Signals for 25 May 2026

Drie strikte signalen over agent-discovered reasoning, AI-builder workflows en modelstackkeuzes.

Researchers let Claude Code discover AI scaling algorithms that humans probably wouldn't have designed

The Decoder

Researchers from UMD, Google, Meta, and other institutions used AutoTTS to let a coding agent independently discover control algorithms for AI reasoning. The found algorithm cuts compute by about 70 percent compared to standard self-consistency while matching its accuracy. Relevant because real AI implementation depends on evaluation, reliability, and understanding new failure modes.

#agent #builder #evals #research-evals

Mad House — Usborne Creepy Computer Games

Simon Willison

Simon Willison used Claude to recreate a 1983 Usborne computer game from a PDF as an interactive JavaScript and HTML artifact. Relevant because the builder layer around AI is becoming more concrete: tools, runtimes, and developer workflows increasingly determine the real leverage.

#tooling-runtime

Anthropic may keep supplying Claude to the NSA despite being flagged as a supply chain risk by the Pentagon

The Decoder

Anthropic may keep supplying AI models to the NSA despite being labeled a supply chain risk, partly because its model reportedly runs on older hardware. Relevant because model choice is increasingly an architecture question around cost, context, latency, hardware, and control.

#evals #models-architecture