Jev: InstructGPT coauthor Diogo Almeida on System One models for prod, not God
Latent Space · rss · 2026-09-22
Latent Space interviews Diogo Almeida, TypeSafe AI CEO and InstructGPT coauthor, whose Jev launch video hit 40M views. Key points:
- Since InstructGPT, innovations like function calling, structured outputs and reasoning have been patches on the string-based seq2seq paradigm; optimizing for human raters (RLHF) breeds hallucination and sycophancy, while RLVR exacerbates jagged intelligence.
- Jev's core innovation is RLCD (Reinforcement Learning for Calibrated Decisions): optimizing epistemically honest probabilities on System One tasks — models built for composable software, not chat.
- The "Bitterest Lesson": tasks and data beat compute; TypeSafe is a data lab, rejects public benchmarks, and optimizes intelligence per dollar.
- Build AI by decomposing workflows into small measurable decisions, replacing giant prompts with structured state; AI should disappear into software like regex.
- Use cases: games/computer use, voice + browser control, the Doom demo, coding agents (linting, tool-call compaction — see his "Tyranny of the KV Cache" note), dark data entity resolution, natural-language search, and AI-as-judge.
Also covers why he wouldn't pre-train with $1B, why most neo-labs get AI wrong, the "inverse SaaS-pocalypse," and contrarian scaling-law views.
More from coding & agent
- Foundation Capital: enterprise agents need identity, boundaries — and the web rebuilt for machines — brucemacv · 2026-09-22
- Question's Gambit lifts deep research agents: GPT-5.5 hits 90.5% on BrowseComp-Plus — omarsar0 · 2026-09-22
- Open-Source Tool Maps E-Drums to Ableton Kits, Designed for Coding Agents to Extend — _Dave__White_ · 2026-09-22
- Chollet: delegate coding, never delegate understanding — beware 'conceptual debt' — round · 2026-09-22
- Dev open-sources modsure npm package for LLM-powered social media moderation — RealDannyhvv · 2026-09-22
- TypeSafe's 2-stage extraction cascade: big-model quality at a fraction of the cost — TheMoonMidas · 2026-09-22