Up to date
Model Follower / Deep dives
Long reads on how open models are built, measured and served. Each one keeps its data, its figures and its animations; the figure stays in view while the text steps through it.
A chatbot writes an answer. A decision model picks one. An animated explainer of typed decision models for people who buy inference: what they are, how three kinds are built, what we measured, and where they slot into a serving stack.
A decision model reads text once and returns a probability for each answer you allow. In three and a half weeks the category went from one vendor to open weights that top the independent board. What that means for inference spend and latency, what our studies measured, and what to do about it.
Claude and Gemini now mark the text they write by tilting which words they pick. We built the same kind of mark with our own key, put 150 web pages through 18 kinds of handling, and measured what survives.
Short pages and planned pieces. Most take up something the decision-models briefing cut.
Three weeks after Jev, there are two public leaderboards, a pile of single-job tests and a few honest calibration studies, and they disagree. What each one measures, what the Decision Index currently says, and what we use instead.
Every model we tested on real agency work, through one method, with what each one can handle alone at 90% accuracy and what it costs per 1,000 decisions.
On Oct 4 we wrote down what each study would show and what result would change a decision. Here is what happened, including the ones that failed.
A decision request is prefill plus a readout, with no decode. What that means for batching, hardware fit, pricing and the engines that now speak /v1/systemone.
One distribution, four "confidence" numbers. Why thresholds do not transfer between models, formulas or question wordings, and how to fit your own.
