RLHF Co-creator: LLMs Are Built for Assistance, Picking the Right Task Is Key to Automation
FinanceYF5 · x · 2026-11-12
At AI Engineer Code Summit, @CompleteSkeptic — who co-created RLHF/ChatGPT at OpenAI and now builds Jev — argued that today's LLMs are optimized for assistance (human in the loop), not automation: "95% Einstein, 5% Mr. Bean" is fine for a chatbot but fatal for back-office decisions. His "Bitterest Lesson": compute beats algorithms, data beats compute, and the bitterest part is picking the right task. Slides invoked the "jagged technological frontier" (Dell'Acqua, Mollick et al. 2023) — AI's capability boundary is jagged, with seemingly equal-difficulty tasks straddling it — and framed instruction following, Copilots, Claude Code as "too good to be true" assistance, while reliability-requiring automation (customer service, accounting, data entry) remains "too bad to be useful." Jev's bet: not bigger, not cheaper, but reliable.
Related event: RLHF co-creator: LLMs are for assistance, software will make a comeback(3 posts)→
More from AGI Musings
- He tried to prove Crunchbase misses deals — it covered 41 of 45, logged 1 day early — YvesMulkers · 2026-11-12
- Critics slam Amodei's cure-all prediction: AI bubble draining investment from real biotech — AlexTensor · 2026-11-12
- Clara Shih: junior workers face the worst job market in 37 years — clarashih · 2026-11-12
- Managed agents platforms are the new chat-with-a-PDF, says developer — intellectronica · 2026-11-12
- He gave up Harvard Law and an Anthropic offer to fight AI takeover — and got mocked — AaronBergman18 · 2026-11-12
- Vertex exec calls AI-cures-all claims "complete BS": 50 AI firms, zero working drugs — AlexTensor · 2026-11-12