What matters in AI, and why.
The one that matters
Sources say Meta is in talks to lease data-center capacity to Anthropic in a deal worth around $10B over two years, as it hires AWS's outgoing compute chief to run its build-out.
Why it matters
It reframes Meta from a frontier-model competitor into a compute landlord, echoing the capacity deals xAI and SpaceX have struck, and hints demand for its own models is lagging its data-center spend.
Learn one thing
View full-size diagram ↗What it is
A reinforcement-learning recipe for fine-tuning language models that removes PPO's separate value (critic) network and instead judges each answer by comparing it against a batch of other answers to the same prompt.
How it works
The runtime samples several answers to one prompt, scores each with a reward signal (often an automatic verifier for math or code), then standardizes each answer's reward inside its own batch (subtract the batch average, divide by the spread) to get its advantage; a clipped policy-gradient step plus a KL leash toward a reference model updates the weights, with no learned value head.
A concrete example
DeepSeek popularized GRPO to train its math and R1 reasoning models with verifiable rewards, where a checker marks each solution right or wrong, and the batch-relative baseline let them scale RL more cheaply than PPO.
Why it matters
Dropping the critic removes one of the largest and most finicky pieces of an RL stack, cutting memory and instability, a big reason open labs can run frontier-scale reinforcement learning on reasoning models.
What people get wrong: That it is simply 'PPO minus the critic.' The real shift is where the baseline comes from, peer answers rather than a trained value function, which is why GRPO leans on sampling and scoring several outputs per prompt.
Today's top 5
Hands-on evals, not just vendor benchmarks, are what tell builders whether an open-weights model is worth self-hosting.
Decoupling the conversational layer from the reasoning model lets voice stay fast while inheriting every frontier upgrade.
A concrete case of a multi-agent system driving a real scientific loop from hypothesis to validated result, not just text.
Consumer platforms are starting to expose agent-first interfaces, not just apps, as a way to reach users.
China is pairing open-weights model releases with diplomacy to make its AI stack the global default.
A concise daily briefing on the AI developments worth understanding. Subscribe on the homepage.