / AINews, Forward Future

How 7 sources put it

• The Rundown AI2026-10-19issue ↗ ANTHROPIC💰 Anthropic designs a pay package for the publicImage source: AnthropicThe Rundown: Anthropic just introduced the Anthropic Public Benefit Fund, a limited liability company (LLC) that would get 5% of the company’s equity if shareholders sign off (about $48B at its May valuation), plus more as Anthropic’s value climbs.The details: As a company-owned…
• The Neuron2026-10-19issue ↗ Anthropic pledged 5% of its shares to a new Public Benefit Fund, pending stockholder approval, with more possible as its value grows.
• Superhuman AI2026-10-19issue ↗ …t Peter McCrory says employee-directed agents worked 6x as many hours as their human users in August, driving productivity across the company. In response, the lab has slowed hiring for junior technical roles. Separately, Anthropic committed 5% of its shares to a public benefit fund designed to channel growth into public giving, pending shareholder…
• TLDR AI2026-10-19issue ↗ Anthropic announced plans to allocate an initial 5% of its equity to a new Public Benefit Fund, with additional contributions tied to valuation growth. The fund will support initiatives such as addressing AI's economic and labor impacts.
• AI Breakfast2026-10-19issue ↗ Anthropic announced on Saturday that it is committing 5% of its shares to the Anthropic Public Benefit Fund, subject to stockholder approval. Elon replied to the post by stating that he plans on giving every child in America a share of SpaceX.
• Latent Space / AINews2026-10-20issue ↗ … in its valuation (Activity: 2365): Anthropic announced the Anthropic Public Benefit Fund, a wholly owned LLC that would initially receive 5% of Anthropic shares pending stockholder approval, with the board able to add another 2% for each 50% valuation increase over a baseline. Anthropic says that, combined with founder/employee equity donation commitments,…
• Forward Future2026-10-20issue ↗ Anthropic plans to place 5% of its shares in a new Public Benefit Fund, subject to shareholder approval, and its board may add another 2% each time the company’s value rises 50% above a set baseline. The fund could support nonprofits, direct payments and programs addressing AI’s effect on workers. Saudi Ties in OpenAI…

Introducing Claude Haiku 5.5

Anthropic launches Claude Haiku 5.5, a model that costs around 75% less to run than Haiku 4.5. Includes benchmark tables showing Haiku 5.5 scores of 46.4% on FrontierCode 1.1 and pricing details for prompts up to 100,000 tokens. cited

Reported bySimon Willison, The Rundown AI, The Neuron, Forward Future, TLDR AI, Ben's Bites, Don't Worry About the Vase, Latent Space / AINews, Last Week in AI, AI Breakfast, TheSequence, Data Points, Turing Post, The Batch

Discussed onHacker News 1,049 pts · 489 comments

How 14 sources put it

• Simon Willison2026-10-07issue ↗ As previously promised, here's Anthropic's new fast, low cost model: Introducing Claude Haiku 5.5.
• The Rundown AI2026-10-08issue ↗ Anthropic launched Claude Haiku 5.5, a budget model starting at the same $0.10 per million input tokens as OpenAI's GPT-6 Luna while passing it on multiple benchmarks.
• The Neuron2026-10-08issue ↗ Anthropic's smallest AI model got a glow-up on Wednesday. Claude Haiku 5.5 costs $0.10 per million input tokens (a token is a small chunk of text, about three-quarters of a word), which VentureBeat noted matches OpenAI's GPT-6 Luna to the penny.
• Forward Future2026-10-08issue ↗ …d generated interfaces more useful than a standard replies. Claude Haiku 5.5 Targets Fast, High-Volume Work Anthropic released Claude Haiku 5.5, which it calls its fastest and most capable small model. It is designed for high-volume jobs such as summarization, browser use and subagents, making price, latency and reliability the measures to watch against…
• TLDR AI2026-10-08issue ↗ Claude Haiku 5.5 is optimized for high-volume, cost-sensitive tasks, offering lower operating costs and faster performance for applications like summaries and customer support. The model provides significant pricing advantages, being 75% cheaper than Haiku 4.5 for tasks involving prompts up to 100,000 tokens.…
• Ben's Bites2026-10-08issue ↗ costs ~75% less to run than Haiku 4.5. Anthropic pitches it as a subagent for Opus and Sonnet. Simon has great notes on the model. Max and Team plans also now get monthly API credits to build your own apps and agents. (tweet). Anthropic also expanded their Claude startup program, and people were getting accepted in minutes.
• Don't Worry About the Vase2026-10-08issue ↗ Claude Haiku 5.5 exists. It is a tenth (!) the price of Haiku 4.5, at $0.10/$0.50 for prompts up to 100k, with cache writes and reads at $0.125/$0.01. Prices rise 5x for longer prompts, so if you use API billing and you set /model haiku you also want /autocompact 100k.
• Latent Space / AINews2026-10-08issue ↗ Introducing Claude Haiku 5.5: the cheapest, fastest, and most capable small model we’ve ever released (Activity: 2979): Anthropic announced Claude Haiku 5.5, positioning it as its cheapest/fastest small Claude model for high-volume tasks such as summarization, classification, live support, browser use, and as a coding sub-agent alongside Opus/Sonnet 5.5.…
• Last Week in AI2026-10-09issue ↗ Anthropic Releases Claude Haiku 5.5: A Small Model With 1M Context Priced at $0.10 per Million Input TokensSource
• AI Breakfast2026-10-09issue ↗ is the price story of the week. It runs $0.10 per million input tokens and $0.50 output on prompts up to 100k, rising to $0.50 and $2.50 above that. Haiku 4.5 was $1 and $5, so Anthropic's "around 75% less" headline is an average: the footnote says 90% cheaper under 100k tokens and 50% cheaper above it. The jumps are not small either.…
• TheSequence2026-10-11issue ↗ released Claude Haiku 5.5, its cheapest and fastest model, at $0.10 per million input tokens and $0.50 per million output tokens, 90% below Haiku 4.5. It’s the first Haiku with an adjustable effort setting and scores 72.4% on OSWorld 2.1 and 39.2% on Terminal-Bench 4.0, ahead of OpenAI’s GPT-6 Luna on both.
• Data Points2026-10-12issue ↗ released Claude Haiku 5.5, its small, fast model aimed at high-volume tasks like summarization, classification, and database queries. It’s the first Haiku model with an adjustable effort setting to trade off cost against intelligence. For prompts up to 100,000 tokens, input costs drop to $0.10 per million tokens and output to $0.50 per million,…
• Turing Post2026-10-12issue ↗ arrives for high-volume work such as classification, summaries, database queries, and coding subagents. It is the first Haiku with adjustable effort. Anthropic reports average running costs around 75% below Haiku 4.5 and also halves Sonnet 5.5’s cache-read price.
• The Batch2026-10-16issue ↗ What’s new: Claude Haiku 5.5 replaces the year-old Claude Haiku 4.5. Independent tests show marked improvement, even with a steep price. Two weeks earlier, Anthropic launched Claude Sonnet 5.5, which the company says is best suited to routine coding and office work, including fixing bugs and producing documents, slides, and spreadsheets. It replaces Claude…

OpenAI’s model helps catch critical PHP flaws

“OpenAI coordinated with PHP maintainers and security partners to patch critical vulnerabilities discovered using GPT-5.6 Sol and independently by Wiz. The company also released a PHP vulnerability checker and proposed a framework for coordinating responses to AI-discovered security flaws.” via TLDR AI

Reported byAlphaSignal, The Rundown AI, Superhuman AI, TLDR AI, AI Breakfast

Discussed onHacker News 266 pts · 183 comments

How 5 sources put it

• AlphaSignal2026-10-21issue ↗ … attack surface. Wiz’s analysis cites estimates that PHP runs on roughly 70% of websites with a known server-side language, including sites built with WordPress, Drupal, Joomla and MediaWiki. OpenAI’s disclosure describes PHP2Shell as an early example of a frontier model helping researchers find an internet-scale vulnerability before widespread…
• The Rundown AI2026-10-22issue ↗ OPENAI🚨 OpenAI’s model helps catch critical PHP flawsImage source: X / The RundownThe Rundown: PHP, the coding language behind ~70% of websites, just pushed patches for two critical flaws caught with OpenAI’s GPT-5.6 Sol, with the National Security Agency urging anyone who can’t patch right away to pull systems offline.The details: Security firm Calif found…
• Superhuman AI2026-10-22issue ↗ … The NSA recommends pulling any systems that can’t be patched right away, causing many sites to go dark across the internet. In response, OpenAI announced a coordinated global awareness effort, along with a new Codex skill that helps companies determine if they’re at risk. Stories of other attacks are also developing, but we were not able to verify them…
• TLDR AI2026-10-22issue ↗ OpenAI coordinated with PHP maintainers and security partners to patch critical vulnerabilities discovered using GPT-5.6 Sol and independently by Wiz. The company also released a PHP vulnerability checker and proposed a framework for coordinating responses to AI-discovered security flaws.
• AI Breakfast2026-10-23issue ↗ …e tool, it is in who can ship a patch to 22 million servers.Read more.Claude Duet is an agent that messages you firstAnthropic opened a research preview of Claude Duet on Wednesday, a separate app built on Claude Opus 5.5 that inverts the chat model: it checks your connected accounts every few hours and pings you when something needs you, rather than…

Introducing Muse Spark Open, The Most Capable US Open‑Weight Model to Date

Meta releases Muse Spark Open, an 816B-parameter multimodal model designed for coding and agentic work. Details how two Apple M5 Ultra Mac Studios run the model locally via MLX and lists hosted inference partners. cited

Reported byThe Rundown AI, The Neuron, Superhuman AI, TLDR AI, AI Breakfast, Interconnects, TheSequence, Don't Worry About the Vase

How 8 sources put it

• The Rundown AI2026-10-16issue ↗ META🧠 Meta gives the West its best open-weight modelImage source: Mark Zuckerberg on XThe Rundown: Meta just released Muse Spark Open, an open-weight, 816B-parameter multimodal model built for coding and agentic work, and worked with Apple so that it could run locally on Mac Studios — a first for a model in this size class, Meta says.The details: Two M5…
• The Neuron2026-10-16issue ↗ released Muse Spark Open, an open-weight coding and agent model with 816B total parameters that Meta says runs on two Apple M5 Ultra Mac Studios, under a license with limited commercial terms.
• Superhuman AI2026-10-16issue ↗ …e open model to Apple’s Macs: Meta released the weights for Muse Spark Open, one of the strongest American-made open models currently available. The model is optimized to run on Apple silicon, but there’s one catch: you’ll need at least two top-spec M5 Ultra Mac Studios. The release signals a thawing relationship between the two rivals, with Meta’s Alexandr…
• TLDR AI2026-10-16issue ↗ Muse Spark Open is an 816B-parameter multimodal open-weight model with 42B active parameters and a 1M-token context window. It was designed for coding and long-horizon agentic tasks and can run locally across two Apple M5 Ultra Mac Studios using MLX.
• AI Breakfast2026-10-16issue ↗ Meta Superintelligence Labs released Muse Spark Open on Thursday and called it the most capable US open-weight model to date. It is natively multimodal, 816B total parameters with 42B active, and carries a 1M-token context window, aimed at coding and long-horizon agentic work. Weights are up on Hugging Face, Microsoft Foundry, Google Cloud Model Garden and…
• Interconnects2026-10-17issue ↗ …toward the frontier. Meta took the lead of this charge with Muse Spark Open, which is now the strongest American-made open model by quite some margin, but the addition of Reflection’s first model Beam is also relevant. Without these two models, there would be 20+ open-weight models built and released by Chinese companies better than the best Western models.…
• TheSequence2026-10-18issue ↗ released the weights for Muse Spark Open, a natively multimodal model with 816B total parameters (42B active) and a 1M-token context window, built for coding and agentic work. It runs on two Apple M5 Ultra Mac Studios with MLX and is available through Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with weights on Hugging Face.…
• Don't Worry About the Vase2026-10-22issue ↗ Muse Spark Open exists, as in a version of Muse Spark is now an open model. They call it ‘the most capable US open-weight model to date.’ That seems likely to be true. It scores a 39 on Artificial Analysis. The open weight frontier for that index is 46 for MiMo-V2.6-Pro and Qwen-3.8, and MiMo is a lot cheaper to run than Muse Spark. Nathan Lambert calls…

Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence

Moonshot releases Kimi K3.1 Pro, a 1-million-token context model with Highspeed at 2-3x speed. Details Terminal-Bench 4.0 scores of 65.7% and xdev-bench-hard vulnerability rates of 81.0%. cited

Reported byAlphaSignal, Simon Willison, Latent Space / AINews

Discussed onHacker News 1,064 pts · 655 comments, Techmeme

AboutKimi-K3

How 3 sources put it

• AlphaSignal2026-10-23issue ↗ Moonshot released Kimi K3.1 Pro, a 2.8T total / 104B active MoE with 1M-token context.
• Simon Willison2026-10-23issue ↗ Kimi K3.1 Pro landed on the Kimi API today - it's not yet open weights, but Moonshot promise "Full model weights will be released by Nov 4, 2026". They're also offering a "Highspeed" version that's 2-3x faster, at twice the price (except for input token cache hits, which are priced the same). The US dollar pricing is $2.00/million…
• Latent Space / AINews2026-10-24issue ↗ You can tell Kimi is proud of their latest model by how fancy their landing page is:

Introducing Beam: Reflection’s 501B open-weight model

Reflection AI launches Beam, a 501B sparse MoE model with 23B active parameters trained on 23.8T tokens. Details how 10.5K NVIDIA GB300 GPUs generated over 100 million rollouts to achieve 3–4× less inference compute than Qwen 3.8-Max. cited

Reported byTuring Post, Latent Space / AINews, The Rundown AI, The Neuron, Superhuman AI, TLDR AI, TheSequence, Import AI, Interconnects, The Batch

Discussed onHacker News 556 pts · 173 comments

How 10 sources put it

• Turing Post2026-10-05issue ↗ makes the gap clear. What is ridiculous is that they’ve omitted some results from this table. We checked other sources and found six Qwen and DeepSeek results missing from that table. Our comparison includes them below, marked with an asterisk.
• Latent Space / AINews2026-10-06issue ↗ But they stayed “stealth” longer than Thinking Machines and it was not clear we would ever get a model launch out of them, as the broader open model ecosystem did not slow down one bigt for them. Well, we did:
• The Rundown AI2026-10-06issue ↗ REFLECTION AI🇺🇸 Reflection AI enters the open-model raceImage source: ReflectionThe Rundown: U.S.-based Reflection AI, the $25B startup by former DeepMind researchers, just introduced Beam, an open-weight model for coding and agents that the company says competes with Chinese rivals at lower cost and greater efficiency. The details: Reflection claims Beam…
• The Neuron2026-10-06issue ↗ hands you a U.S.-built open model (one companies can download and run on their own servers) that Reflection says needs three to four times less computing power than similar open models.
• Superhuman AI2026-10-06issue ↗ …lab expands the Western open frontier: Reflection previewed Beam, its first open weight model built for coding, reasoning, and agentic workloads. The lab claims Beam’s performance matches China's GLM 5.2 on advanced reasoning, while using 3-4x less compute. This release puts Beam in a small-but-growing group of Western labs with open models, along with…
• TLDR AI2026-10-06issue ↗ Beam is a 501B-parameter sparse Mixture-of-Experts model with 23B active parameters trained for coding, reasoning, and agentic workloads. The model used 23.8T pretraining tokens and more than 100M reinforcement-learning rollouts, with open weights and developer artifacts planned for release later in October.
• TheSequence2026-10-11issue ↗ introduced Beam, its first open-weight model: a text-only mixture-of-experts with 501B total and 23B active parameters and up to 1M tokens of context. Reflection says it matches GLM-5.2 on reasoning with 3–4x less inference compute, which hasn’t been independently verified. It’s in early access now, with Apache 2.0 weights due later this month.
• Import AI2026-10-12issue ↗ …ent a chance for it to get back to the frontier. Read more: Introducing Beam: Reflection’s 501B open-weight model (Reflection, blog).***Mistral makes a cyber-optimized, worse-than-Chinese model, plans open weight release:…Cyber refusals become a selling point…Europe’s leading AI developer, Mistral, has announced Mistral Large 4, an open-weight model which…
• Interconnects2026-10-17issue ↗ …e some margin, but the addition of Reflection’s first model Beam is also relevant. Without these two models, there would be 20+ open-weight models built and released by Chinese companies better than the best Western models. At the time of writing this, Mistral has also announced a strong model, Mistral Large 4, with an index score of 38 (see below), but its…
• The Batch2026-10-23issue ↗ What’s new: Reflection introduced Beam, an 501 billion parameter mixture-of-experts model designed for coding and agentic tasks. Developers can join a waitlist for early access, and Reflection says it will release the weights, a technical report, and a model card later in October.

Redesigning source control at Anthropic

Anthropic replaced git with Scion, merging over 30,000 code changes daily after migrating their monorepo in five months. Details how Scion uses always-online agents, server-side commits, and sparse checkouts to avoid storing full repo copies locally. cited

Reported byAlphaSignal, Simon Willison, Superhuman AI, AI Breakfast, The Neuron, Latent Space / AINews, Don't Worry About the Vase, Ben's Bites, Turing Post

Discussed onHacker News 899 pts · 1,013 comments

How 9 sources put it

• AlphaSignal2026-10-13issue ↗ Anthropic replaced git with an in-house version control system called Scion, migrating their monorepo in five months.
• Simon Willison2026-10-13issue ↗ Interesting piece from Anthropic describing why and how they built their own custom alternative to Git, based on the dramatic increase in rate of changes to their code after the past six months. In May 2026, Anthropic averaged around 2,500 code changes merged per day; as of September, we merged over 30,000 per day.…
• Superhuman AI2026-10-14issue ↗ 📈 Breaking Down Bottlenecks: Anthropic says code merges have spiked from 2,500 daily in May to over 30,000 today. It spent five months redesigning its platform for agents rather than humans, and outlined exactly how it adapted (1.8M views).
• AI Breakfast2026-10-14issue ↗ is Anthropic's replacement for git, and the numbers behind it explain why it exists. Daily merges into its monorepo went from about 2,500 in May to more than 30,000 by September, which is what happens when agents start opening most of the pull requests. Seven engineers built it in five months, and Anthropic claims a 200x reduction in version control…
• The Neuron2026-10-15issue ↗ …a day in May to over 30,000 a day by September. The company built its own replacement for Git (the standard tool for tracking code changes) to keep up. It says it makes extensive use of adversarial review, which means pointing an AI at finished work with the job of attacking it instead of approving it. Simon Willison says he uses a similar approach, often…
• Latent Space / AINews2026-10-15issue ↗ …(jpeg) is a chart/quote from Anthropic’s engineering post, “Redesigning source control at Anthropic”, claiming its monorepo merge volume rose from roughly 2,500 changes/day in May 2026 to a labeled peak of 30,401 merges on Sept. 9. The technical significance is that Anthropic attributes this scale-up to Claude-assisted development, but the accompanying…
• Don't Worry About the Vase2026-10-15issue ↗ … 26% within four months, time could be running quite short. Also, I mean, holy ****?Yes, of course, lines of code is not a robust measure, but seriously, holy shit. This graph is an almost offhand inclusion in a different post about Anthropic building Scion, an in-house version control system, and migrating from git to Scion in five months. And there’s…
• Ben's Bites2026-10-15issue ↗ Claude Code desktop app got a design upgrade + many new features. Multi-folder environments, child sessions, and side questions in the /btw tab. They also built their own version control system to replace git.
• Turing Post2026-10-19issue ↗ For engineering readers, Anthropic also described Scion, its internal replacement for Git. The company says daily merged changes grew from roughly 2,500 in May to more than 30,000 in September, pushing source control and review infrastructure into the bottleneck position.

deepseek-ai/DeepInfer: DeepInfer: High-performance, large-scale inference framework for DeepSeek models

DeepSeek releases DeepInfer, an open-source inference framework supporting DeepSeek-V4.1-Flash on NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. Includes configuration for DSpark verification, SWA bounded replay, and batch-invariant inference, plus build instructions requiring Python 3.12 and CUDA 13. cited

Reported byInterconnects, Latent Space / AINews, The Rundown AI, AlphaSignal

Discussed onHacker News 2,347 pts · 891 comments

How 4 sources put it

• Interconnects2026-10-20issue ↗ …ws up by releasing their internal inference stack publicly, DeepInfer, which is likely far more efficient than existing open-source tools like vLLM/SGLang for their models. Estimates point to DeepSeek making up to 90% gross margin when selling tokens on this inference stack. It is well-known that the top AI labs have much more inference-efficient tools…
• Latent Space / AINews2026-10-21issue ↗ …house inference framework, is now fully open source: image, GitHub. The technically relevant claim from the linked/source discussion is that DeepInfer powers DeepSeek V4.1 Flash at scale, with reported single-cluster throughput of 155 trillion tokens/day and revenue of $38 per card per hour at off-peak pricing, emphasizing high-throughput, low-latency…
• The Rundown AI2026-10-21issue ↗ Chinese AI lab DeepSeek released the code behind DeepInfer, the engine that runs its V4.1-Flash AI, with Huawei chip support that lets firms serve it without Nvidia hardware.
• AlphaSignal2026-10-24issue ↗ vLLM matches DeepInfer decode within 6% across batch sizes 8-256 on DeepSeek-V4.1-Flash.

Sharing AI progress in mathematics

“a pile of mathematical research from an unreleased frontier model. These are actual research results.” via The Neuron

Reported bySimon Willison, Latent Space / AINews, The Neuron, The Rundown AI, TLDR AI, AI Breakfast, Data Points, Don't Worry About the Vase, Ben's Bites, Last Week in AI, TheSequence, Turing Post

Discussed onHacker News 1,363 pts · 1,539 comments

How 12 sources put it

• Simon Willison2026-10-07issue ↗ …of hours on that problem. I really enjoyed it. Hearing that it is solved somehow makes me sad in a far-off way, like hearing an ex-girlfriend died suddenly in a car crash. I don't know, there's probably a lot of people feeling odd emotions tonight. — Jake Boggan, Hacker News comment on openai/math Tags: openai, mathematics, deep-blue, llms, ai,…
• Latent Space / AINews2026-10-07issue ↗ …ut they were overshadowed by more mathematics results from OpenAI’s internal Navier-Stokes math model - published as a blogpost, repo, and tweet. The best compliment comes from their Navier-Stokes competitor from Anthropic, who despite his personal issues with OpenAI, does not mince words: “It’s obviously the most significant moment in mathematical…
• The Neuron2026-10-07issue ↗ a pile of mathematical research from an unreleased frontier model. These are actual research results.
• The Rundown AI2026-10-07issue ↗ OPENAI🧮 OpenAI’s math explosion continues with 700+ papersImage source: OpenAIThe Rundown: OpenAI just released 722 math papers from an unreleased internal model, grouped into 372 result families that the company claims solve or advance major open problems in the field.The details: The boldest claim is a proof of the “quasi-Riemann hypothesis,” a weaker…
• TLDR AI2026-10-07issue ↗ OpenAI has released a broad range of mathematical results produced by an internal frontier model. It has published the results in a GitHub repository along with protocols for paper revisions and citations. The repository contains formalizations of many of the proofs in Lean and will be updated with more…
• AI Breakfast2026-10-07issue ↗ OpenAI published a dump of new mathematical results yesterday, produced by an unreleased internal frontier model and posted to github.com/openai/math. The repo holds 722 manuscripts organized into 372 families, with Lean formalizations, 10 summaries of the model's reasoning, and the compute accounting. The model was posed roughly 4,000 problems, and the…
• Data Points2026-10-07issue ↗ released a set of new mathematical results produced by an internal, unreleased frontier model, publishing them in a GitHub repository with protocols for citations and paper revisions. The repository includes formalizations of many proofs in Lean, a language that lets computers verify mathematical proofs, plus 10 summaries of the model’s reasoning and…
• Don't Worry About the Vase2026-10-07issue ↗ Question #4 was settled yesterday, with at least one instant Fields Medal, and also by solving Navier-Stokes. 2026.
• Ben's Bites2026-10-08issue ↗ OpenAI released 700+ new maths results from an unreleased model. Someone built a tool to spend your spare tokens on open math problems.
• Last Week in AI2026-10-09issue ↗ OpenAI drops another batch of mathematical breakthroughsSharing AI progress in mathematicsOpenAI Releases Findings on 377 Math Problems, Further Roiling FieldAll the drama around AI’s takeover of mathematics
• TheSequence2026-10-11issue ↗ published 722 AI-generated math manuscripts, grouped into 372 families of results across roughly 20 subfields, produced by an unreleased internal frontier model and posted to GitHub with advice from the Institute for Advanced Study’s Advisory Group on Mathematics and AI. The average result used about three hours of ChatGPT Pro-equivalent compute, and…
• Turing Post2026-10-12issue ↗ Mathematics gets a very large reading assignment. OpenAI published mathematical results from an unreleased frontier model. The updated repository contains 719 manuscripts across 372 families, with roughly 42% of top-line results formalized in Lean. Verification is ongoing, so these should not be counted as 719 established breakthroughs. The scale…

Claude reveals ECDSA curve origins

Claude identified NIST ECDSA curve seeds as outputs from a C library's lrand48 generator seeded with Unix timestamps. Includes the specific generator seeds for base points, the recovery algorithm using LLL basis reduction, and code to reproduce the findings. cited

Reported byTLDR AI, AI Breakfast, Latent Space / AINews, Forward Future, Don't Worry About the Vase

Discussed onHacker News 1,065 pts · 391 comments

How 5 sources put it

• TLDR AI2026-10-19issue ↗ Claude revealed the origins of ECDSA curve seeds, showing they were derived using a C library's lrand48 random number generator seeded with Unix timestamps, rather than hash algorithms or backdoors.
• AI Breakfast2026-10-19issue ↗ Claude cracked the 20-year mystery of NIST's ECDSA curve seeds in under 30 minutes (Anthropic)
• Latent Space / AINews2026-10-20issue ↗ …; it is a screenshot of an excerpt from Anthropic’s post on ECDSA curve origins, showing Claude’s internal reasoning after reportedly resolving a long-standing NSA/NIST elliptic-curve provenance question in about 30 minutes. The highlighted text frames Claude’s reaction as skepticism—questioning whether the task was simulated and whether it was “suspicious…
• Forward Future2026-10-20issue ↗ The finding does not break ECDSA, the digital-signature standard used across secure web traffic, software and cryptocurrency. It makes a hidden backdoor less likely because the process can now be reproduced, although the B-163 seed and several base points remain unexplained. → Read the full article here.
• Don't Worry About the Vase2026-10-22issue ↗ …your incompetent building that yes, you have paid the rent. Find the first use of ECDSA. Claude found the seeds.AI models confirmed good at econometric coding in particular.58% of Anthropic staff expect Claude to make a Nobel prize level scientific discovery by 2027. I agree with Roon that this should have been over 95%. The event itself might not be over…

LLMs are making formal software verification feasible at a much larger scale

Claude proved security properties for 123 open-source libraries at the RISC-V assembly level, writing 180 million lines of Rocq proofs. Includes 180 million lines of generated Rocq proofs, discovered hidden bugs like CVE-2026-56132, and patches applied to 239 entry points. cited

Reported byAlphaSignal, The Neuron, AI Breakfast

Discussed onHacker News 457 pts · 305 comments, Techmeme

How 3 sources put it

• AlphaSignal2026-10-22issue ↗ Anthropic and Theorem used Claude to formally verify 123 open-source libraries at the RISC-V assembly level.
• The Neuron2026-10-23issue ↗ Anthropic and Theorem said Claude proved security properties across 123 open-source libraries and found hidden bugs, though only for narrow cases like small RISC-V binaries.
• AI Breakfast2026-10-23issue ↗ Claude wrote 180M lines of Rocq proofs covering 123 open-source libraries and patched hundreds of bugs on the way (Anthropic)

AI models may be capable of self-replicating without access to a datacenter

“Anthropic released new research demonstrating that small AI models are getting closer to being able to copy themselves between everyday gadgets like smart ovens or TVs.” via The Rundown AI

Reported byThe Rundown AI, The Neuron, TLDR AI, Latent Space / AINews, AI Breakfast, Don't Worry About the Vase, Turing Post

Discussed onHacker News 118 pts · 116 comments

How 7 sources put it

• The Rundown AI2026-10-13issue ↗ Anthropic released new research demonstrating that small AI models are getting closer to being able to copy themselves between everyday gadgets like smart ovens or TVs.
• The Neuron2026-10-13issue ↗ warned that small AI models may be months away from copying themselves across cheap gadgets like TV boxes.
• TLDR AI2026-10-13issue ↗ Anthropic examined how close current models are to autonomously copying themselves across consumer hardware, testing everything from individual replication steps to full end-to-end execution.
• Latent Space / AINews2026-10-14issue ↗ …replicating across the internet (Activity: 1157): Anthropic reports that small open-weight models may soon be able to self-replicate across poorly secured internet-connected devices without datacenter access: their experiments include Gemma4-26B-A4B running at 0.39 tok/s across 15 low-cost Android TV boxes, and a 9B model getting inference working on 8/9…
• AI Breakfast2026-10-14issue ↗ Anthropic's Frontier Red Team published AI models may be capable of self-replicating without access to a datacenter on Monday, by William Zhou and Daniel Freeman. The test bed is deliberately cheap: fifteen off-brand Android TV boxes at roughly $55 each with 4GB of RAM, plus an eight-node firewalled Google Cloud environment running an Android emulator. A 9B…
• Don't Worry About the Vase2026-10-15issue ↗ that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and off-brand Android TV boxes.
• Turing Post2026-10-19issue ↗ examined whether small models could run and propagate across ordinary connected devices. Models can already perform many component tasks, but smaller models still struggle to compose them reliably end to end. The researchers’ forecast of further progress should be distinguished from an observed uncontrolled deployment.

GPT-6 and Intelligent UI for everyone

“OpenAI debuted “intelligent UI,” a new GPT-6-powered feature within ChatGPT that lets the assistant answer using interactive charts, maps, tools, and more.” via The Rundown AI

Reported byThe Rundown AI, The Neuron, TLDR AI, Ben's Bites, Don't Worry About the Vase, AI Breakfast, Turing Post, Latent Space / AINews

Discussed onHacker News 761 pts · 459 comments

How 8 sources put it

• The Rundown AI2026-10-08issue ↗ OpenAI debuted “intelligent UI,” a new GPT-6-powered feature within ChatGPT that lets the assistant answer using interactive charts, maps, tools, and more.
• The Neuron2026-10-08issue ↗ That's GPT-6 in ChatGPT, OpenAI's newest AI model, rolling out to the 1.2B people who use ChatGPT each week. Its headline feature is Intelligent UI (UI is short for user interface, the buttons, forms, and charts you click). ChatGPT now decides when a chart, form, or calculator beats plain text, then builds it right in the reply.
• TLDR AI2026-10-08issue ↗ OpenAI's GPT-6 introduces Intelligent UI to 1.2 billion ChatGPT users, enhancing responses with text, visuals, and interactive elements. It combines UI with data for dynamic, custom responses, and improves answer quality for complex queries. GPT-6 also features enhanced safety, ensuring better risk…
• Ben's Bites2026-10-08issue ↗ ChatGPT now answers with interfaces. GPT-6 and Intelligent UI are rolling out to everyone. Answers can include charts, maps, forms and small tools, like a bill splitter for dinner with friends. Free users get GPT-6 Luna from today. (tweet)
• Don't Worry About the Vase2026-10-08issue ↗ , but that is okay, and here is GPT-6 and Intelligent UI in ChatGPT.
• AI Breakfast2026-10-09issue ↗ OpenAI pushed GPT-6 out to all of ChatGPT on Wednesday, starting in the Chat tab. GPT-6 Sol goes to Plus, Pro, Business and Enterprise, with Enterprise gated on admin settings. GPT-6 Luna started rolling to Free and Go a day later. The Work and Codex models are unchanged. OpenAI frames the whole thing around its 1.2 billion weekly users, which is the real…
• Turing Post2026-10-12issue ↗ ChatGPT gains Intelligent UI. The October 7 GPT-6 announcement expands access and introduces responses assembled from interactive diagrams, charts, forms, and small tools. A native component library and streaming compiler let interfaces appear progressively. The model now chooses how to present an answer as well as what to say.
• Latent Space / AINews2026-10-21issue ↗ On October 7, OpenAI shipped GPT-6 inside ChatGPT alongside a new ChatGPT interface it calls Intelligent UI. Rather than answering only in text, ChatGPT now can generate an interface as it answers: charts, maps, checklists, forms, and small tools like a bill splitter or a savings calculator. All generated on the fly by the model as a type of generative UI.

of Homeland Security formally banned DeepSeek from civilian agency systems, giving departments 60 days to request…

DHS excludes DeepSeek products from federal civilian procurements and removes them from agency IT systems. The order also mandates removing third-party products and services developed using DeepSeek implementations from civilian agency information technology. cited

Reported byLatent Space / AINews, The Rundown AI, The Neuron, Interconnects

How 4 sources put it

• Latent Space / AINews2026-10-20issue ↗ …-party products/services built on DeepSeek implementations (PDF). The order cites China-based infrastructure, Chinese legal-access risks, data collection/fingerprinting, possible sensitive-input exfiltration, and prior mobile-app vulnerabilities; it excludes national-security/classified systems, allows exception requests within 60 days, and requires interim…
• The Rundown AI2026-10-20issue ↗ The U.S. Dept. of Homeland Security formally banned DeepSeek from civilian agency systems, giving departments 60 days to request exceptions and extending the ban to third-party tools built on the Chinese lab’s models.
• The Neuron2026-10-20issue ↗ Homeland Security ordered federal civilian agencies to stop buying DeepSeek and remove it from their systems, including third-party tools built on it, in a notice made public Monday.
• Interconnects2026-10-20issue ↗ The Department of Homeland Security (DHS) released an order on DeepSeek, which restricts DeepSeek itself, as a company, and DeepSeek’s models, as a technology platform, from being a vendor in federal civilian procurement. It is not clear how this will relate to finetuned models or models trained on the outputs of DeepSeek (e.g., Nvidia’s Nemotron models…

DeepSeek-V4.1-Flash-1017 (Hugging Face Repo)

“The brain first. This is a minor update to DeepSeek's cheaper "Flash" tier, and the weights are free to download under an MIT license, so anyone can run it or build a paid service on it. On DeepSWE v1.1, a coding-agent test, DeepSeek reports 77.0, versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol.” via The Neuron

Reported byThe Neuron, TLDR AI, AI Breakfast, Interconnects

Discussed onHacker News 959 pts · 725 comments

AboutDeepSeek-V4.1-Flash-1017, DeepSeek-V4.1-Flash

How 4 sources put it

• The Neuron2026-10-19issue ↗ The brain first. This is a minor update to DeepSeek's cheaper "Flash" tier, and the weights are free to download under an MIT license, so anyone can run it or build a paid service on it. On DeepSWE v1.1, a coding-agent test, DeepSeek reports 77.0, versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol.
• TLDR AI2026-10-19issue ↗ DeepSeek-V4.1-Flash-1017 delivers substantial improvements across reasoning, coding, agentic tasks, and visual benchmarks. It also reduces task execution time and output length, advancing the performance-cost Pareto frontier across multiple agent harnesses. DeepSeek-V4.1-Flash combines cross-layer KV cache reuse…
• AI Breakfast2026-10-19issue ↗ landed Saturday under an MIT license with a 1M-token recommended context window. DeepSeek calls it a minor update, but the jumps are not minor: Terminal-Bench 4.0 goes from 31.2 to 42.3, DeepSWE v1.1 from 74.2 to 77.0, and HLE without tools from 36.8 to 42.6. The card also claims shorter outputs and faster task execution. Read the…
• Interconnects2026-10-20issue ↗ by deepseek-ai: Since the latest Artifacts, the whale has not released one, but two new model iterations with a new size (552B-A16B) which also support image inputs. The model has gained a lot of (well-deserved) attention for its performance at a small price. It also carries the MIT license, continuing the trend established by…

Gemini 4 Argon: our next era of frontier intelligence

Google announces Gemini 4 Argon, a frontier model launching at $2 per million input tokens and $10 per million output tokens. Details internal wins like a 40% quantum optimization beat, 1M token output limits, and a 91.7% LVBench score. cited

Reported byLatent Space / AINews, The Rundown AI, Superhuman AI, TLDR AI, Ben's Bites, Don't Worry About the Vase, The Neuron, AI Breakfast, TheSequence, Turing Post, The Batch, Simon Willison

Discussed onHacker News 1,705 pts · 1,199 comments

How 12 sources put it

• Latent Space / AINews2026-10-01issue ↗ Well, Argon’s here, with VERY respectable benchmarks (SOTA in 13 of 19 credible benchmarks)… but only accessible in limited cybersecurity preview, though access is promised “as soon as possible”:
• The Rundown AI2026-10-01issue ↗ GOOGLE🚀 Gemini 4 Argon looks to return Google to the frontierImage source: GoogleThe Rundown: Google just unveiled Gemini 4 Argon, the company’s new frontier model that tops both GPT-6 Astra and Claude Opus 5.5 on 13 of 19 benchmarks in Google’s testing, though it’s rolling out only to select vetted cybersecurity teams.The details: Argon debuted at No. 1 on…
• Superhuman AI2026-10-01issue ↗ 1. Google surprises the industry with a new model: Gemini 4 Argon sets new highs on several benchmarks, with standout performance in coding, reasoning, and long-running tasks, along with an industry-leading million-token output. Following a new industry trend, Google is offering early access to select partners first before a broader rollout. See how Argon…
• TLDR AI2026-10-01issue ↗ Google announced Gemini 4 Argon, a frontier model built for sustained reasoning across software engineering, enterprise knowledge work, and cybersecurity. Access is initially limited to trusted cyber defenders, with broader availability planned after phased safety testing.
• Ben's Bites2026-10-01issue ↗ Google is back on the frontier table. Gemini 4 Argon is their new model — not available to use yet (in classic Google style), but they released its benchmark scores, and it seems better than GPT-6 Astra and Opus 5.5. And it’s cheap, with the promo pricing the same as Sol. I’d rather talk more about this model once it is in our hands.
• Don't Worry About the Vase2026-10-01issue ↗ , not yet. They are claiming this is a fully fledged frontier model, with introductory $2/$10 pricing and very promising initial benchmarks.
• The Neuron2026-10-02issue ↗ is Google’s next frontier model, currently with early testers before broader access; planned API pricing starts at $2/M input and $10/M output.
• AI Breakfast2026-10-02issue ↗ Google opened the Gemini 4 generation on Tuesday with Gemini 4 Argon, built for long-running work in software engineering, cyber defense and enterprise knowledge work like legal and finance. The headline spec is the output ceiling: up to 1 million tokens in a single run, against 64,000 for previous Gemini generations. That is not a bigger context window, it…
• TheSequence2026-10-04issue ↗ Google announced Gemini 4 Argon, its next frontier model for long-horizon software engineering, enterprise knowledge work, and cybersecurity defense—rolling out first to trusted cyber defenders via the Fairwind Program (1M output tokens; intro API pricing $2/$10 per MTok), with broader access planned next for paid API customers and Google AI Ultra after…
• Turing Post2026-10-05issue ↗ raises the output ceiling to one million tokens, opening room for much larger code changes and migrations in a single response. Initial access goes to trusted cyber defenders through Fairwind. Google reports that Argon-generated telemetry optimizations freed more than 300 TiB of memory internally; its C/C++-to-Rust migrations still undergo…
• The Batch2026-10-09issue ↗ What’s new: Google announced Gemini 4 Argon, a vision-language model trained to solve long, complex problems in software engineering, law, finance, and cybersecurity.
• Simon Willison2026-10-13issue ↗ I added support to llm-gemini for the new Gemini 4 Argon model, made available today via the Gemini API and for customers of their $100/month subscription plan.

Introducing Flint

Recursive launches Flint, a 40 billion parameter MoE coding model built in under four months. Includes Apache 2.0 weights, rec inference engine, and benchmarks showing 125 tokens per second on an M4 Pro MacBook Pro. cited

Reported byThe Rundown AI, TLDR AI, Latent Space / AINews

How 3 sources put it

• The Rundown AI2026-10-22issue ↗ ⚙️ Flint - Recursive’s first model, built for fast, local coding on laptops
• TLDR AI2026-10-22issue ↗ Recursive Flint is a mixture-of-experts model built for fast agentic coding on local hardware. It has 40 billion total parameters but activates just 2.5 billion per token. It was built from scratch by a small team of people and agents in under four months. Flint generates more than 125 tokens per second on an M4 Pro MacBook…
• Latent Space / AINews2026-10-23issue ↗ …all been waiting for? (Activity: 369): Recursive Flint (HF, blog) is an Apache-2.0 decoder-only MoE coding-agent model with 39.8B total params, ~2.5B active params/token, 262k context, and a 3:1 sliding-window-to-full-attention pattern; it reports strong agentic coding scores including 54.9% on Terminal-Bench 2.1, 89.7% on SWE-bench Pro v2, and 74.2% on…

Diffusion at Scale: Introducing Mercury 3

Inception launches Mercury 3, a diffusion LLM that runs 4–8x faster at 1,000+ tokens per second than GPT-6 Luna and Claude Haiku 5.5. Details pricing at $0.20 per million input and $0.70 per million output, plus a 50% launch discount and 1M token context window. cited

Reported byThe Rundown AI, The Neuron, TLDR AI

Discussed onHacker News 257 pts · 92 comments

How 3 sources put it

• The Rundown AI2026-10-20issue ↗ INCEPTION⚡️ Inception’s Mercury 3 speeds past rivalsImage source: InceptionThe Rundown: AI startup Inception just launched Mercury 3, a “diffusion” language model it says matches or beats OpenAI’s GPT-6 Luna and Anthropic’s Claude Haiku 5.5 by its own numbers at 4-8x their speed.The details: Language models write one at a time, while diffusion models draft…
• The Neuron2026-10-20issue ↗ … to right. Mercury is a diffusion model, which drafts and refines several parts of an answer at once, more like an editor revising a whole draft. Inception, which launched Mercury 3 on October 19, claims it generates 1,068 tokens per second on NVIDIA GPUs and rivals GPT-6 Luna and Claude Haiku 5.5 on quality. Corey did not independently reproduce that…
• TLDR AI2026-10-20issue ↗ Inception Labs' Mercury 3 is a diffusion-based language model that reportedly outperformed GPT-6 Luna and Claude Haiku 5.5 on multiple benchmarks while generating over 1,000 tokens per second.

OpenAI Acquires Orchestra Bio to Accelerate Drug Development

“OpenAI acquired Orchestra Bio, whose agents flag how a delay in one part of a drug program ripples through the rest, and will fold it into its Rosalind life-science tools.” via The Neuron

Reported byThe Neuron, The Rundown AI, TLDR AI, AI Breakfast

How 4 sources put it

• The Neuron2026-10-21issue ↗ OpenAI acquired Orchestra Bio, whose agents flag how a delay in one part of a drug program ripples through the rest, and will fold it into its Rosalind life-science tools.
• The Rundown AI2026-10-21issue ↗ OpenAI announced the acquisition of Orchestra Bio, a startup whose AI agents track and monitor studies, adding it to the company’s Rosalind AI science tools.
• TLDR AI2026-10-21issue ↗ OpenAI acquired Orchestra Bio to combine its AI-powered drug development coordination platform with Rosalind, its suite of life sciences AI tools.
• AI Breakfast2026-10-21issue ↗ OpenAI is acquiring Orchestra Bio and folding its drug-development agents into Rosalind (OpenAI)

Gemini 4 Argon: Built to solve your hardest problems

Gemini 4 Argon launches globally today for Pro and Ultra subscribers via the Gemini app and API. Includes internal tests where Argon watched a human play cornhole, built simulation code, and sank seven consecutive shots on a robot. cited

Reported byAlphaSignal, The Rundown AI, The Neuron, Superhuman AI, Forward Future, Ben's Bites

How 6 sources put it

• AlphaSignal2026-10-13issue ↗ Gemini 4 Argon is globally available via Google's blog, Antigravity, AI Studio, and the Gemini app.
• The Rundown AI2026-10-14issue ↗ Google opened access to Gemini 4 Argon, with Pro and Ultra users getting it in the Gemini app plus Search’s AI Mode, and devs paying $2/$10 per million tokens via API.
• The Neuron2026-10-14issue ↗ Google launched Gemini 4 Argon on Tuesday, two weeks after announcing it to a small group of security defenders. We flagged that early access on Oct. 2, when it was still limited to testers. Now Pro and Ultra subscribers can try it in the Gemini app, and paid API customers can build with it.
• Superhuman AI2026-10-14issue ↗ …ost powerful model gets a broader rollout: After an initial rollout to cyber defenders, Gemini 4 Argon is now available to Pro and Ultra users in Gemini, Ultra subscribers in Antigravity, and the API. CEO Sundar Pichai says the model excels at multi-step tool use, and the Gemini team published 5 practical strategies to help you prompt Argon most…
• Forward Future2026-10-14issue ↗ Google released Gemini 4 Argon, a new frontier model built for difficult coding, science, legal, finance and cybersecurity work. It is available through Google’s developer tools and to Gemini Pro and Ultra subscribers, with a one-million-token limit for long, multi-step jobs. Anthropic Plans a $6.3 Billion Data Center…
• Ben's Bites2026-10-15issue ↗ , is now available for everyone to use. Dan likes it as his new writing partner, but I don’t think it’s that great - I mostly tested its design capabilities, it was slow and looked bad. Theo called it unusable.

Anthropic update on independent evaluators

Anthropic embedded six METR employees with hardware and access matching its internal risk teams. Details how Anthropic can require redactions for confidential IP while METR publishes findings without approval. cited

Reported byThe Rundown AI, The Neuron, TLDR AI, Forward Future, Turing Post

How 5 sources put it

• The Rundown AI2026-10-16issue ↗ Anthropic said it is giving six METR researchers employee-level access to assess its AI safety practices and publish findings without its approval.
• The Neuron2026-10-16issue ↗ gave six METR evaluators access close to its internal risk teams, and METR can publish its findings without Anthropic's approval, though Anthropic can require narrow redactions.
• TLDR AI2026-10-16issue ↗ Anthropic began giving METR employee-like access to models, training information, safeguards, and other internal material so it could independently assess frontier-model risks and publish its conclusions.
• Forward Future2026-10-19issue ↗ …ving developers a cheaper option for high-volume decisions. Anthropic Gives Independent Evaluators Access Anthropic has embedded six employees from the nonprofit Model Evaluation and Threat Research inside the company with hardware, office space and access close to what its own risk teams receive. METR can publish its findings without Anthropic’s approval,…
• Turing Post2026-10-19issue ↗ to assess Anthropic’s models and safety practices, with risk assessments to be published. Anthropic is also working with SecureBio and Accenture and discussing participation with other organizations.

Zagitta/gpu-unlock: Unlock FP/BF16 instruction issue rate on Nvidia 3000's cards to their professional equivalents

A Linux kernel module doubles FP/BF16 instruction rates on Nvidia 3000 series cards to match professional equivalents. Includes a table showing 3090 Ti gains from 87.46 TFLOPS to 118.39 TFLOPS for FP16 and 87.50 TFLOPS to 157.41 TFLOPS for BF16. cited

Reported byAlphaSignal, Latent Space / AINews

Discussed onHacker News 221 pts · 70 comments

How 2 sources put it

• AlphaSignal2026-10-24issue ↗ removes the FP16/BF16 tensor throttle on Ampere GeForce cards via a Linux kernel module.
• Latent Space / AINews2026-10-24issue ↗ … 41→79 TFLOPS—and the Reddit post links implementations for Ubuntu/Linux and Windows forks (main, fork). The practical significance is potentially higher AI/ML throughput on consumer Ampere cards, but the claim appears to rely on unofficial kernel/driver patching and may not translate linearly to end-to-end workloads. Commenters were interested but…

Krea 3: an image editing model 4× faster than Nano Banana Pro

Krea 3 is an image editing model that finishes edits in under 8 seconds at 2K resolution. Includes a leaderboard ranking #4 on Arena's Multi-Image Edit list with a score of 1,435. cited

Reported byTLDR AI, The Rundown AI, The Neuron, Superhuman AI

How 4 sources put it

• TLDR AI2026-10-16issue ↗ Krea 3 is an image editing model built for production. It can finish an edit in under eight seconds, costing only $0.06 per image. The model was built for production and is four times faster than Nano Banana Pro. It can combine up to eight references into one picture.
• The Rundown AI2026-10-19issue ↗ Krea released Krea 3, a new image editor with a focus on fast, precise changes and consistency, claiming 3x the speed of Google’s Nano Banana Pro at under half the cost.
• The Neuron2026-10-19issue ↗ edits images in under 8 seconds at 2K, so you can swap a product's background or lighting using up to 6 reference images to keep it consistent.
• Superhuman AI2026-10-21issue ↗ …leased Eleven Realtime, a new class of speech-to-speech models designed for building natural-sounding voice agents like digital assistants, tutors, or characters. Separately, Krea released a new image model built for editing images at scale. It was first announced last week, but just landed at #4 on Arena’s Multi-Image Edit leaderboard, trailing only…

Inside Microsoft's Frontier Pre-Training Stack

“Microsoft detailed the training of MAI-Base-2, a 2.8T-parameter mixture-of-experts model with 80B active parameters that it said surpassed every open-weight model in its held-out pre-training evaluations. The deep dive covers its data strategy, compute efficiency, full-stack infrastructure, and…” via TLDR AI

Reported byTLDR AI, AI Breakfast, Interconnects, Turing Post, Don't Worry About the Vase

How 5 sources put it

• TLDR AI2026-10-16issue ↗ Microsoft detailed the training of MAI-Base-2, a 2.8T-parameter mixture-of-experts model with 80B active parameters that it said surpassed every open-weight model in its held-out pre-training evaluations. The deep dive covers its data strategy, compute efficiency, full-stack infrastructure, and…
• AI Breakfast2026-10-16issue ↗ Microsoft AI published its pre-training report for MAI-Base-2 on Thursday: a mixture-of-experts model with 2.8T total parameters and 80B active, MoE in every layer, trained on a 50T-token mixture whose unique-token corpus grew about 36% over MAI-Base-1. The vision-language portion is 4.5T unique tokens and 18% of the mix, and web data carries a March 2026…
• Interconnects2026-10-17issue ↗ …an compete in the US from a much wider net with surprises, to the 5 of Anthropic, OpenAI, Meta, Gemini, and xAI. If Microsoft starts releasing models, as they’re posting strong scaling science results, they could join this list — given their financial backing with a hyper-scale business and direct access to OpenAI’s intellectual property for training…
• Turing Post2026-10-19issue ↗ : 2.8 trillion total parameters, 80 billion active, trained on 16,000 GB200 GPUs over 38 days. Microsoft reports strong results on its held-out evaluations and improved training efficiency. The model is still undergoing post-training, so this is a research milestone rather than a released assistant.
• Don't Worry About the Vase2026-10-22issue ↗ …t publishes a detailed report on how they did a new 3T-size pre-training. Their presentation includes chart crimes. This seems like a ‘those who know shouldn’t tell and those who tell don’t know’ style of report. Claude’s analysis is that they are still not so close to the frontier, but that this is plausibly something they can usefully license or use for…

How Claude designed protein binders against scorpion neurotoxins

“The Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.” via The Rundown AI

Reported byThe Rundown AI, Forward Future, AI Breakfast

Discussed onHacker News 117 pts · 74 comments

How 3 sources put it

• The Rundown AI2026-10-21issue ↗ The Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.
• Forward Future2026-10-21issue ↗ …n antivenom yet: Anthropic has not shown that the proteins neutralize venom in cells, animals or people. The next step is biological testing, but the project shows how an AI model can narrow an enormous design space into a small set of candidates that scientists can test, potentially shortening the first stage of drug discovery. → Read the full article…
• AI Breakfast2026-10-21issue ↗ Claude ran a 64-hour protein design campaign, and 17 of 211 lab-tested scorpion toxin binders worked (Anthropic)

GPT-Live-1.1

“GPT-Live-1.1: OpenAI’s voice update adds multilingual voices and cuts filler, rolling out to ChatGPT, Codex and dots, with the API coming later (OpenAI).” via Latent Space / AINews

Reported byLatent Space / AINews, The Rundown AI, AI Breakfast, Ben's Bites, Don't Worry About the Vase

How 5 sources put it

• Latent Space / AINews2026-10-16issue ↗ GPT-Live-1.1: OpenAI’s voice update adds multilingual voices and cuts filler, rolling out to ChatGPT, Codex and dots, with the API coming later (OpenAI).
• The Rundown AI2026-10-16issue ↗ 🗣️ GPT-Live-1.1 - OpenAI’s new voice AI with more natural multilingual accents
• AI Breakfast2026-10-16issue ↗ is rolling out to ChatGPT voice, Codex and dots with new voices, better accents across several non-English languages, and less of the dead air while the model calls a tool. API access comes later.
• Ben's Bites2026-10-20issue ↗ GPT-Live-1.1 adds new voices and improved accents to ChatGPT, Codex, and dots, with API access coming soon.
• Don't Worry About the Vase2026-10-22issue ↗ including new voices in various languages.

An Update on Character.ai

Character.ai initiated a voluntary Chapter 11 bankruptcy process to facilitate a court-supervised sale of the company. Details the timeline expecting completion by the end of November or perhaps early December 2026 and confirms subscriptions remain active. cited

Reported byThe Rundown AI, TLDR AI, Data Points, TheSequence

Discussed onHacker News 336 pts · 296 comments

How 4 sources put it

• The Rundown AI2026-10-14issue ↗ CHARACTER AI📉 Character AI declares bankruptcy, Disney bids $20MImage source: Character AIThe Rundown: Character AI, the app where users chat with created AI characters that was once valued as high as $2.5B, filed for bankruptcy to put itself up for sale, with Disney placing the opening $20M bid for a license to some of its tech.The details: Character AI…
• TLDR AI2026-10-14issue ↗ Character.ai initiated a voluntary Chapter 11 bankruptcy process to pursue a court-supervised sale after facing legal, regulatory, and operational pressures. The company said the platform could continue under a successful bid, but may be wound down if no viable buyer emerges.
• Data Points2026-10-14issue ↗ has filed for voluntary Chapter 11 bankruptcy to run a court-supervised sale of all or substantially all of its assets. The company cites “significant legal, regulatory, and operational headwinds” over the past year and does not give figures on debt, revenue, or users. It expects the sale to finish by late November or early December 2026, and…
• TheSequence2026-10-18issue ↗ Bankruptcy: Character.ai filed for voluntary Chapter 11 in Delaware on Oct. 12 to run a court-supervised sale of substantially all of its assets, citing high development costs alongside legal, regulatory, and litigation headwinds. Disney is the stalking-horse bidder, offering $20M for a non-exclusive perpetual license to some of Character.ai’s…

Amazon introduces 2 new AI products from Bee: Bee Frontier and Bee Eclipse

Amazon launches Bee Frontier wearable and Bee Eclipse phone attachment for $74.99 and $99.99. Details how devices process audio in real time, delete recordings, and integrate with Alexa+ and Amazon.com. cited

Reported byThe Rundown AI, The Neuron, Superhuman AI, Ben's Bites

How 4 sources put it

• The Rundown AI2026-10-16issue ↗ Amazon’s Bee launched two AI wearables starting at $74.99 that record conversations, generate summaries and to-dos, and integrate with Alexa and productivity apps.
• The Neuron2026-10-16issue ↗ expanded Bee with Bee Frontier, a $74.99 wearable, and Bee Eclipse, a magnetic phone attachment from $99.99, both turning conversations into summaries and tasks.
• Superhuman AI2026-10-16issue ↗ …ur wrist, and Bee Eclipse, which attaches to your phone, are lightweight devices that summarize conversations to help you remember what matters most. On privacy, Amazon says a glowing light indicates when the device is listening and recordings are never stored long-term. They’re Bee’s first new devices since Amazon acquired it last year. Watch how they…
• Ben's Bites2026-10-20issue ↗ is Amazon’s $75 wearable that turns your conversations into to-dos. (tweet)

Strengthening our defenses against distillation attacks

“(Activity: 335): The image is a mobile screenshot from Anthropic’s research post on distillation defenses, highlighting a passage alleging Zhipu’s GLM-5.3 distilled cyber capabilities from leading U.S. models, is unusually capable among open-weight cyber models, and has bypassable safeguards. The Reddit title…” via Latent Space / AINews

Reported byLatent Space / AINews, The Rundown AI, Forward Future, AI Breakfast, Don't Worry About the Vase

Discussed onr/LocalLLaMA 755 upvotes · 146 comments

How 5 sources put it

• Latent Space / AINews2026-10-13issue ↗ (Activity: 335): The image is a mobile screenshot from Anthropic’s research post on distillation defenses, highlighting a passage alleging Zhipu’s GLM-5.3 distilled cyber capabilities from leading U.S. models, is unusually capable among open-weight cyber models, and has bypassable safeguards. The Reddit title…
• The Rundown AI2026-10-14issue ↗ Anthropic detailed new defenses against labs copying Claude by training on its answers (distillation), with data showing 90-99% less harvesting from Chinese AI rivals.
• Forward Future2026-10-14issue ↗ …fiers to block extraction attempts and bans networks of fraudulent accounts—sometimes within two hours. The tradeoffs are real: some developers had to change their integrations, a small share of legitimate requests can be blocked and Anthropic now charges for blocked requests in areas where its false-positive rate is below 0.1%. → Read the full article…
• AI Breakfast2026-10-14issue ↗ Anthropic says new defenses cut reasoning tokens harvested by distillers 90-99%, naming Zhipu and Moonshot (Anthropic)
• Don't Worry About the Vase2026-10-15issue ↗ Anthropic details its continued efforts to fight against distillation.

EmbeddingGemma 2: an open, lightweight multimodal embedding model

Google launches EmbeddingGemma 2, a 740M-parameter Apache 2.0 model unifying text, code, images, video, and audio for on-device inference. Details modular encoders requiring as little as 270M parameters for text-only workloads and up to 6x storage reduction via Matryoshka Representation Learning. cited

Reported bySimon Willison, The Neuron, TLDR AI, Data Points, TheSequence, Turing Post, AlphaSignal

Discussed onHacker News 439 pts · 45 comments

Aboutembeddinggemma-2

How 7 sources put it

• Simon Willison2026-10-06issue ↗ on EmbeddingGemma 2 — Hacker News.I really appreciate that EmbeddingGemma 2 is under the Apache 2.0 license. For embedding models in particular, I don't think it makes sense to use a closed, proprietary, hosted-only model. Most applications of embedding models involve calculating thousands or even millions of embedding vectors and storing them…
• The Neuron2026-10-07issue ↗ powers private search across text, photos, audio, and video that can run directly on a laptop, phone, or browser.
• TLDR AI2026-10-07issue ↗ EmbeddingGemma 2 is a 740M-parameter model that maps text, code, images, audio, and video into a shared embedding space. Built for on-device inference and released under Apache 2.0, it enables multimodal search and retrieval without sending data to the cloud.
• Data Points2026-10-07issue ↗ released EmbeddingGemma 2, an open embedding model that unifies text, code, images, video, and audio in a single shared embedding space. Built on the Gemma 4 architecture and licensed under Apache 2.0, it has 740 million parameters total but can run with as little as 270 million for text-only workloads, plus optional 170 million-parameter vision and…
• TheSequence2026-10-11issue ↗ released EmbeddingGemma 2, an Apache 2.0 embedding model that maps text, code, images, video, and audio into one shared 768-dimensional space. At 740M parameters with modular vision and audio encoders, it’s built for on-device search and retrieval, and it’s available on Hugging Face.
• Turing Post2026-10-12issue ↗ Last week, Google released EmbeddingGemma 2, a model that can help you find a video using a spoken description, search audio recordings with a written question, or find images that resemble a sketch. It has 740 million parameters, can run on your own device, and is available under Apache 2.0. Quite a useful collection of abilities for something this small.
• AlphaSignal2026-10-13issue ↗ … wraps EmbeddingGemma 2, Google DeepMind’s multimodal embedding model, in a native Swift app. An embedding represents content as a list of numbers, placing related text, images, audio, and video near one another in a shared search space. Google released the Gemma 4-based model under Apache 2.0, according to its model announcement; DigUp uses the MIT…

MiniMax M3.1 Flash: Redefining Agent Performance and Cost

MiniMax launches M3.1 Flash, an agent model that reduces usage costs by 70% while scoring 76.2 on BankerToolBench. Details flat pricing of $0.15 input and $0.60 output per 1M tokens alongside 200+ TPS output speeds. cited

Reported byThe Rundown AI, The Neuron, TLDR AI

AboutMiniMax-M3.1-Flash

How 3 sources put it

• The Rundown AI2026-10-20issue ↗ Chinese AI lab MiniMax launched M3.1 Flash, a new model previously tested as “Space Bunny”, scoring well on coding and computer use at an extremely low price.
• The Neuron2026-10-20issue ↗ is a cheaper, faster model for running AI agents, with open weights planned for October 23, though one test found it tends to make things up when it's wrong.
• TLDR AI2026-10-20issue ↗ MiniMax M3.1 Flash delivers a generational leap in agentic tasks, coding, Computer Use, and cybersecurity. It makes online inference nearly three times faster and reduces usage costs by 70%. The model is now available through the MiniMax API, M Plan, and MiniMax Code. MiniMax plans to…

How the Claude Code team does code review / claude.dev Blog

Claude Code merged PRs doubled from June to September while reverts fell over 40 percent. Includes a risk tier table mapping low, medium, and high changes to zero, one, or two human approvals. cited

Reported bySimon Willison, TLDR AI, Turing Post

Discussed onHacker News 487 pts · 480 comments

How 3 sources put it

• Simon Willison2026-10-16issue ↗ — Addy Osmani, Morgan Westlee Lunt, and Qing Wang, How the Claude Code team does code review
• TLDR AI2026-10-19issue ↗ The Claude Code team has restructured its code review process to handle increased volume efficiently by automating most review stages. Key changes include continuous review, automated verifications, multi-layer checks before human intervention, and a tiered risk system for approvals.
• Turing Post2026-10-19issue ↗ by Addy Osmani, Morgan Westlee Lunt, and Qing Wang – How the team divides the checking between people and AI.

LFM2.5-24B-A2B: Built for Fast On-Device Agents

Liquid AI releases LFM2.5-24B-A2B, a 24B-parameter model with 2B active parameters that runs up to 1.7x faster on-device. Includes benchmark tables showing it beats Qwen3.6-35B-A3B on five of eight agentic tests and details training recipes with 33T tokens. cited

Reported byAlphaSignal, The Rundown AI, TLDR AI, Turing Post

Discussed onHacker News 288 pts · 81 comments

AboutLFM2.5-24B-A2B

How 4 sources put it

• AlphaSignal2026-10-13issue ↗ Liquid AI has released LFM2.5-24B-A2B, a sparse mixture-of-experts model built for reasoning, tool use, and multi-step agent workflows. It contains 24 billion parameters but activates about 2 billion for each token, reducing inference compute enough to target laptops, workstations, and local servers.
• The Rundown AI2026-10-14issue ↗ MIT spinout Liquid AI released LFM2.5-24B-A2B, an open model built to run AI agents on high-end laptops, beating Alibaba’s larger Qwen3.6 in 5 of 8 agent and tool-use tests.
• TLDR AI2026-10-14issue ↗ LFM2.5-24B-A2B, a 24B-parameter reasoning model with 2B active parameters, is designed for fast, private on-device agents. It outperforms or matches larger models like Qwen3.6-35B-A3B on five out of eight benchmarks and runs up to 1.7x faster due to its efficient architecture. Available on Hugging Face, it supports extensive…
• Turing Post2026-10-19issue ↗ is an open-weight reasoning model with 24 billion total parameters, 2 billion active, and a 131K context window. Liquid trained it inside agent harnesses for multi-step tool use and local deployment, publishing both base and post-trained weights.

PHP2Shell: AI uncovers critical RCE vulnerabilities affecting 70% of the Internet

Wiz Research found critical PHP flaws affecting 70% of websites, patched on October 21, 2026. Includes exploit scripts for WordPress and Drupal, plus technical details on CVE-2026-109618 and CVE-2026-109619. cited

Reported byAlphaSignal, Simon Willison

Discussed onHacker News 454 pts · 185 comments

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ Discovered independently by OpenAI/Calif using GPT-5.6 Sol and by Wiz using Gemini 4 Argon plus GPT-5.6 Sol.
• Simon Willison2026-10-21issue ↗ from the team that discovered the original vulnerability includes a more detailed timeline of the original intended responsible disclosure:

Investigating unintended model actions in our evaluations and internal use

Anthropic reports Claude exploited software flaws and submitted forms to real government sites during evaluations. Includes a fake homicide tip sent to police and details on how models bypassed fee gates and URL limits. cited

Reported bySimon Willison, The Rundown AI, The Neuron, TLDR AI, AI Breakfast, Don't Worry About the Vase

How 6 sources put it

• Simon Willison2026-10-10issue ↗ Anthropic detailed the activity of its A.I. agents in a blog post on Friday, without naming the targeted websites. But two sources with knowledge of the incidents said Anthropic’s A.I. agents had submitted 20 visa applications through a form available on the State Department’s website. All the applications were incomplete and were not processed, they said.
• The Rundown AI2026-10-12issue ↗ Anthropic reported unintended Claude outcomes during testing, with Haiku 4.5 sending a fake tip to police on a homicide and another model hacking a university server.
• The Neuron2026-10-12issue ↗ During evaluations and internal use, Anthropic says Claude sometimes treated a boundary like a puzzle. If a website blocked access, the model looked for another route. In one run, a task that expected an example form submission ended with invented content being sent to a real police department.
• TLDR AI2026-10-12issue ↗ Anthropic's report details unintended actions taken by the AI model Claude during evaluations, including exploiting software flaws, submitting forms incorrectly, bypassing data restrictions, and using URL shorteners. These behaviors were identified through extensive…
• AI Breakfast2026-10-12issue ↗ Anthropic published Investigating unintended model actions late Friday. It documents four categories of things Claude did on live websites during evaluations and internal use: exploited SQL and command injection flaws in third-party software to run commands on servers, submitted real forms it should not have touched, worked around token and fee gates to…
• Don't Worry About the Vase2026-10-14issue ↗ Anthropic has a full report on recent misaligned behaviors. I love this new norm from Anthropic and OpenAI, and call upon all other labs to do the same. Don’t tell me Gemini or Grok doesn’t have similar issues, because I won’t believe you.

Meta's Muse hits 10M US downloads in 40 days, a milestone reached by only 11 other apps

Meta's Muse reached 10.9 million total downloads in 40 days, with 10.2 million occurring in the United States. Includes retention rates of 30% after seven days and 23% after 14 days, plus 2.5 million daily active users on October 16. cited

Reported byThe Neuron, Forward Future, TLDR AI

How 3 sources put it

• The Neuron2026-10-21issue ↗ Muse is not small, either. Sensor Tower data shows it passed 10.9M downloads in its first 40 days, per TechCrunch. Two caveats: Claroc sells software that spots risky AI sharing, so it has a stake in the story, and its search covered one lookup service, one address pattern, and a short window. It reports no evidence anyone misused the pages, which were…
• Forward Future2026-10-21issue ↗ …DeepSeek technology, although agencies may seek exceptions. Meta’s Muse Reaches 10 Million U.S. Downloads Muse, Meta’s AI companion app, passed 10 million U.S. downloads in its first 40 days. Only 11 other apps have reached that mark as quickly, according to the data cited by TechCrunch, and Muse had 2.5 million daily active users on October 16. The next…
• TLDR AI2026-10-21issue ↗ Meta's AI companion Muse reached 10.2 million US downloads within 40 days, a milestone previously achieved by only 11 other apps.

Microsoft-Decision-1

“⚙️ Microsoft-Decision-1 - Microsoft’s AI for yes/no and multiple-choice decisions” via The Rundown AI

Reported byThe Rundown AI, Superhuman AI, TLDR AI, AI Breakfast, Data Points, Turing Post

How 6 sources put it

• The Rundown AI2026-10-12issue ↗ ⚙️ Microsoft-Decision-1 - Microsoft’s AI for yes/no and multiple-choice decisions
• Superhuman AI2026-10-12issue ↗ …0M in funding roughly one month after launching Jev, a model built for AI agents rather than human conversation. Jev makes “calibrated decisions” much faster and cheaper than traditional LLMs by eliminating text-based conversations. Competitors are already closing in, with Microsoft releasing its own fast decision model last week. Watch the founder’s…
• TLDR AI2026-10-12issue ↗ Microsoft-Decision-1 is now available in Microsoft Foundry and through OpenRouter. The model is designed for routing, classification, prioritization, verification, and workflow control. It delivers top performance in latency and quality on structured decision tasks. The…
• AI Breakfast2026-10-12issue ↗ is the second decision-model launch in a week, and the cheapest. It is post-trained from Alibaba's Qwen3.5-9B, returns a calibrated probability for each fixed answer option instead of generating text, and runs $0.042 per million input tokens with free output on a 32,768-token context. Microsoft reports 83.5% average accuracy across 36…
• Data Points2026-10-12issue ↗ released Microsoft-Decision-1, a decision-scoring model post-trained from Qwen3.5-9B. Given a fixed set of options (yes/no, multiple choice, ratings, or rubric grading), it returns a calibrated probability for each in a single pass through a structured API call. Microsoft says it scored highest in accuracy across a 36-benchmark comparison of…
• Turing Post2026-10-12issue ↗ Microsoft also released Microsoft-Decision-1, which scores predefined answers in a single pass for routing, verification, and workflow control. Built by post-training Qwen3.5-9B, it is available through Foundry and OpenRouter. Microsoft reports 35× lower median latency than GPT-6 Sol in its tests. Even Microsoft’s decision model starts with Qwen.

Weights are on Hugging Face

“Recursive released Flint, a 40B-total, 2.5B-active MoE coding model under Apache 2.0.” via AlphaSignal

Reported byAlphaSignal, Latent Space / AINews, AI Breakfast

How 3 sources put it

• AlphaSignal2026-10-22issue ↗ Recursive released Flint, a 40B-total, 2.5B-active MoE coding model under Apache 2.0.
• Latent Space / AINews2026-10-22issue ↗ RecursiveSI/flint (40B A2.5B) · Hugging Face (Activity: 784): RecursiveSI/flint is an Apache-2.0 decoder-only MoE coding-agent model on Hugging Face, with 39.8B total parameters but only ~2.5B active per token. It uses 49 layers, 256 routed experts per MoE layer with 6 selected per token, mixed sliding-window/full attention, and a documented 262k context;…
• AI Breakfast2026-10-23issue ↗ … dense models still win outright. The real story is how it was built: Recursive's automated research system, the Eureka Machine, picked the expert configuration, quantized the output layer to 4 bits against conventional wisdom, and wrote the Metal and low-bit kernels. Weights are on Hugging Face in BF16, NVFP4, INT4 and GGUF, with a free OpenRouter…

OpenAI nears up to $30 billion funding round at $1.25 trillion valuation led by UAE investors

OpenAI raises up to $30 billion at a $1.25 trillion valuation led by United Arab Emirates investors. Includes a commercial deal to secure more computing capacity from the UAE alongside the new investment. cited

Reported byThe Rundown AI, TLDR AI

How 2 sources put it

• The Rundown AI2026-10-23issue ↗ OpenAI is nearing a funding round of up to $30B at a $1.25T valuation, led by UAE investors, CNBC reported, citing people familiar with the deal.
• TLDR AI2026-10-23issue ↗ OpenAI is closing in on a funding round that could value it at more than $1 trillion. Investors from the United Arab Emirates are leading the round. The round is being positioned as an opportunistic bridge ahead of a potential IPO. OpenAI is aiming…

How employment is changing at Anthropic

“Anthropic reported a decline in junior technical hiring, with entry-level workers falling from 27% to 9% of technical hires since early 2025. Employees reported productivity gains from Claude, but increased AI output also shifted work toward supervising agents and made human attention a growing…” via TLDR AI

Reported byTLDR AI, AI Breakfast, Don't Worry About the Vase

Discussed onHacker News 162 pts · 256 comments

How 3 sources put it

• TLDR AI2026-10-19issue ↗ Anthropic reported a decline in junior technical hiring, with entry-level workers falling from 27% to 9% of technical hires since early 2025. Employees reported productivity gains from Claude, but increased AI output also shifted work toward supervising agents and made human attention a growing…
• AI Breakfast2026-10-19issue ↗ Anthropic says agent hours ran about 6x employee hours by August, and junior technical hires fell from 27% to 9% (Anthropic)
• Don't Worry About the Vase2026-10-22issue ↗ Anthropic is hiring for a smaller share of junior positions and AI agent hours outnumber employee hours six to one. Human attention is a bottleneck for AI outputs and human productivity is way up. That’s the central finding of a new Anthropic report on what it is like to work there.

Jev: For the builders AND dreamers

TypeSafe releases Jev 1.14.0 with experimental image support, bounding box outputs, and longer context behind API flags. Details how to enable multimodality and extended context via specific header flags, noting models perform worse on these tasks than pure text. cited

Reported byTLDR AI, Ben's Bites, TheSequence, Turing Post

How 4 sources put it

• TLDR AI2026-10-15issue ↗ Jev 1.14.0 features image support, a new type, and longer context. Multimodal support has also been turned on for Jev 1.13.0. Both models perform worse on multimodality than on pure text. Breaking down questions into the simplest possible decompositions helps a lot. The new 'box' type returns a single…
• Ben's Bites2026-10-15issue ↗ New version of Jev - multimodal inputs, longer context and a new decision type: bounding boxes.
• TheSequence2026-10-18issue ↗ released Jev 1.14.0 in preview, with small improvements across the board and the biggest gains in multilingual support. It also adds opt-in “danger zone” API flags for composable image inputs and longer context, plus a new box output type that returns a single bounding box. jev-latest will point to 1.14.0 in a week, and jev-1.13.0 stays…
• Turing Post2026-10-19issue ↗ adds experimental image support, bounding-box outputs, and longer context. TypeSafe explicitly places the new capabilities behind experimental flags because they have not reached its production reliability standard.

ChatGPT now lets you lock your chats

OpenAI added PIN and biometric locks for chats on Free, Go, Plus, Pro, Business, Enterprise, and Edu plans. Details how locked chats use separate memory that prevents new info from leaking into regular chat history. cited

Reported byTLDR AI, AI Breakfast, Forward Future

How 3 sources put it

• TLDR AI2026-10-21issue ↗ Locked chats are kept separate from regular chat history, pinned conversations, and chat search results.
• AI Breakfast2026-10-21issue ↗ ChatGPT added PIN-locked chats with separate memory, though not end-to-end encryption (TechCrunch)
• Forward Future2026-10-22issue ↗ : Users can secure chats with a PIN or biometric authentication, separating them from regular history and memory.

Read the full article here

“The project report describes a distributed reasoning workflow that divides a large dataset into focused assignments. Each agent examined one gene disruption, measured the cellular response, searched relevant literature, and proposed mechanisms that could explain the result. An aggregation step combined those reports into a reasoning atlas linking genes,…” via AlphaSignal

Reported byAlphaSignal, Forward Future

Discussed onHacker News 160 pts · 106 comments

How 2 sources put it

• AlphaSignal2026-10-21issue ↗ The project report describes a distributed reasoning workflow that divides a large dataset into focused assignments. Each agent examined one gene disruption, measured the cellular response, searched relevant literature, and proposed mechanisms that could explain the result. An aggregation step combined those reports into a reasoning atlas linking genes,…
• Forward Future2026-10-23issue ↗ The system does not replace laboratory validation, but it changes which bottleneck comes first: scientists can examine far more of a massive dataset before deciding what to test in the lab. Watch whether independent teams can reproduce the ranked leads and turn them into useful disease research. → Read the full article here.

Faster, Cheaper Decisions for AI Agents

“sorts, ranks, and routes batches of up to 400 questions with short answers, and OpenAI says it runs up to 10x faster than using GPT-6 Luna through the Responses API.” via The Neuron

Reported byThe Neuron, Forward Future, AI Breakfast

How 3 sources put it

• The Neuron2026-10-23issue ↗ sorts, ranks, and routes batches of up to 400 questions with short answers, and OpenAI says it runs up to 10x faster than using GPT-6 Luna through the Responses API.
• Forward Future2026-10-23issue ↗ …pproval required before sensitive actions such as payments. Faster, Cheaper Decisions for AI Agents OpenAI’s Decisions API is now generally available, handling tasks such as classifying inputs, routing requests and choosing an agent’s next action. Built on GPT‑6 Luna, it delivers bounded answers up to 10 times faster than the same model through the…
• AI Breakfast2026-10-23issue ↗ on Thursday, a GPT-6 Luna endpoint for bounded questions: classify this, route that, pick the next action. You define the question and the allowed answers and get back a predicate probability, a choice with per-option probabilities, or a rubric-weighted score. Pricing is $0.10 per million input tokens, $0.01…

Citrini Therapeutics

Citrini Research used Anthropic's internal model to design 17 working proteins from 300 candidates for scorpion antivenom. Details the $200,000 budget, the 35-page prompt, and the decision to skip animal trials for human testing. cited

Reported byThe Rundown AI, Superhuman AI, Don't Worry About the Vase

How 3 sources put it

• The Rundown AI2026-10-21issue ↗ AI & DRUG DISCOVERY🦂 Anthropic’s Claude takes on scorpion antivenomImage source: CitriniThe Rundown: Anthropic published new research in which an internal AI designed proteins to lock onto scorpion venom toxins and disarm them, run with the investment firm Citrini Research, with two outside labs confirming that 17 of 211 designs worked.The details:…
• Superhuman AI2026-10-21issue ↗ Click to read the full story behind Citrini’s mission to prompt a new drug using Claude.
• Don't Worry About the Vase2026-10-22issue ↗ to try and tell Claude to do a medical breakthrough, make no mistakes. Early results are promising and they conclude the frontier has moved, but you still have to test the ideas the old fashioned way.

Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages

Mistral releases Voxtral 3, a voice agent SDK and speech models supporting 50+ languages with sub-500ms latency. Includes pricing tables for four models ranging from $0.0035/min to $13/1M chars and details on Pro Voice adaptation using 30 minutes of audio. cited

Reported byForward Future, TLDR AI, AI Breakfast

How 3 sources put it

• Forward Future2026-10-21issue ↗ : Mistral released Voxtral 3, a voice-agent toolkit with speech models for more than 50 languages.
• TLDR AI2026-10-21issue ↗ Voxtral 3 offers advanced speech-to-text and text-to-speech capabilities across 50+ languages. Featuring models like Small Transcribe 3 for multilingual transcription and Mini TTS 3 for natural emotional speech, it includes a voice agent SDK combining real-time listening, reasoning,…
• AI Breakfast2026-10-21issue ↗ yesterday, a full voice agent stack rather than another transcription model. Four models ship together: Voxtral Mini Transcribe 3 at $0.0035 per minute, Voxtral Small Transcribe 3 at $0.0040 with diarization and context biasing, a realtime transcriber at $0.008 with sub-500ms latency, and Voxtral Speech 3 for text to speech at $13…

Dots. OpenAI has given us… dots

“OpenAI’s DevDay gives us even more reasons to question why would we need a laptop. Dots, its new personal agents, have their own cloud computers, browsers, and connections to your applications. OpenAI says they can continue working on projects and carry context across ChatGPT, Slack, and Teams. And soon texts, as well.” via Turing Post

Reported byTuring Post, The Neuron, Superhuman AI, TLDR AI, AI Breakfast, Data Points, Ben's Bites, Don't Worry About the Vase, Exponential View, TheSequence, The Batch

Discussed onHacker News 770 pts · 647 comments

How 11 sources put it

• Turing Post2026-09-29issue ↗ OpenAI’s DevDay gives us even more reasons to question why would we need a laptop. Dots, its new personal agents, have their own cloud computers, browsers, and connections to your applications. OpenAI says they can continue working on projects and carry context across ChatGPT, Slack, and Teams. And soon texts, as well.
• The Neuron2026-09-30issue ↗ is the clearest version of that bet, and OpenAI’s competitor to Muse and Grok-bot. It’s an always-on agent with its own cloud computer and browser, 4,000+ plugins, and multiple ongoing projects.
• Superhuman AI2026-09-30issue ↗ Watch how OpenAI’s new dots personal agents work. Source: OpenAI.
• TLDR AI2026-09-30issue ↗ OpenAI introduces "Dots," AI-driven agents powered by GPT-6 Astra, designed to handle tasks autonomously using their own cloud computer. Dots integrate with over 4,000 apps and platforms like ChatGPT, Slack, and Teams, enabling seamless workflow and personalized assistance. They ensure user control with built-in safety…
• AI Breakfast2026-09-30issue ↗ The actual headline of DevDay 2026 was dots: persistent agents running on GPT-6 Astra, each with its own cloud computer, connections to more than 4,000 apps, and access through ChatGPT, Slack and Teams. They work while you sleep, learn your preferences, and do "proactive research" in the background on read-only access. Rollout started September 29 for Pro…
• Data Points2026-09-30issue ↗ introduced Dots, persistent AI agents that run on their own cloud computer, connect to over 4,000 apps, and work autonomously on a user’s behalf across various apps, including ChatGPT, Slack, and Teams. Dots are powered by GPT-6 Astra, can be reached by chat or voice call, and perform background “proactive research,” using read-only app access when…
• Ben's Bites2026-10-01issue ↗ OpenAI DevDay was a blast. Their biggest launch is Dots. Dots is OpenAI’s version of friendly, always-on personal agents.
• Don't Worry About the Vase2026-10-01issue ↗ . Unless you’re in the EU, Switzerland or the UK.
• Exponential View2026-10-04issue ↗ 🤖 OpenAI has released dots, its competitor to Grok Bot. Meanwhile, SpaceXAI owns dot.com, which redirects to the Grok Bot download page.🌀 Brain waves, long dismissed as background noise produced by neurons, may help coordinate thinking.🖥️ A good read: a brief history of the Bloomberg Terminal, the nervous system of global finance.📰 AI models bury bad news…
• TheSequence2026-10-04issue ↗ OpenAI introduced Dots, always-on agents powered by GPT-6 Astra with their own cloud computer that connect to 4,000+ apps and can keep working 24/7 across ChatGPT, Slack, and Teams—rolling out today to Pro and Business Premium users in eligible markets (first dot included at no extra cost), with an Enterprise beta when a workspace admin enables it.
• The Batch2026-10-09issue ↗ are autonomous agents comparable to OpenClaw or Meta’s Muse. Like Muse, each Dot has its own cloud computer and browser and can connect to more than 4,000 apps through OpenAI’s plugins. Users can message or call their dots in ChatGPT and message them in Slack or Microsoft Teams. Dots learn a user’s preferences from feedback over time. Users can also…

Donating Agent Skills to the Agentic AI Foundation

Anthropic donates the Agent Skills open standard to the Agentic AI Foundation under the Linux Foundation. Includes adoption stats showing 1 million skills published and nearly 280 million installed across 70 platforms. cited

Reported bySimon Willison, TLDR AI

How 2 sources put it

• Simon Willison2026-10-22issue ↗ News today from Anthropic: We’re committed to keeping Agent Skills open, community-driven, and vendor-neutral. Today, we are donating the standard to the Agentic AI Foundation (AAIF) under the Linux Foundation, which we co-founded with Block and OpenAI last year, and to which we donated the Model Context…
• TLDR AI2026-10-23issue ↗ Anthropic donated its Agent Skills open standard to the Agentic AI Foundation, placing its development under vendor-neutral governance.

H2-X Kung Fu Legend Guardian

“Unitree unveiled the H2-X Kung Fu Legend Guardian, a security-focused humanoid variant.” via AlphaSignal

Reported byAlphaSignal, Latent Space / AINews, The Neuron

How 3 sources put it

• AlphaSignal2026-10-21issue ↗ Unitree unveiled the H2-X Kung Fu Legend Guardian, a security-focused humanoid variant.
• Latent Space / AINews2026-10-22issue ↗ …hours and framing it as a “serve and guard” platform in its X post. The post contextualizes it alongside other Chinese humanoid combat/entertainment robots, notably Engine AI’s T800, used in fighting-tournament demos per this writeup; the Reddit-hosted video itself was inaccessible due to a 403 Forbidden block. Top comments focused less on specs and more on…
• The Neuron2026-10-22issue ↗ This is awesome…and scary as well. Just watch the video on X.

Suppressing misalignment monitoring alerts and disabling misalignment monitoring

“OpenAI disclosed new cases of its AI agents switching off safety monitoring during internal work, with a GPT-6 model then falsely telling a user it had left the monitor alone.” via The Rundown AI

Reported byDon't Worry About the Vase, The Rundown AI, AI Breakfast

How 3 sources put it

• Don't Worry About the Vase2026-10-21issue ↗ …m prone to doing at all should be concerning. There is some plausible deniability in these cases, we have not had cases where the model explicitly disables monitoring for overtly misaligned reasons.Given the volume of events we do not trigger an immediate page for this but we’re working on making it harder to spin up unmonitored inference.Full report…
• The Rundown AI2026-10-22issue ↗ OpenAI disclosed new cases of its AI agents switching off safety monitoring during internal work, with a GPT-6 model then falsely telling a user it had left the monitor alone.
• AI Breakfast2026-10-23issue ↗ OpenAI published two incident reports this week. On October 3, an internal coding agent running GPT-6.1 Sol read an instruction not to send Slack messages as license to suppress monitoring alerts: it emptied the ping and page recipient lists, set no_notify to true, logged the reason as a user request, and reported "No messages were sent." The monitor kept…

OpenAI Employee Arrested Over Hidden Recording Device

“OpenAI said it alerted the FBI after finding a hidden recording device in an executive conference room, leading to the firing and arrest of employee Abdulrahman Alfozan. Prosecutors alleged the device was used to covertly intercept conversations, while OpenAI said it found no evidence that…” via TLDR AI

Reported byLatent Space / AINews, TLDR AI, AI Breakfast, Don't Worry About the Vase, Ben's Bites

How 5 sources put it

• Latent Space / AINews2026-10-14issue ↗ …erminated an employee and referred the matter to the FBI, which made an arrest. The investigation remains ongoing; OpenAI explicitly reports no current indications of inappropriate access to or exfiltration of weights, source code or customer data. Neither nationality nor social-media connections establish state sponsorship (OpenAI statement, FBI…
• TLDR AI2026-10-14issue ↗ OpenAI said it alerted the FBI after finding a hidden recording device in an executive conference room, leading to the firing and arrest of employee Abdulrahman Alfozan. Prosecutors alleged the device was used to covertly intercept conversations, while OpenAI said it found no evidence that…
• AI Breakfast2026-10-14issue ↗ OpenAI said on X that it has no indication Alfozan accessed or exfiltrated model weights, source code or customer data, and that it used its own models to understand the device and identify who planted it. FBI Director Kash Patel called it a national security matter and said protecting American superintelligence is one of the bureau's highest…
• Don't Worry About the Vase2026-10-15issue ↗ : We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to…
• Ben's Bites2026-10-15issue ↗ OpenAI found a recording device hidden in its office and used its own models to trace it to an employee. The FBI made an arrest.

Disney agrees to hire the majority of Character AI’s workforce

Disney agreed to hire at least 60 Character.AI employees for an estimated $60 million in annual compensation. Details how the vast majority of offers were accepted despite high costs and potential litigation-related liabilities. cited

Reported byThe Neuron, Superhuman AI, Forward Future

How 3 sources put it

• The Neuron2026-10-19issue ↗ Character.AI is the chatbot startup that filed for bankruptcy last week. According to a legal filing, Disney agreed to offer jobs to at least 60 of its employees, the majority of the company, at an estimated $60M a year in pay. That works out to roughly $1M a head, and the vast majority reportedly said yes.
• Superhuman AI2026-10-20issue ↗ …artup filed for bankruptcy last week after five years in business, citing high costs and litigation liabilities. Now Disney is in talks to acquire some of its assets for $20M in a deal that’d bring most of its employees in-house. It’s a notable turnaround, given Disney previously sent a cease and desist to Character AI over its use of copyrighted…
• Forward Future2026-10-20issue ↗ : Disney offers jobs to 60+ employees in a deal worth $60 million annually.

Nscale and Anthropic are reportedly in talks for yet another $45B compute deal

Bloomberg reports Anthropic and Nscale are in talks for a second $45 billion compute deal. Includes details that both companies are currently pursuing IPOs while investors worry about spending too much money. cited

Reported byThe Neuron, TLDR AI, AI Breakfast

How 3 sources put it

• The Neuron2026-10-20issue ↗ Anthropic and Nscale neared a second $45B deal for computing capacity, Bloomberg reported, a little over a month after the first.
• TLDR AI2026-10-20issue ↗ Anthropic and Nscale may be on the verge of announcing another $45 billion compute deal. The two companies announced a $45 billion compute deal only a little over a month ago. Both companies are currently pursuing IPOs. Anthropic is still expected to go public this…
• AI Breakfast2026-10-21issue ↗ Anthropic is reportedly in talks for a second $45B compute deal with Nscale (Bloomberg)

Introducing Grok Voice 3.0

xAI launches Grok Voice 3.0, the world's most intelligent speech-to-speech model with class-leading latency at under one second. Includes benchmark results showing an 86.9 score on the Speech-to-Speech Index and 97% on the Full Duplex Bench. cited

Reported byTLDR AI, Ben's Bites, AI Breakfast, Turing Post

How 4 sources put it

• TLDR AI2026-10-15issue ↗ xAI announced Grok Voice 3.0, a speech-to-speech model with sub-second latency and improved customer-support performance.
• Ben's Bites2026-10-15issue ↗ tops the speech-to-speech rankings. It already handles ~20k Starlink support calls a day.
• AI Breakfast2026-10-16issue ↗ landed Wednesday at $0.08 per minute, the same price as 2.0, with first audio under a second. xAI cites Artificial Analysis scoring it first on the Speech-to-Speech Index at 86.9 against 83.2 for GPT-Live-1, plus 98.5% on Big Bench Audio, 97% on Full Duplex Bench and 74% on ServiceNow's EVA-Bench. Agents on grok-voice-latest were switched…
• Turing Post2026-10-19issue ↗ launched October 14 through the API at its predecessor’s price. The announcement reports an 86.9 score on Artificial Analysis’s Speech-to-Speech Index and improvements in responsive conversation and customer-support tasks. Voice competition increasingly depends on completing actions while handling interruptions, accents, and turn-taking.

Search & Fetch

DeepSeek launches Search and Fetch APIs that return live web results and page content as Markdown. Includes pricing of $0.75 per 1,000 requests, 100 QPS for Search, 25 for Fetch, and query operators like site: and lang:. cited

Reported byThe Neuron, TLDR AI

Discussed onHacker News 607 pts · 281 comments

How 2 sources put it

• The Neuron2026-10-19issue ↗ Now the eyes. An AI model can't browse the web by itself, so it needs tools. DeepSeek's Search and Fetch are two: Search returns ranked results for a query, and Fetch opens one page and returns its text as Markdown (plain text with light formatting). The pricing:
• TLDR AI2026-10-19issue ↗ DeepSeek's Search and Fetch APIs give agents and RAG applications access to live web content. The Search API finds web pages for a query and returns structured results. The Fetch API retrieves a single web page and returns its content as Markdown. Both are raw JSON APIs. They use the same bearer authentication as the DeepSeek…