🤖Latent Space: InstructGPT Coauthor Diogo Almeida on Jev
TL;DR
Diogo Almeida, an InstructGPT coauthor and TypeSafe AI CEO, explains why his Jev 'System One' models skip chat entirely. Jev trains with RLCD, which rewards calibrated probabilities instead of human preference, so AI can sit inside software like regex.
Diogo Almeida, an InstructGPT coauthor and TypeSafe AI CEO, explains why his Jev 'System One' models skip chat entirely. Jev trains with RLCD, which rewards calibrated probabilities instead of human preference, so AI can sit inside software like regex.
Key Points
Recorded Sept 21, 2026 with swyx; Jev's launch video passed ~40M views
RLCD (Reinforcement Learning for Calibrated Decisions) is unpublished and targets honest confidence scores
Almeida argues RLHF causes sycophancy and RLVR worsens jagged intelligence
TypeSafe rejects public benchmarks and optimizes for intelligence per dollar
He says he wouldn't pre-train even with $1B, calling TypeSafe a data lab rather than a model lab
Why It Matters
If calibrated small decisions beat giant prompts, the next wave of AI engineering looks like classic software with learned functions, not chatbots.
Quick Facts
Comments
Be the first to comment
Enjoyed this article?
Get it daily. 7am. Free. Reads in 5 minutes.