Jev swaps text generation for instant calibrated decisions, plays Doom at 7 req/s
JeremyCMorgan · x · 2026-09-23
A breakdown of TypeSafe's Jev model: instead of generating text token by token, it takes a state (text/JSON) plus named questions and instantly returns calibrated numerical answers. Three question types: choice (pick from a list, returns probabilities + confidence), score (position on an ordered rubric), and noul (probability 0–1 that a claim is true). The pitch is speed and reliable calibration — traditional LLMs are orders of magnitude slower and their self-reported confidence rarely matches reality. In the demo, Jev predicts button presses in real time at 7 requests/second to navigate and clear Doom, something text-generation LLMs can't do. Includes a horse-race handicapping API example.
More from coding & agent
- Stripe explains how agents detect and use checkout via WebMCP — jeff_weinstein · 2026-09-23
- Stripe demos WebMCP integration: a deep dive through a checkout page's new tools — jeff_weinstein · 2026-09-23
- Building AI agent harnesses that improve across 6 editable control dimensions — MaryamMiradi · 2026-09-23
- Drop launches: rootless Linux sandbox that isolates coding agents at the OS level — aronchick · 2026-09-23
- Sentry Bets Big on Evals: 1,800 PR-Triggered Runs, 20-50% Token Cost Cuts — zeeg · 2026-09-23
- MiMo-V2.6-Flash Tool-Call Failures Are vLLM Bugs: Empty Streaming Replies, Lost Reasoning, Hidden 2048-Token Cap — mamolengo · 2026-09-23