AI Genius Predicts the Next Critical Tipping Point
Joe Lonsdale and Vivek Gopalan talk with Rayan Krishnan, founder of Vals AI, about independently testing AI models, tracking progress toward recursive self-improvement, Chinese model risks, and why he stays optimistic.
The gist
-
Krishnan says reading Zero to One drew him to Palantir, where as an intern he used CLIP to build a reverse image search for Gotham.
-
Krishnan argues labs shouldn't grade themselves and direct regulation slows progress, so a market of independent evaluators, possibly government-licensed, should test models.
-
Vals AI's RSI index, comparing models to human researchers, projects Anthropic reaching full recursive self-improvement by July or August 2027, with Gemini rising fastest.
-
In Vals AI testing, closed US models beat Chinese open models at decompiling code and multi-turn chats, yet GLM-5.3 showed little pro-China bias.
-
Krishnan stays optimistic, citing labs agreeing with the White House on embedded evaluators, and AI taking away busy work while unlocking new discoveries.
"Around the same time, while I was at Microsoft, I read Zero to One, and it became increasingly clear that people at Palantir were trying to construct the technology in a way that protected civil liberties or Western democratic ideals at large."
▶ Play at 3:375 creators recommend it, including David Senra, Mark Pincus, and Matan Grinberg.
See all on the Zero to One page"Essentially, I'd done some work with CLIP, this joint image-text embedding model."
▶ Play at 4:302 creators recommend it, including Will Marshall.
See all on the CLIP pageWhat they recommend · 2 picks
"Around the same time, while I was at Microsoft, I read Zero to One, and it became increasingly clear that people at Palantir were trying to construct the technology in a way tha..."
"Essentially, I'd done some work with CLIP, this joint image-text embedding model."
Also mentioned
Related episodes
Episodes covering similar topics
More from Joe Lonsdale