What matters in AI, and why.
The one that matters
After a week as the industry's lone holdout, Anthropic publishes its position: it says it has never backed an open-weights ban, defends chip export limits to China, and calls for global model testing.
Why it matters
Open weights are the substrate of local AI and self-hosting; where the most safety-cautious frontier lab draws this line will shape what regulation targets next.
Learn one thing
View full-size diagram ↗What it is
In-context learning is a model's ability to pick up a task from examples placed in the prompt itself — no training run, no weight updates. 'Few-shot prompting' is the everyday form: show two or three worked input-to-output pairs, then a new input, and the model continues the pattern.
How it works
Attention lets the model read earlier prompt tokens while predicting the next one. Circuits that researchers call induction heads find repeated structure — 'last time A was followed by B' — and copy it forward. Demonstrations mostly help the model locate a task it already knows rather than teach it new knowledge.
A concrete example
Ask a model to turn 'March 5th, 1999' into a date code and you may get any format back. Show it two examples — 'July 4th, 1776 -> 1776-07-04' and 'Jan 2, 2001 -> 2001-01-02' — and it emits '1999-03-05'. The examples pinned down format, delimiter, and zero-padding in one shot.
Why it matters
Few-shot examples are the cheapest adaptation lever you have: they beat instructions when a task is easier to show than tell — format conversions, house style, fuzzy labeling rules. They cost context tokens every call, so long example blocks pair naturally with prompt caching and long context.
What people get wrong: That the model 'learns' from your examples the way training does. Weights never change — and Min et al. (2022) showed models do nearly as well with randomly wrong demonstration labels: the win comes from showing format and label space. Task recognition, not task learning.
Today's top 5
The last holdout lab put its policy in writing — this document is what open-model regulation fights will cite.
A frontier-class open release with a real technical report — the strongest local-model option shifts again.
Post-Hugging-Face-breach, AI security is becoming an ecosystem play — and notably without OpenAI, Google, or Anthropic.
Concrete evidence of week-scale autonomous coding — and a price tag that reframes what agent labor costs.
A rare hands-on check of a headline AI-coding claim — worth reading before repeating anyone's rewrite numbers.
A concise daily briefing on the AI developments worth understanding. Subscribe on the homepage.