Gui Valente — Signal on AI

Wed · April 29, 2026

◆ Signal · 04 · 29

The interesting story this week isn't another model release — it's that three separate labs quietly stopped publishing their evals. Capability is moving inside again. Read →

◆Today · Headline Filed 09:12 ET

USAISI hearing room with parallel placards for OpenAI and Anthropic, papers on the witness table.

Anthropic and OpenAI will share frontier-model safety cases with USAISI — but on different terms.

Both labs signed parallel agreements giving the U.S. AI Safety Institute pre-release access to frontier evals. The fine print is the story: one side reserves the right to redact training data; the other doesn't.

The Human Perspective — Gui

"Pre-release access" is the wrong frame. What just happened is that the U.S. government has, for the first time, a legible way to ask what does this model know before it ships — and two of the three labs that matter agreed. The third will be the story by July.

5 sources · 4 min read 14 outlets · 3 framings

§Substack · Latest essay Pub. 04 / 27

Late-night research lab — a single analyst at a workstation, colleagues meeting beyond glass under a 'CONFIDENTIAL: PHASE CHANGE' whiteboard.

The labs are getting quieter, and that's the most interesting thing about them.

A long read on capability secrecy, the slow death of the public benchmark, and what it means that the model we're talking about is no longer the model that's shipping.

From the essay

"For ten years the field competed in public. The new pattern is to compete in private and announce only the conclusions. We should treat that as a phase change, not a press strategy."

2,400 words · 12 min Read on Substack ↗