← All speakers

Diogo Almeida

Conference affiliation: CEO · TypeSafe AI · 2026

Diogo Almeida is the cofounder and CEO of TypeSafe AI, where he leads the team behind Jev, a model that gives software typed judgments and probabilities it can use to decide whether to act. At OpenAI, he coauthored InstructGPT, which showed how much better language models could follow instructions when trained with human feedback. Jev grew out of a question that work left open for him: how should a model convey uncertainty when no one is there to read its answer and catch a mistake?

On this page

From neural networks to instruction following

Diogo worked as a software engineer at Amazon and Stuph.co before becoming a senior data scientist at Enlitic in 2014. He moved to research at Google Brain in 2017 and joined OpenAI in 2020. His early research included the coauthored Resnet in Resnet, a two-stream architecture that generalized residual and conventional convolutional networks.

At OpenAI, Diogo was first author on a 2021 study of learned optimizers with Clemens Winter, Jie Tang, and Wojciech Zaremba. The team wanted a learned optimizer that would still be useful on problems unlike those it had trained on. They left a conventional optimizer in charge of updating the model’s weights and trained a separate controller to adjust its settings as training progressed. The team reported a twofold speedup on ImageNet and a 2.5-fold speedup on a language modeling task that used more than five orders of magnitude more compute than the tasks used to train the controller.

The InstructGPT team started from GPT-3, fine-tuned it on demonstrations, and used human rankings of candidate responses to train a reward model for reinforcement learning from human feedback, or RLHF. Diogo and Ryan Lowe wrote candidate prompts for a baseline that tried to improve GPT-3 through prompting alone. Diogo’s instruction prefix scored highest on the reward model’s validation set and was used for that comparison. On the study’s evaluation prompts, human raters preferred answers from the 1.3-billion-parameter InstructGPT model to those from the 175-billion-parameter GPT-3, even with prompting. A separate example in Figure 9 shows a limitation of the 175-billion-parameter InstructGPT model itself: it could take a question’s false premise at face value and answer accordingly. Diogo went on to contribute to ChatGPT and coauthor the GPT-4 technical report.

When the program has to make the call

Diogo initially saw chat models as a possible route to artificial general intelligence. He came to question how little routine work they had automated, even as he continued to value ChatGPT and coding assistants. He left OpenAI in 2024 and cofounded TypeSafe that August to build models for decisions inside software.

Customer service makes his distinction between assistance and automation concrete. A chatbot can help a customer find the relevant policy while the business still asks a person to authorize a costly action. Diogo argues that a reward model trained on preferences can spot hesitation more easily than it can spot a plausible mistake. That gives a model reason to sound certain even when several answers are plausible. For an automated workflow, he treats calibrated decision-making as a training goal in its own right: when the model assigns high probabilities, those judgments should prove more dependable in practice. The application can then decide whether the evidence is strong enough for a particular action.

Diogo introduced Jev in September 2026 as TypeSafe’s first System One Model, trained with what the company calls Reinforcement Learning for Calibrated Decisions. The design gives up arbitrary text generation. A parallel sampler produces structured outputs together rather than generating a response token by token. The interface puts two of Diogo’s design choices in the hands of developers:

  • Typed questions keep the policy in code. Jev evaluates each question independently against the same input state. Choice selects among supplied options, Score evaluates an ordered rubric, and Noul estimates whether a statement is true. In TypeSafe’s startup pitch example, the application asks separately about market size, technical feasibility, and differentiation, then weights the scores in code. A developer can give feasibility more weight by changing a coefficient while leaving the questions alone.
  • Probability distributions make review thresholds possible. Choice and Score both return distributions, but they calculate confidence differently. For a three-option Choice with probabilities of 0.6, 0.3, and 0.1, Jev reports confidence of 0.4. That statistic is distinct from the leading option’s probability of 0.6. Score also accounts for how far the probability mass lies from the most likely level on the ordered scale. A team can use those outputs to route uncertain cases to a person and test the threshold on real cases, where it can weigh the cost of a mistake against the cost of review.

Diogo’s ambition for these narrow judgments reaches beyond any one API. He imagines a linter that checks whether code preserves an intended property that would be hard to express as a syntax rule. He applies the same thinking to agent workflows: split work into stages with outputs that can be checked, and manage shared state explicitly so one agent can reuse what another has learned about a codebase. In both proposals, the model handles a specific judgment while the surrounding program keeps track of the work and decides what happens next.

Read the topics behind these talks

2 conference talks

Loading more talks and highlights…

References