Ex-OpenAI Researcher Jerry Tworek: RL Not the Key to AGI, Need New Architecture
机器之心 · wechat · 2026-08-05
Former OpenAI researcher Jerry Tworek, once a strong believer in RL, now argues it won't lead to AGI. He co-founded Core Automation with ex-Google Gemini pretraining lead Rohan Anil to explore architectures that enable continual learning after deployment. The article analyzes limitations of in-context learning and fine-tuning, and proposes meta-learning and deeper computation as directions.
More from AGI Musings
- Berkeley Paper Explores How Goal-Directed Mechanisms Affect AI and Human Behavior — berkeley_ai · 2026-08-05
- Predicts AI Will Replace All Jobs by 2028, Calls for Universal High Income — davidpattersonx · 2026-08-05
- Charity Majors: AI Demands More Engineering Discipline, Not Less — mipsytipsy · 2026-08-05
- Opinion: AI Agents Haven't Taken Off Because They Lack a 'Desktop' UI Metaphor — Scobleizer · 2026-08-05
- Thought Experiment: Could AI Trained Only on Pre-1940 Text Invent the Microwave? — chkbd1102 · 2026-08-05
- Developer Insight: Pivot from Agentic Payments to Compute Loansharking — arthurcolle · 2026-08-05