Researcher Accuses Laya Author of Overstating Work as Jev-Like System
Paimaamu · x · 2026-09-20
A technical dispute over research hype: @shantanugoel argues the Laya author is cleverly (or naively) overstating and broadening claims, since their 2025 work was not "Jev/Jev-like" at all.
- Jev is not merely "a classifier + RL": it takes arbitrary state plus typed questions (yes/no, choice, score) and returns calibrated probability distributions in parallel; RLCD trains for calibrated decisions and the same model generalizes across task types.
- Their sales RL agent was just embeddings → small network → PPO → a single 0-1 sales conversion probability, far from Jev's generality, and not novel.
The quoted @aarondfrancis adds a different angle: doing cool work isn't enough if you don't tell people about it.
More from AGI Musings
- Oliver Curry argues Effective Altruism never gave a good reason to be altruistic — mjdramstead · 2026-09-20
- Security veteran to Hinton: centralized AI is far more dangerous than open source — nptacek · 2026-09-20
- Beliefs are shaped by local incentives, not rational analysis: when to trust people's convictions — nabla_theta · 2026-09-20
- Polymarket debate: why Jensen Huang's AGI skepticism isn't actually suspicious — nabla_theta · 2026-09-20
- Blogger: 'Autonomous' AI Agent Hype Distracts From What Are Plain Felonies — mjdramstead · 2026-09-20
- Reddit debate: We're building AI agents at the wrong scale — aim for persistent AI organizations — Druss_ · 2026-09-20