TypeSafe launches Jev, a judgment-only model 200x faster and 400x cheaper, to rethink RLHF
oran_ge · x · 2026-09-22
TypeSafe released Jev, its first model that outputs only calibrated judgments instead of text — 200x faster and 400x cheaper than frontier LLMs. CEO Diogo Almeida's deeper thesis, laid out in a pre-launch talk:
- Peak capability, trough trust: enterprises let AI solve math but not customer decisions; human-in-the-loop software can't own outcomes or price on results.
- RLHF is the root cause: optimizing for human approval keeps a human in the loop, precluding full automation.
- Side effects: sycophancy (praising a fart sound as 'atmospheric', Gemini worse) and hallucination (hesitation penalized in reward models).
- Automation requires models that make decisions and take responsibility, not models that please people.
More from AGI Musings
- "AI is only trained on human data" is outdated: RL exploration now drives frontier training — aran_nayebi · 2026-09-22
- AI Slop Is a Productivity Tax: Execs Say Enterprises Are Drowning — sebpaquet · 2026-09-22
- It Was Never About Coding: A Veteran Engineer's Take on the Agent Era — blaizedsouza · 2026-09-22
- Veteran Dev: Best Programmers Have Language Degrees, Don't Outsource Writing to AI — blaizedsouza · 2026-09-22
- Vinod Khosla: the right agent won't be an app — users will live in the chat with their agents — petergyang · 2026-09-22
- Math Isn't Solved: AI Turns a Universe We Walked Into Wormholes — ignite_intelligence · 2026-09-22