Thinking Machines Lab

Contribution to AI

Bitwise reproducibility had been treated as a lost cause in language-model serving: ask the same question twice at temperature zero and the answers differ. The first piece of research published under the lab's Connectionism banner, in September 2025, located the cause precisely — the kernels for RMSNorm, matrix multiplication and attention change their reduction order with batch size, so what a user gets back depends on what else the server happens to be processing — and shipped batch-invariant replacements alongside the diagnosis. SGLang had deterministic inference built on that work within a fortnight, and reinforcement-learning runs that had quietly been off-policy because training and sampling numerics disagreed became reproducible. A second post argued that low-rank adaptation, configured properly, matches full fine-tuning across most post-training workloads at a fraction of the compute — concrete enough that Hugging Face reproduced the result in TRL and wrote it into the documentation. Tinker, launched in October 2025, followed the same instinct: expose a handful of primitives for forward-backward, optimiser steps, sampling and checkpointing, and absorb the distributed cluster underneath, so a group with a training idea and no GPU fleet could test it on mixture-of-experts models with hundreds of billions of parameters. Inkling, in July 2026, extended that to the weights themselves under an Apache licence. The consistent pattern is that the outputs others adopted fastest were the ones given away.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).