Diogo Almeida helped invent the training method that powers nearly every major AI chatbot today, then spent two years building a company premised on why that method has fundamental limits. As a member of OpenAI's post-training team, Almeida co-authored the InstructGPT paper and contributed to the GPT-4 technical report, helping develop the human-feedback techniques that became RLHF (Reinforcement Learning from Human Feedback). 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. It's not an outside critic making the argument, but one of the people who helped prove RLHF worked in the first place. His position gives his critique a credibility that purely external skeptics wouldn't carry.

What's Wrong With RLHF, According to Its Own Creator?

Almeida's core argument rests on a sharp distinction: assistance versus automation. "Today's AI was designed for assistance through optimizing for human preference," he has 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. "Why do all LLMs require a human in the loop? The simple answer is we literally put them in the loop," Almeida explained. 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 cuts to the heart of a problem many AI users have noticed: "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," Almeida stated. This 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.

His diagnosis of the industry's last several years is blunt: "The AI industry has produced cheaper software, not smarter software." Coding tools have accelerated how fast software gets written, but the software's actual capabilities have stayed within the constraints of what he calls the "assistance era".

How Is TypeSafe AI Building a Different Kind of AI Model?

TypeSafe AI's answer to this diagnosis is a training method called RLCD, or Reinforcement Learning for Calibrated Decisions. 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, called Jev, doesn't generate text at all. Instead, 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 public launch on September 15, 2026. The company spent roughly two years building toward this specific bet before showing it publicly.

Steps to Understanding the RLHF Versus RLCD Debate

• Recognize RLHF's Design Goal: RLHF optimizes for human rater approval as the training signal, which works well for assistance tasks where humans check outputs but creates structural problems for autonomous decision-making.
• Understand the Confidence Problem: RLHF-trained models can be confidently wrong because they're rewarded for sounding convincing, not for being correct, creating a gap between how certain they sound and how accurate they actually are.
• See the Alternative Approach: RLCD trains models to return structured decisions with calibrated confidence values, designed for automation where software acts on outputs directly without human review.
• Evaluate the Trade-offs: Jev can't generate free-form text like ChatGPT, but that limitation is intentional; it prevents the hallucination problem by design rather than trying to fix it after the fact.

Is the Industry Actually Moving Away From RLHF?

It's worth being precise about which parts of Almeida's story are his framing and which are independently verifiable. RLHF's mechanics genuinely do optimize for human rater approval as the training signal, and it's well-documented in the machine learning literature that this can produce sycophancy and overconfidence as side effects. That's not a controversial technical claim.

What remains more contested is whether this makes RLHF-trained models fundamentally unsuited for automation, versus merely imperfect at it. A difference exists 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, remains genuinely unresolved. 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.

Almeida's own responses in that discussion were notably candid. He agreed 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 Does This Matter Beyond One Founder's Story?

Almeida's specific position, having helped build RLHF and 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 departures from consensus by prominent AI researchers 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 points to a real tension in how the AI industry has evolved. The question isn't whether RLHF works; it clearly does. The question is whether it was ever the right tool for the job the industry actually wants to solve.