The lab's research blog, unusually load-bearing for its reputation before any model shipped. Notable posts: Defeating Nondeterminism in LLM Inference (September 2025, with the 1K+-star batch_invariant_ops release, tracing nondeterminism to batch-variant kernels), LoRA Without Regret (when LoRA matches full fine-tuning), On-Policy Distillation, and Modular Manifolds (optimizer geometry). The posts read as the methodological groundwork for Tinker and Inkling's post-training stack.

researchpost-traininginfrastructure

Related