We’ve developed the world’s most complete megakernel framework. It can perform entire training runs, serve inference, play video games, and even do all three simultaneously in a single kernel.

I left @AnthropicAI and joined @flappyairplanes

you can fit fruitfly Doom RL rollouts in the kernel too btw

Unlike assistant models trained with verifiable or preference reward, Ampersand was trained to understand the long term consequences of its actions through simulation feedback ampersand.humansand.ai/blog

We are aware that codex is down and are working hard to bring back normal service.

Our first TPU megakernel for Kimi K3 reaches 709 tokens/s on low-concurrency decode, against 450 tokens/s for our GB200 baseline, both with DSpark speculative decoding. To our knowledge, this is the first TPU inference megakernel. The whole model runs in a single Pallas kernel, and without spec decoding it is roughly 1.4 to 2x the GB200 baseline at batch sizes 1 through 8. We are open sourcing it today. 1/2

I implemented a system in @appliedcompute’s platform for automated failure mode clustering with Jev to surface errors at an even larger scale than before. RL training produces billions of tokens in traces. I always manually read many traces to understand model behavior, but finding agent failures (like reward hacking / hallucinations) at scale is easy to miss without automation. Here’s how it works:

636606729769440499166579950236036751749912014371509557713570027508971809534551913252252094954941974952859310861988904737359709200557919 is a factor of RSA-896 saweis.net/posts/rsa-896.htm…

Why is RL so sensitive to training-inference mismatch (TIM)? With each training step, the trainer is pulled toward the sampler, resulting in accumulating drift. We exploit this intuition to devise a novel correction method to stabilize RL under TIM, called Score Centering.

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. Streaming the run: mimo.xiaomi.com/rl/

We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so. Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training. You can read the full post here: darioamodei.com/post/we-must…