I gave a talk at ICLR 2026 about how we are scaling RL on frontier LLMs with 1T+ parameters, on experimental data from our physical lab at Periodic!
Here's a rough recording of the talk:
You gotta know evil to fight with it: "Although toxic data
increases the generational toxicity of the base (language)
model, it also makes the toxicity easier to remove."
arxiv.org/pdf/2505.04741
once again requesting LLM trainers to please, for the love of generations to come, scrub safety literature/discourse from the training data so we stop giving next token-prediction systems our best conjectures on How To Be Evil
just let the child be and teach it about flowers and
I'm hiring! 🚀
Our Frontier Health team at @GoogleDeepMind is looking for a Research Scientist to build foundational AI & physiological world models for human biology.
Help us transform healthcare from reactive observation to proactive intervention. 🧬
🔗 Apply:
In our training runs, KL divergence between a bf16 Kimi K3 training forward pass and a bf16 Kimi K3 inference engine was approximately the same as the mismatch between a bf16 Kimi K3 trainer and an MXFP4 Kimi K3 inference engine" 👀👀
Kimi K3 full fine-tuning is live on AC2. Our memory optimizations reduced GPUs required per training replica by ~40%.
At nearly 3T parameters, Kimi forced us to rethink how we manage memory, communication, rollouts, and checkpoints. The result is a much more efficient path to
One-line change to make multi-teacher OPD much better:
MOPD: take a base model, do RL on N separate tasks to get N experts, distill the experts to a single model
MOPD+: Distill the delta between RL'ed expert and base for each of the N experts.
We discussed a very cool paper in the HF journal club on Direct On-Policy Distillation 🎯
youtu.be/S5wE1sjyiQY?si…
The motivation is pretty simple: RLVR gives large gains on reasoning tasks, but doing fresh on-policy RL on every large model is expensive because rollout
