The Hallway Track
12 tracked signals on distillation.
Competing open letters expose industry fracture over open-weight AI models and development pacing
“We request that the U.S. government support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development.”
OpenAI disrupted a coordinated adversarial campaign to extract protected model reasoning via distillation
DeepSeek 4 Pro achieves near-frontier quality with MIT-licensed open weights, pressuring closed AI labs.
“DeepSeek has MIT licensed open weights. Anyone can run the exact same model at their own price.”
China commits to AI openness as Xi's WAIC speech and Qwen's open-weight pivot signal strategic shift
“Xi gave his speech where he directly committed to openness and open source as a strategy.”
Ben Thompson proposes US law legalizing AI training data collection and distillation to compete with China
“The U.S. should pass a law that (1) makes explicit that collecting data for training models is fair use, and (2) bars terms of service that forbid distillation, for U.S. companies at a minimum.”
2026 frontier post-training has shifted to Multi-Teacher On-Policy Distillation (MOPD), merging many specialist models into one.
“The shape of a post-training recipe has changed more in the last year than in the prior three.”
Distillation is structurally unpreventable because all models converge toward a shared platonic representation of intelligence
“distillation is effectively unpreventable. Because once you've found the shape of intelligence, you know what it looks like.”
Microsoft Foundry adds reinforcement learning and a low-level training API to turn production agents into cheaper, faster models.
“Think of it as PyTorch as a service.”
uRun's Helios model generates video at 1/100th the cost of frontier models with comparable quality
“at least 40 models with real-time capabilities and long horizon generation capabilities released this year”
On-policy distillation uses RL policy shift signals to train student models without full RL cost
“the student might be stronger than the teacher in this case”
Hugging Face released LFM2.5 Q4_0 checkpoints via quantization-aware distillation
Applied Compute is bringing continual learning to enterprises via a distillation spectrum from offline to online
“this is sort of the holy grail of continual learning where I have a model that's serving production traffic, it does a rollout, it creates a trace, we figure out how to learn from that trace, we update the model”