JEV, the 'decision model' that writes nothing, ships — but 193x speed claims need discounting
大模型之路 · wechat · 2026-09-28
TypeSafeAI opened early access to JEV, a 'decision model' that skips autoregressive decoding and returns calibrated yes/no, choice, and score answers in 70-500ms. Founder Diogo Almeida is an InstructGPT co-author back with a $40M seed round; JEV hit 1,800+ points on Hacker News. The claimed 193x speedup and 444x cost saving come from self-run benchmarks vs. averaged frontier-model answers — independent tests land at 5-25x. Trained with RLCD for calibrated confidence, it suits high-frequency agent judgments (routing, moderation, risk gating) but fails at writing, reasoning, and auditability, and accuracy degrades with noisy context.
Related event: TypeSafe's Jev: A Text-Free Decision Model That's Fast and Dirt Cheap(6 posts)→
More from coding & agent
- Setting /autocompact to 400k saved 29% of weekly Claude usage without hurting performance — rickasaurus · 2026-10-10
- Eazo hands-on: one-prompt apps with 6 design variants, Stripe payments, 50-70% referral cuts — vista8 · 2026-10-10
- Nested parallelism: running Hermes agents inside Grok bot VMs for compute offload — alexcovo_eth · 2026-10-10
- Brian Holt sold his 42U homelab — a Framework Desktop plus coding agents does more — film_girl · 2026-10-10
- Group-Evolving Agents wins COLM 2026 award, hits 71% on SWE-bench Verified — xwang_lk · 2026-10-10
- Asking an agent to fix a bug you don't understand is continuous paperclip maxxing — brandon_xyzw · 2026-10-10