Diogo Almeida

Founder & CEO, TypeSafe AI

About

Diogo Almeida is the founder and CEO of TypeSafe AI, and was the fourth author on InstructGPT (2022), the OpenAI paper that made RLHF the standard way a language model is turned into an assistant. He worked on the post-training behind ChatGPT and GPT-4, and was at Google Brain before that. He now argues that the technique he helped build is what holds automation back. RLHF optimises for human preference, and a reward model can detect visible uncertainty far more easily than it can check whether an answer is correct — so the objective rewards sounding confident. On his account hallucination is not a defect waiting to be patched but the objective working as specified, and the assistance era that followed ChatGPT was a "weird detour" away from software that does real work. TypeSafe's answer is RLCD, reinforcement learning for calibrated decisions, and a model called Jev that returns a typed value with a probability attached instead of prose.

Key Contributions

Fourth of twenty authors on InstructGPT (2022), the paper that made RLHF the standard method for turning a language model into an assistant
Worked on the post-training behind ChatGPT and GPT-4 at OpenAI, after Google Brain
Argues that RLHF's preference objective makes hallucination intrinsic rather than incidental: a reward model detects visible uncertainty far more easily than it verifies correctness, so training rewards confident delivery
Separates post-training into three branches by what each optimises for — RLHF for human preference, RLVR for pure correctness, and TypeSafe's RLCD for calibrated decisions
Calls the assistance era after ChatGPT a "weird detour" from real automation, arguing that SaaS has barely changed since 2019 except for a chatbot latched onto the side
Founded TypeSafe AI and launched Jev (September 2026), which returns a value matching a schema defined in advance, with what the lab calls a calibrated probability, rather than free text

Moments Timeline View all moments

Sep 2026

Jev and System One Models

TypeSafe AI, after two years in stealth, publishes Jev along with a name for the category it claims to have opened: System One Models. The pitch is a change of contract rather than a change of substrate. Where a language model takes a prompt and returns free text one token at a time, Jev takes an input and a schema and returns a value of that type with a calibrated probability attached, in what the lab calls a single query. Founder Diogo Almeida — fourth of twenty authors on InstructGPT, the paper that made RLHF standard — frames it as a third branch of post-training, sorting the field by what each objective optimises for: RLHF for human preference, RLVR for pure correctness, and TypeSafe's RLCD for calibrated decisions. His argument is that preference optimisation makes hallucination intrinsic, because a reward model can detect visible uncertainty far more easily than it can verify correctness, so the training rewards confident delivery. The intelligence underneath is still pre-trained: Almeida calls pre-training "phenomenal" and says the problem is "how we unearth it", describing the method as mainlining the intelligence of pre-trained models into something useful for software. Whose base model, and whether TypeSafe pre-trained it themselves, is not disclosed, along with parameter count and context window. Jev's own figures are TypeSafe's, measured by TypeSafe on an early-access model released the same day: responses of 70 to 500 milliseconds, and $0.042 per million input tokens with output free. The 3-to-329-second frontier-model range they are set against is not theirs — it is lifted from a third-party benchmark site. And the 0% hallucination rate is not a measurement at all: the lab says plainly that the number "is not empirical", and that guaranteed schema matching is what lets them put a zero on the chart. It describes conformance to a schema, not truth. Early access is open; no independent numbers have been published.

Videos & Interviews

Why Jev — Diogo Almeida, TypeSafe AI

Almeida gives the argument behind Jev weeks before the model ships, while TypeSafe is, in his words, "still kind of stealthy." The case is that RLHF optimises for human preference, that a reward model finds visible uncertainty far easier to punish than incorrectness is to detect, and that hallucination therefore follows from the objective rather than from any missing patch — his illustration is ChatGPT calling a recording of fart noises "a very eerie vibe atmosphere piece." He splits post-training into three branches by what each targets, RLHF for preference, RLVR for correctness, and TypeSafe's RLCD for calibrated decisions, and defends the part he is not attacking: pre-training is "phenomenal," and the problem is "how we unearth it." Worth watching for what he declines to answer, too — asked directly why not simply train a classifier head on top of pre-training, he says it's complicated and moves on.

View Details

Papers & Publications

Training language models to follow instructions with human feedback

2022

The InstructGPT paper, on which Almeida is the fourth of twenty authors. It showed that a 1.3B model tuned on human comparisons was preferred to the 175B GPT-3, and in doing so made RLHF the default last step in building an assistant. Almeida now argues the same objective is why those assistants overclaim.

Read Paper

Connections

Yoshua Bengio

Kindred

AI Pioneer & Safety Researcher

A resonance rather than a documented exchange — though an audience member at Almeida's 2026 talk put them side by side, asking whether one could do roughly what Bengio suggests, and Almeida called it complicated and moved on. Bengio's Scientist AI is built to explain the world 'as opposed to taking actions in it to imitate or please humans,' carrying explicit uncertainty against overconfident predictions; TypeSafe's post-training aims at what Almeida calls 'calibrated decision-making.' Both reach calibration as the way out of people-pleasing, from opposite motives: one to make agents safe, the other to make automation trustworthy.

arxiv.org · youtube.com

Richard Sutton

In contrast

Founder, Oak Lab & Professor, University of Alberta

Both call the present a wrong turn, and disagree about where it turned. Sutton holds that language models are 'about mimicking people' and get by 'without having a goal or a sense of better or worse,' so the mistake is imitation itself; Almeida calls pre-training 'phenomenal' and blames only the final step, optimising for human preference. In the same talk he invoked Sutton's Bitter Lesson and amended it: data matters more than compute, and doing the right task matters more than data.

youtube.com · dwarkesh.com

Amanda Askell

Collaborated

Philosopher & AI Alignment Researcher

Both are authors of InstructGPT (2022), the OpenAI Alignment team paper that made reinforcement learning from human feedback the standard last step in building an assistant — Almeida among its primary authors, Askell in the wider list. The paper took its helpful, honest, harmless framing from Askell's earlier work, then conceded that honesty could not be measured in a black-box model and measured truthfulness instead. Almeida's later argument starts in that gap: an objective built on human preference rewards sounding right more easily than being calibrated.

arxiv.org