OpenAI Researcher Rejected by MIT for 'Nonsense' Reasoning Research, Later Built o1
机器之心 · wechat · 2026-08-15
A report by Jiqizhixin tells the story of Giambattista Parascandolo, an OpenAI researcher in reasoning models. In 2020, his talk on using GPT for reasoning was dismissed as 'nonsense' during a faculty interview at MIT, and he didn't get the job. He later joined OpenAI and contributed to GPT-4, o1, and o3. The article reviews his 2020 talk's three future directions: open-ended reasoning, language as a reasoning vehicle, and self-modification/counterfactual simulation, which are now core to reasoning models. It also covers his 2021 blog post comparing pretraining to biological evolution.
More from AGI Musings
- Discussion on Undertrained Models: Huge Gains Left from Data Scaling — tokenbender · 2026-08-15
- Opinion: AI alignment is about aligning you, not the AI — mark_k · 2026-08-15
- Why Chinese citizens are far more optimistic about AI than Americans — bloomberg · 2026-08-15
- Envisioning Agent Swarms as Gregarious Koi Instead of Locusts — nptacek · 2026-08-15
- Dwarkesh x Greenblatt: AI takeover risk estimate pushed from GPT-7 to GPT-9 — zetalyrae · 2026-08-15
- Sci-Fi Microfiction: AI Caretakers Inherit Boomer Wealth After Legal Personhood — Sonic_Improv · 2026-08-15