25 Aug 2026

Digest: Ngo on alignment's capabilities legacy, AlphaEvolve math advance

1. What just happened? Pragmatism and Pessimization Recommended

   Richard Ngo — an alignment researcher formerly at OpenAI and Google DeepMind — published the second installment of his retrospective on the field, and it is the sharpest part yet: a documented, name-by-name history of how 'pragmatic alignment' work advanced frontier capabilities at OpenAI, DeepMind and Anthropic, from RLHF engineering that was indistinguishable from what 'a prescient capabilities-maximizer would have been doing', to scaling GPT-2/GPT-3 justified internally as enabling safety research, to WebGPT — staffed by alignment-motivated researchers — serving as the direct predecessor to ChatGPT. The community is treating it as a major document (roughly 300 karma within a day, far above a typical frontpage post), but the top comment carries serious pushback: alignment researcher Jan Kulveit argues that the high-integrity strategy Ngo prescribes is harder to live than to preach — nobody assigns credit for capability ideas deliberately left unpublished (his own group's restraint cost it funding and people), and the field's very exemplars of integrity, Daniel Kokotajlo and Ngo included, earned their platforms precisely by working at the labs first. Ngo's reply partly bites the b

   Richard_Ngo via LessWrong
2. Google DeepMind and university researchers push the frontier of matrix multiplication using AlphaEvolve

   Researchers from Google DeepMind, Carnegie Mellon, Columbia and MIT report an improvement to the theoretical bound on the matrix multiplication exponent — the constant governing how fast matrices can in principle be multiplied, one of theoretical computer science's central open quantities — achieved first via a new optimization approach and then pushed further by using DeepMind's AlphaEvolve to evolve the optimization algorithm itself. Where AlphaEvolve's earlier headline results found faster concrete algorithms for small matrix sizes, this moves a frontier theoretical bound, adding to the growing pattern of AI systems contributing to research-level mathematics rather than just competition problems.

   arXiv
3. Researchers unveil SPADE, a self-play framework for automatically generating training environments for LLMs

   A multi-university team (University of Washington, Stanford, MIT, CMU) introduces SPADE, a self-play framework in which an LLM alternates between designing executable training environments and solving them, letting a model bootstrap diverse synthetic training data for itself; on Qwen3 models up to 30B — mid-size open-weight models well below the frontier — it improved game and tool-use benchmark scores over fixed-environment baselines. The gains are modest and small-scale, but environment design is one of the main human bottlenecks in agentic RL training, and automating it is a concrete ingredient of self-improving training pipelines; Jack Clark's Import AI 470 flags it as one of the week's notable results.

   arXiv