What matters in AI, and why.
The one that matters
The White House OSTP director says Moonshot AI distilled Anthropic's Fable to build Kimi K3, and the Treasury secretary put sanctions on the table for 'distillation attacks'.
Why it matters
If Washington starts treating distillation as sanctionable IP theft, the open-weights ecosystem builders rely on - much of it Chinese - could change shape fast.
Learn one thing
View full-size diagram ↗What it is
The knobs that control how a model turns predictions into a reply. At each step it assigns a probability to every token in its vocabulary; sampling parameters decide how that table becomes one picked token - locked to the favourite at one extreme, happy to gamble at the other.
How it works
Temperature rescales the model's scores before they become probabilities: low values sharpen the distribution, high values flatten it, zero always takes the top token. Top-p keeps only the smallest set of tokens whose probabilities sum past p and samples inside that nucleus, which resizes itself with the model's confidence. The dice are rolled again for every token.
A concrete example
Ask twice for a weekend project idea at temperature 1.0 and you usually get two different answers: the first token already forks ('Build...' vs 'Try...') and each fork compounds. At temperature 0 a coding agent's retries become nearly reproducible, so agent frameworks pin it low while brainstorming tools turn it up.
Why it matters
A lot of 'flaky LLM' debugging is just sampling doing its job. Use low temperature for extraction, classification and tool calls where consistency matters, and higher settings where variety is the point. It also changes evals: single runs at temperature 1.0 mostly measure luck, not prompts.
What people get wrong: That temperature 0 makes a model deterministic and more truthful. GPU batching and floating-point rounding still cause nondeterminism unless the serving stack is batch-invariant, and temperature reshapes confidence, not knowledge - a wrong fact stays wrong at every setting.
Today's top 5
A new front in the US-China AI fight aimed directly at open-weights models.
A serious hedge against Nvidia dependence at gigawatt scale.
A frontier-scale open-weights promise, with the key specs still withheld.
The incident is becoming a reference case for agent containment failures.
Enterprise agents are moving from demos into core business systems.
A concise daily briefing on the AI developments worth understanding. Subscribe on the homepage.