Diogo Almeida spent years helping build the exact thing he now says the industry needs to move past. As a member of OpenAI's post-training team, he co-authored the InstructGPT paper, contributed to the GPT-4 technical report, and helped develop the human-feedback techniques that became ChatGPT — the work that made RLHF (Reinforcement Learning from Human Feedback) the standard training recipe behind essentially every major chat model since. Now, as founder of TypeSafe AI, he argues the field will eventually look back on this entire era as "a weird detour."
That's a striking claim from someone with his specific credentials — not an outside critic, but one of the people who helped prove RLHF worked in the first place. His argument, and the company he built around it, are worth understanding in more depth than explainx.ai's earlier coverage of TypeSafe's Jev launch had room for.
TL;DR
| Question | Answer |
|---|---|
| Who is he? | Former OpenAI post-training team member; co-author of InstructGPT, GPT-4, and ChatGPT's underlying RLHF work; previously at Google Brain |
| What's his core claim? | RLHF optimizes for human preference, not task correctness — making it structurally suited for assistance, not automation |
| What does he mean by "weird detour"? | That the current scaling-and-chat-centric era of AI will eventually be seen as a detour from the real goal: reliable automation |
| What's he building now? | TypeSafe AI, co-founded with Erik Gafni and Sasha Sheng in mid-2024 — makers of Jev, a "System One Model" |
| What training method replaces RLHF in his framing? | RLCD — Reinforcement Learning for Calibrated Decisions |
| Is this view widely accepted? | No — genuinely contested; RLHF (and its descendants like RLAIF and Constitutional AI) remains the dominant approach industry-wide |
The case, in his own words
Almeida's argument rests on a distinction he draws sharply: assistance versus automation. "Today's AI was designed for assistance through optimizing for human preference," he's said — a design choice that works well when a human is in the loop checking every output, and fails precisely when you try to remove that human. His explanation for why it fails is specific and mechanical, not just philosophical: "Why do all LLMs require a human in the loop? The simple answer is we literally put them in the loop" — meaning the RLHF training process itself uses human raters judging outputs, so the resulting model is optimized to produce outputs a human rater would approve of, not outputs that are independently, verifiably correct.
The sharpest version of his critique is this line: "Overpromising is a feature...by design. By construction, every RLHF model will always have a big difference [between confidence and correctness]... No matter how wrong the models are, they will look right." That's a direct claim that hallucination isn't a bug RLHF happens to have — it's close to an inevitable consequence of what RLHF optimizes for. A model rewarded for sounding convincing to a human rater will learn to sound convincing, whether or not the underlying content is actually true, because "sounding convincing" and "being correct" are correlated but not identical signals, and the training process can only optimize for the signal it actually receives.
His diagnosis of the industry's last several years, put plainly: "The AI industry has produced cheaper software, not smarter software." Coding tools have accelerated how fast software gets written, in his framing, but the software's actual capabilities have stayed within the constraints of what he calls the "assistance era" — useful with supervision, unreliable without it.
Watch: "What's Next After RLHF?"
Almeida laid out this argument at length in a talk that's worth watching directly for the full reasoning rather than the condensed version above — including his three-part structure: today's AI is great at assistance but poor at automation, RLHF was designed for assistance by construction, and tomorrow's AI needs to be built around calibrated decision-making instead.
From critique to product: RLCD and Jev
TypeSafe AI's answer to this diagnosis is a training method it calls RLCD — Reinforcement Learning for Calibrated Decisions — explicitly designed around a different objective than RLHF's human-preference signal.
Where RLHF asks "would a human rate this response highly," RLCD is built to ask something closer to "how confident should this system actually be, and is that confidence accurate." The resulting model, Jev, doesn't generate text at all — it returns structured decisions (a choice, a score, or a yes/no probability) with a calibrated confidence value attached, computed in parallel rather than through autoregressive generation. That architectural choice is a direct expression of Almeida's underlying philosophy: if the goal is automation rather than assistance, the output shouldn't be free-form text a human has to read and judge — it should be a decision a piece of software can act on directly, with an honest confidence signal attached.
TypeSafe was co-founded by Almeida alongside Erik Gafni and Sasha Sheng in mid-2024, operating largely in stealth until Jev's September 15, 2026 public launch — meaning the company spent roughly two years building toward this specific bet before showing it publicly.
Where the argument holds up, and where it's contested
It's worth being precise about which parts of this story are Almeida's framing and which are independently verifiable:
Well-supported: RLHF's mechanics genuinely do optimize for human rater approval as the training signal, and it's well-documented in the ML literature that this can produce sycophancy and overconfidence as side effects — this isn't a controversial technical claim.
More contested: whether this makes RLHF-trained models fundamentally unsuited for automation, versus merely imperfect at it — a difference between "structurally broken" and "needs more work." The rest of the industry hasn't abandoned RLHF; it's evolved it, through techniques like RLAIF (AI feedback instead of purely human feedback), Constitutional AI, and RLVR (Reinforcement Learning with Verifiable Rewards) for domains like math and code where a ground-truth answer exists to check against. Whether that evolutionary path eventually solves the calibration problem Almeida describes, or whether it requires the more fundamental architectural break TypeSafe is betting on, is genuinely unresolved.
Not yet fully proven: Jev's own launch drew substantial technical pushback on Hacker News — commenters challenged whether "frontier model" was an earned label for a system that can't generate text, and whether "can't hallucinate" overstates what schema-guaranteed output actually guarantees (see explainx.ai's full breakdown of that debate). Almeida's own responses in that discussion were notably candid — agreeing directly that Jev "can be confidently wrong," which is consistent with intellectual honesty but also means the calibration story isn't a solved problem, just a different, arguably more honest way of being uncertain.
Why this profile matters beyond one founder's story
Almeida's specific position — having helped build RLHF, then building a company premised on its limits — gives his critique a credibility that a purely external critic wouldn't have. It's the same reason Yann LeCun's departures from consensus or Ilya Sutskever's Safe Superintelligence venture draw outsized attention: insiders who reverse course on their own prior work, or leave to pursue a different bet entirely, are a meaningfully different signal than outside skeptics making the same argument. Whether TypeSafe's specific bet on RLCD and non-generative decision models turns out to be the right fix, or one useful piece of a larger puzzle, Almeida's underlying diagnosis — that optimizing for "looks right to a human" and optimizing for "is actually correct" are different objectives that current training methods conflate — is a genuinely useful frame for evaluating any AI system's suitability for unsupervised, high-stakes work, independent of whether you buy his specific solution.
FAQ
Who is Diogo Almeida? A former OpenAI post-training team member who co-authored InstructGPT, GPT-4, and ChatGPT's RLHF work, previously at Google Brain, now co-founder and CEO of TypeSafe AI.
Why does he call ChatGPT-era AI a "weird detour"? Because RLHF optimizes for human preference rather than task correctness, producing systems suited for supervised assistance but not unsupervised automation.
What specifically does he say is wrong with RLHF? That "overpromising is a feature by design" — models trained to satisfy human raters learn to look correct rather than be correct, making hallucination close to inevitable.
What is TypeSafe AI building instead? Jev, a "System One Model" trained with RLCD (Reinforcement Learning for Calibrated Decisions), returning structured decisions with calibrated confidence instead of generated text.
Is this view widely accepted? No — it's contested. RLHF and its descendants (RLAIF, Constitutional AI, RLVR) remain the dominant industry approach.
Has his argument been demonstrated to work in practice? Partially — Jev launched September 15, 2026 with real capability claims but also drew substantial technical pushback, including from Almeida's own candid admission that the model "can be confidently wrong."
Related reading
- TypeSafe AI's Jev: a "System One Model" that never hallucinates
- Scalable oversight: RLHF, DPO, Constitutional AI, weak-to-strong generalization
- What is fine-tuning an LLM? A complete guide
- Structured output and JSON mode prompting: a complete guide
- What is an embedded evaluator in AI safety?
- Are AI labs now hoarding solved math problems to avoid backlash?
- Official: TypeSafe AI · Diogo Almeida on X
Quotes and biographical details in this piece reflect public talks, interviews, and TypeSafe AI's own published materials as of September 16, 2026. Views described are Almeida's own and represent one contested position within an active industry debate.