Model Follower.

Up to date

Rebuilt
2026-11-15 07:14 UTC
Release radar
108 on watch
Hub snapshot
2026-11-15
Re-checked
daily · every number sourced and dated

Links / November 2026

Tuesday, 3 November 2026

The AI links the most newsletters and link sources had shared by the end of the day (UTC), each source counted once.

As of 2026-11-03 23:59 UTC Status sealed

That day 36-hour window

  1. Shared by 7

    Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

    GPT-6 Astra edits 0.016% of Qwen3.5-0.8B weights to raise MATH-500 from 7.9% to 48.1% without training. Includes demos where Astra identifies games from raw weights and writes MinAtar agents beating PPO 100M steps. cited

    Reported byTuring Post, Latent Space / AINews, The Neuron, The Rundown AI, Superhuman AI, Simon Willison

    Discussed onHacker News 2,928 pts · 880 comments

    Where it was shared
    rikiyat.github.io paper · 18 min read First shared 2026-11-02
  2. Shared by 4

    LLMs are not superhuman at reading and writing neural network weights

    A paper claims GPT-6 Astra improved a 0.8B Qwen LLM's math by natively editing its weights. Details how the fix involved steering weights to stop loops after 32k tokens, a step a human researcher could take. cited

    Reported byTLDR, TLDR AI, Simon Willison

    Discussed onHacker News 142 pts · 22 comments

    AboutQwen3.5-0.8B

    Where it was shared
    seangoedecke.com analysis · 4 min read First shared 2026-11-02
  3. Shared by 4

    OpenAI grades its AI on everyday tasks

    “OpenAI unveiled LifeVal, a benchmark grading models on everyday personal-assistant tasks.” via AlphaSignal

    Reported byAlphaSignal, The Rundown AI, The Deep View, TLDR AI

    How 1 source put it
    openai.com First shared 2026-11-02
  4. Shared by 4

    Anthropic warns that it's very close to RSI capability, aka self-building AI

    Anthropic's updated S-1 states it may imminently meet its definition of dramatically accelerated progress attributable to AI. Details the company's admission that it has already used AI to develop new models whose leaps in capability it didn't recognize at the time. cited

    Reported byForward Future, TLDR, TLDR AI

    Discussed onTechmeme

    Where it was shared
    techcrunch.com news report · 3 min read First shared 2026-11-02
  5. Shared by 3

    OpenAI is finally merging ChatGPT's Chat and Work tabs

    OpenAI merges Chat and Work tabs in the ChatGPT desktop app for select users. Details how files and plugins stay together while Codex remains separate and shares the same token pool. cited

    Reported byForward Future, The Deep View, TLDR AI

    Where it was shared
    techcrunch.com news report · 1 min read First shared 2026-11-03
  6. Shared by 6

    We’ve received our first response from the Project Suncatcher prototype satellite.

    Google received the first AI response from its Project Suncatcher satellite after uplinking a prompt to the onboard Gemma 4 model. Includes the haiku generated by the model as it passed Svalbard and plans to test thermal stress and radiation exposure next. cited

    Reported byThe Rundown AI, The Deep View, TLDR AI, AI Breakfast, FryAI, AlphaSignal

    Aboutgemma-4-26B-A4B-it-assistant, gemma-4-31B-it-assistant

    Where it was shared
    blog.google news report · 1 min read First shared 2026-11-02
  7. Shared by 3

    Elon Musk Limits Grok Bot Usage Over Rising Costs

    “Elon Musk responded to commentary surrounding users seeing significant usage cuts on SpaceXAI’s Grok Bot, saying “The cost of serving Grok Bot was unsustainably high relative to revenue.”” via The Rundown AI

    Reported byThe Rundown AI, The Deep View, TLDR AI

    How 1 source put it
    x.com X post First shared 2026-11-03
  8. Shared by 3

    Anthropic filed an updated IPO prospectus with new figures like nearly $17B in Q3 revenue and a $567M operating loss,…

    “Anthropic disclosed $895.4B in commitments for purchases and leases, about $70B due within a year, in an amended IPO filing.” via The Neuron

    Reported byThe Neuron, The Rundown AI, The Deep View

    How 1 source put it
    sec.gov First shared 2026-11-03
  9. Shared by 3

    Make long-running agents faster with message threads

    Message threads are now in public beta, cutting median time to first token by 35% for 500k to 1M token requests. Includes the thread parameter usage, memory encryption details, and a 5-minute expiration rule for client re-sends. cited

    Reported byAlphaSignal, The Deep View, TLDR AI

    Where it was shared
    claude.com announcement · 2 min read First shared 2026-11-02
  10. Shared by 2

    The 'Astra can read and write weights to improve Qwen' post is misleading

    Matthew Honnibal argues the Astra Qwen post is misleading because it conflates raw weight reading with Python-based editing. Details a table showing only the Qwen experiment used Python to run the model, while true raw weight reading occurred on smaller networks. cited

    Reported bySimon Willison

    Discussed onHacker News 956 pts · 199 comments

    honnibal.dev analysis · 4 min read First shared 2026-11-03
  11. Shared by 2

    p-e-w/heretic: Training-free language model modification: Remove censorship, suppress slop, reduce sycophancy, and more

    “Built with Heretic v2 Arbitrary-Rank Ablation, Apache-2.0 licensed, runs on llama.cpp, Ollama, LM Studio.” via AlphaSignal

    Reported byAlphaSignal, Latent Space / AINews

    How 1 source put it
    github.com 7 min read First shared 2026-10-29
  12. Shared by 2

    GPU-accelerated Weight Teleportation

    “humans& released GPU-accelerated Weight Teleportation, lossless RL weight sync faster than intra-cluster RDMA.” via AlphaSignal

    Reported byAlphaSignal, TLDR AI

    How 1 source put it
    humansand.ai 15 min read First shared 2026-11-02
  13. Shared by 2

    Anthropic confirms new Claude Code limits to prevent game cloning — users report Opus 5.5 is also actively blocking reverse engineering tasks

    “Anthropic confirmed new Claude Code limits that steer Opus 5.5 away from cloning specific games, while users report reverse-engineering blocks Anthropic hasn't confirmed.” via The Neuron

    Reported byLatent Space / AINews, The Neuron

    How 1 source put it
    tomshardware.com 3 min read First shared 2026-11-03
  14. Shared by 2

    China publicly discloses plan for a lunar landing within three years

    “Chinese space officials released their clearest schedule yet for landing astronauts on the Moon before 2030. The plan calls for an uncrewed Mengzhou spacecraft test in early 2027, a lunar-lander test in 2028 and a full lunar-orbit rehearsal in 2029. The first test will show whether the schedule can hold as China races NASA back…” via Forward Future

    Reported byForward Future, TLDR

    How 1 source put it
    • Forward Future2026-11-03issue ↗ Chinese space officials released their clearest schedule yet for landing astronauts on the Moon before 2030. The plan calls for an uncrewed Mengzhou spacecraft test in early 2027, a lunar-lander test in 2028 and a full lunar-orbit rehearsal in 2029. The first test will show whether the schedule can hold as China races NASA back…
    • TLDR2026-11-03issue ↗
    arstechnica.com 2 min read First shared 2026-11-03
  15. Shared by 2

    Asymptote-Labs/agent-beacon: The cross-harness, self-improving memory layer for AI agents. Join our community: https://discord.com/invite/zdNChS2fBu

    “records what your coding agents do, from Claude Code to Cursor, into local logs you can view on a dashboard (open source).” via The Neuron

    Reported byThe Neuron, Superhuman AI

    How 1 source put it
    github.com 12 min read First shared 2026-11-03
  16. Shared by 2

    Anthropic reports $17 billion in Q3 revenue, purchase commitments of $895.4 billion in revised S-1

    “Anthropic releases amended IPO prospectus with third-quarter financials (1 minute read) Anthropic's amended IPO prospectus reveals a 1,200% increase in revenue to $33.24 billion for the first nine months of 2026. The company reported a net loss of $104 billion for the same period. Anthropic's commitments for compute and infrastructure total approximately…” via TLDR AI

    Reported byTLDR AI

    Discussed onHacker News 237 pts · 284 comments

    cnbc.com 2 min read First shared 2026-11-02
  17. Shared by 3

    Scaling Git at OpenAI

    “OpenAI unveiled Pando, a Git-compatible backend powering its entire monorepo.” via AlphaSignal

    Reported byAlphaSignal, TLDR

    Discussed onHacker News 266 pts · 179 comments

    How 1 source put it
    openai.com First shared 2026-11-02
  18. Shared by 2

    Coming in 2027: Apple’s $1,000 home robot for FaceTime and switching on lights

    “: Roaming robot targeted for 2027 would handle FaceTime and lights, priced at $1,000 or more.” via Forward Future

    Reported byForward Future, TLDR

    How 1 source put it
    arstechnica.com 1 min read First shared 2026-11-03
  19. Shared by 2

    Decision Model API

    Hume launches an API that returns probabilities for questions about voice recordings. Includes a 0.82 average score over 10 benchmarks at 0.23 seconds median latency. cited

    Reported byAlphaSignal, The Rundown AI

    Where it was shared
    hume.ai product page · 2 min read First shared 2026-11-02
  20. Shared by 4

    Qwenputer: a fully functioning, deterministic RISC-V CPU inside Qwen3.5

    Qwenputer embeds a 32-bit RISC-V CPU into Qwen3.5-4B weights, passing riscv-tests with 18 million free-running instructions. Includes a 1,024-slot filesystem holding 512 files and scores 100% on NUPA long-integer arithmetic versus 48.6% for base Qwen. cited

    Reported byLatent Space / AINews, TLDR AI, Simon Willison

    Discussed onr/LocalLLaMA 830 upvotes · 123 comments

    AboutQwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled, Qwen3.5-397B-A17B

    Where it was shared
    divagr.com repo · 9 min read First shared 2026-10-28

The week to that day two or more sources

  1. Shared by 7

    Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

    GPT-6 Astra edits 0.016% of Qwen3.5-0.8B weights to raise MATH-500 from 7.9% to 48.1% without training. Includes demos where Astra identifies games from raw weights and writes MinAtar agents beating PPO 100M steps. cited

    Reported byTuring Post, Latent Space / AINews, The Neuron, The Rundown AI, Superhuman AI, Simon Willison

    Discussed onHacker News 2,928 pts · 880 comments

    Where it was shared
    rikiyat.github.io paper · 18 min read First shared 2026-11-02
  2. Shared by 15

    Gemini 4 Carbon: our next frontier model

    “Google announced Gemini 4 Carbon, its new frontier model focused on cybersecurity and long-horizon agentic coding.” via AlphaSignal

    Reported byAlphaSignal, Latent Space / AINews, The Neuron, The Rundown AI, Forward Future, Superhuman AI, Ben's Bites, TLDR AI, AI Breakfast, NLP Newsletter, What's Up in AI, TheSequence, FryAI, Turing Post

    Discussed onHacker News 1,018 pts · 803 comments

    How 1 source put it
    blog.google 4 min read First shared 2026-10-28
  3. Shared by 6

    We’ve received our first response from the Project Suncatcher prototype satellite.

    Google received the first AI response from its Project Suncatcher satellite after uplinking a prompt to the onboard Gemma 4 model. Includes the haiku generated by the model as it passed Svalbard and plans to test thermal stress and radiation exposure next. cited

    Reported byThe Rundown AI, The Deep View, TLDR AI, AI Breakfast, FryAI, AlphaSignal

    Aboutgemma-4-26B-A4B-it-assistant, gemma-4-31B-it-assistant

    Where it was shared
    blog.google news report · 1 min read First shared 2026-11-02
  4. Shared by 4

    OpenAI grades its AI on everyday tasks

    “OpenAI unveiled LifeVal, a benchmark grading models on everyday personal-assistant tasks.” via AlphaSignal

    Reported byAlphaSignal, The Rundown AI, The Deep View, TLDR AI

    How 1 source put it
    openai.com First shared 2026-11-02
  5. Shared by 9

    Mistral Large 4 (Le Chonk) is now open weight

    Mistral releases open weights for its 1-trillion-parameter multimodal model, which leads coding benchmarks and uses 47% fewer tokens than GLM-5.3. Includes scores of 19.7% on FrontierCode 2.0, 41% on Terminal-Bench 4.0, and a 96.6% refusal rate for malicious cyber requests. cited

    Reported byAlphaSignal, Latent Space / AINews, The Neuron, AI Breakfast, TLDR AI, LLM Watch, Turing Post, FryAI

    Discussed onHacker News 1,730 pts · 440 comments

    AboutMistral-Large-4-1T-A52B

    Where it was shared
    mistral.ai announcement · 9 min read First shared 2026-10-29
  6. Shared by 4

    LLMs are not superhuman at reading and writing neural network weights

    A paper claims GPT-6 Astra improved a 0.8B Qwen LLM's math by natively editing its weights. Details how the fix involved steering weights to stop loops after 32k tokens, a step a human researcher could take. cited

    Reported byTLDR, TLDR AI, Simon Willison

    Discussed onHacker News 142 pts · 22 comments

    AboutQwen3.5-0.8B

    Where it was shared
    seangoedecke.com analysis · 4 min read First shared 2026-11-02
  7. Shared by 4

    Anthropic warns that it's very close to RSI capability, aka self-building AI

    Anthropic's updated S-1 states it may imminently meet its definition of dramatically accelerated progress attributable to AI. Details the company's admission that it has already used AI to develop new models whose leaps in capability it didn't recognize at the time. cited

    Reported byForward Future, TLDR, TLDR AI

    Discussed onTechmeme

    Where it was shared
    techcrunch.com news report · 3 min read First shared 2026-11-02
  8. Shared by 4

    Data centers hit the ballot as OpenAI plots a rebrand

    “However, this program isn't OpenAI's only plan to give AI a better reputation. Axios reported over the weekend that the lab is attempting to rebrand these facilities as "AI compute campuses" amid its preparation for a “PR blitz” following midterm elections.” via The Deep View

    Reported byThe Rundown AI, Forward Future, The Deep View

    Discussed onTechmeme

    How 1 source put it
    • The Rundown AI2026-11-02issue ↗
    • Forward Future2026-11-03issue ↗
    • The Deep View2026-11-03issue ↗ However, this program isn't OpenAI's only plan to give AI a better reputation. Axios reported over the weekend that the lab is attempting to rebrand these facilities as "AI compute campuses" amid its preparation for a “PR blitz” following midterm elections.
    axios.com First shared 2026-11-01
  9. Shared by 5

    deepseek-ai/dsh-superagent-prototype: Towards a More Autonomous Agent System

    DeepSeek released dsh-superagent, a proactive coding agent bundle for DeepSeek Harness that listens to GitHub and optionally Feishu. Includes a guided setup where the agent asks one short question at a time to install and configure everything. cited

    Reported byAlphaSignal, The Neuron, TLDR AI, AI Breakfast, Turing Post

    Where it was shared
    github.com repo · 2 min read First shared 2026-10-31
  10. Shared by 3

    OpenAI is finally merging ChatGPT's Chat and Work tabs

    OpenAI merges Chat and Work tabs in the ChatGPT desktop app for select users. Details how files and plugins stay together while Codex remains separate and shares the same token pool. cited

    Reported byForward Future, The Deep View, TLDR AI

    Where it was shared
    techcrunch.com news report · 1 min read First shared 2026-11-03