ChatGPT Co-author Says RLHF Optimizes the Wrong Goal, Ships Calibrated-Decisions Product Jev

At the AI Engineer conference, ChatGPT/GPT-4/InstructGPT co-author Diogo Almeida laid out a systematic critique of RLHF: its objective is wrong, and the way forward is calibrated decision-making. His company TypeSafe shipped Jev on September 15 to implement the idea. Blogger ccerrato147 relayed the full argument in a long thread.

Confirmed

• Diogo Almeida co-authored ChatGPT, GPT-4 and InstructGPT; the views come from his AI Engineer talk as relayed by ccerrato147
• TypeSafe released Jev on Sept 15, replacing the RLHF objective with typed, calibrated decisions
• Almeida's analysis of RLHF: a two-step process of collecting then optimizing human preferences, where the goal was never "running software correctly" but "getting humans to nod" — hence every LLM needs a human in the loop

Arguments

• ccerrato147 disputes that Claude Code represents an automation era, calling it a terminal assistant still driven by RLHF and user-pleasing; models get stronger at agentic tasks yet follow instructions less
• "Overpromising is not a bug, it's the loss function": feeding ChatGPT a fart sound file as "music" gets praised as "atmospheric experimental music," so enterprises learn never to let AI decide on risky matters
• Garry Tan's "golden age of just-in-time software" is double-edged: writing software got automated but software didn't get smarter — still 2019 building blocks and if/else with a chatbot attached
• The thread opens with: "vibe-coded SaaS is just 2019 software written faster"
• Positive claim: smart software is not a flowchart with a chatbot bolted on, but a flowchart whose nodes think — each hardcoded rule box becomes a calibrated decision (routing, scoring, approving or escalating)
• Conclusion: the way out requires a different objective function — not human preference (RLHF), not verified correctness (RLVR), but calibrated decisions: confidence tracks accuracy; act when confident, escalate to humans when not

Why it matters

This is a critique from inside the RLHF lineage (an InstructGPT co-author) targeting the foundational flaw of mainstream alignment, already realized as a product (Jev), offering an engineering path distinct from RLHF/RLVR for safe autonomous AI decision-making in enterprises.

2026-09-20 ~ 2026-09-20 · 9 related posts

• ChatGPT co-author: your vibe-coded SaaS is just 2019 software, faster — ccerrato147 · 2026-09-20
• "Your vibe-coded SaaS is 2019 software, written faster" — ChatGPT co-author thread — ccerrato147 · 2026-09-20
• [source] ChatGPT co-author: RLHF optimizes for approval, so AI still can't be trusted to issue a refund — ccerrato147 · 2026-09-20
• [source] Why every LLM needs a human in the loop: we literally put them there via RLHF — ccerrato147 · 2026-09-20
• Overpromising is the loss function: ChatGPT praises a file of fart sounds as music — ccerrato147 · 2026-09-20
• Claude Code is assistance, not automation — RLHF optimizes for approval — ccerrato147 · 2026-09-20
• Garry Tan's just-in-time software golden age: cheaper to write, exactly as dumb — ccerrato147 · 2026-09-20
• Smart software is a flowchart where the nodes think, not a chatbot bolted on — ccerrato147 · 2026-09-20
• [source] TypeSafe ships Jev: typed calibrated decisions, no RLHF and no text generation — ccerrato147 · 2026-09-20