PostHog's Jeeves: a 9B decision model scoring 0.935 on JevBench
petrusenko_max · x · 2026-09-29
PostHog introduced Jeeves, a 9B Jev-like decision model built on Qwen3.5-9B with a LoRA adapter, pointer head, and a block-4 diffusion drafter, trained via SFT and CISPO. It scores 0.889 on held-out test data vs 0.822 for Kev-9B and 0.857 for Jev, and 0.935 vs 0.866 on JevBench public tiers.
More from Fun
- Polymarket Puts 10% Odds on OpenAI Announcing Another Full Training Pause — Polymarket · 2026-09-29
- Opus 5.5 Generates a Gorgeous Scrollable 3D Explainer on Chipmaking — threepointone · 2026-09-29
- The Zvi on 'Doomtrolling': lying about your LLM's danger to juice stock price has consequences — TheZvi · 2026-09-29
- Short code demos are deliberate practice, creator says in defense of AI experiments — techartist_ · 2026-09-29
- Sci-fi novel foresaw AIs trading compute to evade legal intelligence caps — neuroecology · 2026-09-29
- Yacine says caffeine has a Ballmer Peak: past it, you procrastinate again — yacineMTB · 2026-09-29