# Diogo Almeida

> Source: https://aiwiki.ai/wiki/diogo_almeida
> Updated: 2026-09-16
> Fact-checked: 2026-09-16
> Categories: AI Research, OpenAI, People
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Diogo Almeida." aiwiki.ai, 16 Sept 2026. https://aiwiki.ai/wiki/diogo_almeida
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Diogo Almeida** (full name Diogo Moitinho de Almeida) is a machine learning researcher and the co-founder and chief executive of [TypeSafe AI](https://aiwiki.ai/wiki/typesafe_ai), a San Francisco lab that came out of stealth on 15 September 2026 with about $40 million in seed funding led by DCVC and a first model, [Jev](https://aiwiki.ai/wiki/jev), built to return typed decisions to software rather than text to people [1][3]. Before TypeSafe he spent about four and a half years at [OpenAI](https://aiwiki.ai/wiki/openai), by his own account, where he was one of the primary authors of the 2022 [InstructGPT](https://aiwiki.ai/wiki/instructgpt) paper, is listed among the contributors to [ChatGPT](https://aiwiki.ai/wiki/chatgpt), and appears in three sections of the [GPT-4](https://aiwiki.ai/wiki/gpt_4) technical report's author contributions [4][5][6][21]. Earlier he worked at the medical-imaging startup Enlitic and, according to conference listings from 2017, as a research engineer at [Google Brain](https://aiwiki.ai/wiki/google_brain) [12][14]. TypeSafe, DCVC and Almeida himself describe him as a co-inventor of [RLHF](https://aiwiki.ai/wiki/rlhf) and ChatGPT; this article records that description as theirs and sets out what the public record shows [1][22][23].

This page is about the AI researcher. The English Wikipedia articles titled "Diogo Almeida" are about Portuguese footballers and are unrelated.

## Early life and competitions

The official records of the International Mathematical Olympiad list a contestant named Diogo Miguel Moitinho de Almeida on the Philippine team at the 2008 IMO, with a score of 15 and a bronze medal [15]. A speaker biography published with his 2016 RE.WORK Deep Learning Summit talk says he was an IMO medalist "ending a 13-year losing streak for the Philippines", that he received a top prize in the Interdisciplinary Contest in Modeling, and that he completed his undergraduate degree at Rensselaer Polytechnic Institute [13]. Those three statements come from a conference-supplied biography rather than an independent record; only the IMO medal is confirmed by a primary source here.

He entered machine learning through [Kaggle](https://aiwiki.ai/wiki/kaggle). In the 2013 ChaLearn cause-effect pairs challenge, which asked entrants to decide from observational data alone whether variable A causes variable B, his entry finished second on the public leaderboard and first on the private leaderboard, according to the abstract of his chapter "Pattern-based Causal Feature Extraction" in the Springer volume *Cause Effect Pairs in Machine Learning* (Guyon, Statnikov and Batu, eds., 2019) [16]. The chapter describes a feature-extraction pipeline that generated more than 20,000 features with little manual work and a slimmed-down version that matched its performance with 324 features [16]. The TWIML AI Podcast, in an episode published in October 2016 and recorded at the O'Reilly AI and Strata conferences in New York, introduced him as "a past 1st place Kaggle competition winner" and linked the same challenge [12].

## Enlitic (2015 to 2016)

By late 2015 Almeida was a data scientist at Enlitic, the San Francisco medical-imaging deep learning company; his papers from that period carry the address diogo@enlitic.com [8][9][10][11]. TWIML introduced him in October 2016 as "senior data scientist at healthcare startup Enlitic", recorded at the O'Reilly AI and Strata conferences in New York, where his talk was titled "Deep Learning: Modular in theory, inflexible in practice" [12]. He gave a talk with the same title at RE.WORK's Deep Learning Summit London in 2016 and an introductory deep learning talk at GOTO Copenhagen in October 2016 [13][30].

Four preprints came out of this period:

| Year | Paper | Co-authors | Venue / note |
|---|---|---|---|
| 2015 | GradNets: Dynamic Interpolation Between Neural Architectures | Nate Sauder (Enlitic) | arXiv:1511.06827, "under review" for ICLR 2016 [8] |
| 2016 | Resnet in Resnet: Generalizing Residual Architectures | Sasha Targ, Kevin Lyman | ICLR 2016 workshop track; reports a new CIFAR-100 state of the art [9] |
| 2016 | Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units | Wenling Shang, Kihyuk Sohn, Honglak Lee | ICML 2016; introduces the CReLU activation [10] |
| 2016 | Genetic Architect: Discovering Genomic Structure with Learned Neural Architectures | Laura Deming, Sasha Targ, Nate Sauder, Chun Jimmie Ye (UCSF) | arXiv:1605.07156 [11] |

GradNets and Resnet in Resnet both mark Almeida as an equal first author [8][9]. He is also a named inventor, with Li Yao, Devon Bernard, Kevin Lyman and Enlitic founder Jeremy Howard, on US patent 10,340,044, "Medical scan image analysis system", assigned to Enlitic with a 2017 priority date [17].

## Google Brain (2017)

The GOTO Amsterdam 2017 conference listed "Diogo Moitinho de Almeida, Research Engineer at Google Brain" for a June 2017 session titled "Deep Learning: What It Is and What It Can Do For You" [14]. TypeSafe's team page, the AI Engineer speaker page and the Databricks Data + AI Summit speaker page all say he was at Google Brain before OpenAI [2][19][20]. The arXiv listings under his name include no paper with a Google affiliation, so the dates and content of that work are not documented in the public sources used here [18].

## OpenAI

Almeida's earliest publication with an OpenAI affiliation is "A Generalizable Approach to Learning Optimizers" (June 2021), on which he is first author with Clemens Winter, Jie Tang and [Wojciech Zaremba](https://aiwiki.ai/wiki/wojciech_zaremba). The paper trains an LSTM controller to adjust optimizer hyperparameters rather than model parameters directly, and reports that the learned optimizer beat Adam on every neural-network task tried, with a 2x speedup on ImageNet and a 2.5x speedup on a language modeling task that used more than five orders of magnitude more compute than the training tasks [7]. This, not the 2017 Google Brain paper "Learned Optimizers that Scale and Generalize" (on which he is not an author), is the learned-optimizer work his AI Engineer biography refers to [7][19].

### InstructGPT

He is the fourth listed author of "Training language models to follow instructions with human feedback" (Ouyang et al., arXiv:2203.02155, March 2022), the paper that introduced InstructGPT [4]. The paper marks nine of its twenty authors, including Almeida, with an asterisk as "Primary authors", and notes that it "was a joint project of the OpenAI Alignment team" led by Ryan Lowe and [Jan Leike](https://aiwiki.ai/wiki/jan_leike) [4]. A footnote records one concrete contribution: to find the few-shot prefix used for the paper's prompted GPT-3 baseline, "authors RL and DA held a prefix-finding competition: each spent an hour interacting with GPT-3 to come up with their two best prefixes", scored by reward-model score on the validation set, and "DA won" [4]. InstructGPT's headline result was that a 1.3-billion-parameter model fine-tuned with human feedback was preferred by labelers to the 175-billion-parameter GPT-3 [4].

### ChatGPT and GPT-4

OpenAI's ChatGPT announcement of 30 November 2022 credits no individual authors, but its acknowledgments list Almeida among 87 named contributors [6]. In the GPT-4 technical report (March 2023) he appears in the author list and in three of the alphabetical contribution groups: "Dataset contributions" and "Foundational RLHF and InstructGPT work" under "Reinforcement Learning & Alignment", and "Instruction following and API evals" (with Carroll Wainwright and Marvin Zhang) under "Evaluation & analysis" [5]. He is not among the named "core contributors" with individual role titles in that report [5].

His Google Scholar profile, which lists his affiliation only as "ex-OpenAI", recorded about 63,000 citations in September 2026, the large majority from the InstructGPT paper and the GPT-4 report [18]. In his 2026 AI Council talk he said he had been at OpenAI "for about four and a half years" [21].

## TypeSafe AI (2024 to present)

TypeSafe AI was founded in 2024 in San Francisco by Almeida (CEO) with Erik Gafni (CTO) and Sasha Sheng (COO) [1][2]. The company stayed in stealth until 15 September 2026, when it announced approximately $40 million in seed funding led by DCVC, with general partner James Hardiman quoted in the release, and opened early access to Jev [1][22]. Almeida's launch post describes "two years in stealth" and a new stack built around what the company calls System One Models, trained with a method it names Reinforcement Learning for Calibrated Decisions (RLCD) [3]. His launch tweet, which had been viewed about 18.8 million times by the next day, claimed Jev is "20-200x faster" and "40-400x cheaper" than frontier models; those multipliers are the company's own measurements on its own workflow evaluations, and the launch post concedes that the workflows "were made by individuals on our model capabilities team, so some bias could exist" [3][23].

In the press release Almeida said: "TypeSafe was founded to pursue an alternative path for AI research, focused on machine-native AI. I spent years working on models designed to make AI better at interacting with people. But if AI is going to fundamentally change how work gets done, people can't be the only consumers of intelligence. Most intelligence should eventually live inside software, running quietly in the background." [1] The Register's Thomas Claburn, covering the launch, noted that the company's "hallucination-free" claim rests on Jev returning typed structured values rather than natural language, which "does not preclude the possibility of being incorrect" [24].

The company's manifesto, "Composable AI: Build Prod, Not God", argues that current models are trained to be helpful assistants for a human on the other side and that this is why they need humans in the loop; its stated mission is "to pave the shortest path to an AI-based economic revolution" by making intelligence composable [28]. He pitched the company on the Founders You Should Know showcase in March 2026; the video's description says TypeSafe is building "large language models designed explicitly for true no-human-in-the-loop automation, not conversational assistants" [29].

## Views on RLHF and benchmarks

Almeida's public argument since leaving OpenAI is that [RLHF](https://aiwiki.ai/wiki/rlhf), the technique he helped establish, optimizes models to please the person in the loop and therefore does not prepare them for unsupervised automation. His talk "What's next after RLHF?" at the AI Engineer World's Fair 2026 contrasts assistance with automation, argues that preference optimization penalizes visible uncertainty more than confident mistakes, and calls for a third objective beyond human preference and verifiable rewards: calibrated decision-making [19]. He gave a keynote titled "AI: too good to be true, too bad to be useful" at AI Council in 2026 (posted to the TypeSafe blog on 19 June 2026), where he framed the gap between benchmark results and real automation as the difference between "human in the loop assistance tasks" and "machine in the loop automation tasks" [21]. He was also listed for a Data + AI Summit 2026 session with the same "What's Next After RLHF?" title [20].

Two TypeSafe blog posts from the week before launch extend the argument. "The Bitterest Lesson" (10 September 2026), written in the first person and referring to "when making InstructGPT/RLHF" at OpenAI, argues that choosing the right task matters more than data, which matters more than compute, which matters more than algorithms, and cites InstructGPT's figure showing a much smaller model trained on the right task beating GPT-3 [26]. "Lies, Damned Lies, and Benchmarks" (11 September 2026) says TypeSafe will publish no standard benchmark table in its model releases and will instead post dated evaluation snapshots that are retired once published [27]. In a July 2026 post on X he described poor tool use as "a smell of too much on-policy RL" that makes it hard to transfer engineering practices such as decomposition, abstraction and state management to language models [19].

## The "co-inventor of RLHF and ChatGPT" description

TypeSafe's team page says Almeida "co-invented RLHF and InstructGPT, the methods that lead to ChatGPT and GPT4"; the Business Wire release calls him "co-inventor of RLHF/ChatGPT"; DCVC writes that he "co-invented some of the core techniques and products that launched the current boom: InstructGPT, ChatGPT, and GPT-4"; his X profile says he "co-created RLHF/ChatGPT" at OpenAI; and AI Council billed him as "Co-Inventor of ChatGPT" [1][2][21][22][23]. Two points of context apply. Reinforcement learning from human preferences as a method was introduced in 2017 by Paul Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg and Dario Amodei in "Deep reinforcement learning from human preferences", five years before InstructGPT [25]. What the public record documents for Almeida is primary authorship of the InstructGPT paper, which applied that method to GPT-3 and is the direct technical ancestor of ChatGPT, plus contributor credits on ChatGPT and the GPT-4 report [4][5][6]. The AI Engineer speaker page puts it more narrowly: he "helped develop the human-feedback techniques behind modern conversational AI" [19].

## Selected publications

| Year | Title | Affiliation | Reference |
|---|---|---|---|
| 2015 | GradNets: Dynamic Interpolation Between Neural Architectures | Enlitic | arXiv:1511.06827 [8] |
| 2016 | Resnet in Resnet: Generalizing Residual Architectures | Enlitic | arXiv:1603.08029 [9] |
| 2016 | Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units | Enlitic | ICML 2016 [10] |
| 2016 | Genetic Architect: Discovering Genomic Structure with Learned Neural Architectures | Enlitic / UCSF | arXiv:1605.07156 [11] |
| 2019 | Pattern-based Causal Feature Extraction | (chapter) | Cause Effect Pairs in Machine Learning, Springer [16] |
| 2021 | A Generalizable Approach to Learning Optimizers | OpenAI | arXiv:2106.00958 [7] |
| 2022 | Training language models to follow instructions with human feedback | OpenAI | arXiv:2203.02155, NeurIPS 2022 [4] |
| 2023 | GPT-4 Technical Report | OpenAI | arXiv:2303.08774 [5] |

## References

1. "TypeSafe AI Emerges From Stealth With $40M in Funding With New Model for Composable AI", Business Wire, 15 September 2026. https://www.businesswire.com/news/home/20260915525333/en/
2. "Team", TypeSafe AI, accessed 16 September 2026. https://typesafe.ai/team
3. Diogo Almeida, "Introducing System One Models & Jev", TypeSafe AI blog, 15 September 2026. https://typesafe.ai/blog/introducing-system-one-models-and-jev
4. Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin et al., "Training language models to follow instructions with human feedback", arXiv:2203.02155, 4 March 2022 (author list, primary-author footnote and footnote 6). https://arxiv.org/abs/2203.02155
5. OpenAI, "GPT-4 Technical Report", March 2023, author contributions section ("Reinforcement Learning & Alignment"). https://cdn.openai.com/papers/gpt-4.pdf (also arXiv:2303.08774)
6. OpenAI, "Introducing ChatGPT", 30 November 2022, acknowledgments ("Contributors"). https://openai.com/index/chatgpt/ (checked against the Wayback Machine capture of 31 December 2023)
7. Diogo Almeida, Clemens Winter, Jie Tang, Wojciech Zaremba, "A Generalizable Approach to Learning Optimizers", arXiv:2106.00958, June 2021. https://arxiv.org/abs/2106.00958
8. Diogo Almeida, Nate Sauder, "GradNets: Dynamic Interpolation Between Neural Architectures", arXiv:1511.06827, November 2015. https://arxiv.org/abs/1511.06827
9. Sasha Targ, Diogo Almeida, Kevin Lyman, "Resnet in Resnet: Generalizing Residual Architectures", arXiv:1603.08029, March 2016. https://arxiv.org/abs/1603.08029
10. Wenling Shang, Kihyuk Sohn, Diogo Almeida, Honglak Lee, "Understanding and Improving Convolutional Neural Networks via Concatenated Rectified Linear Units", ICML 2016, arXiv:1603.05201. https://arxiv.org/abs/1603.05201
11. Laura Deming, Sasha Targ, Nate Sauder, Diogo Almeida, Chun Jimmie Ye, "Genetic Architect: Discovering Genomic Structure with Learned Neural Architectures", arXiv:1605.07156, May 2016. https://arxiv.org/abs/1605.07156
12. "Deep Learning: Modular in Theory, Inflexible in Practice with Diogo Almeida", The TWIML AI Podcast, episode 8, 23 October 2016. https://twimlai.com/podcast/twimlai/deep-learning-modular-theory-inflexible-practice-diogo-almeida/
13. RE.WORK, "Deep Learning: Modular in Theory, Inflexible in Practice - Diogo Moitinho de Almeida #reworkDL", YouTube, posted 29 April 2016 (video description with speaker biography). https://www.youtube.com/watch?v=DGgv14_3-QE
14. GOTO Amsterdam 2017, session "Deep Learning: What It Is and What It Can Do For You", speaker listing. https://gotoams.nl/2017/sessions/94 (recording: https://www.youtube.com/watch?v=B4hFqlQdmJE)
15. International Mathematical Olympiad, Philippines team results, IMO 2008 (contestant Diogo Miguel Moitinho de Almeida: bronze, 15 points). https://www.imo-official.org/results/team/year/2008/country/PHI/ (individual record: https://www.imo-official.org/results/contestant/17909/)
16. ChaLearn, "Cause-effect pairs in machine learning" (table of contents and chapter abstracts for Guyon, Statnikov and Batu, eds., Cause Effect Pairs in Machine Learning, Springer, 2019; Chapter 10). https://causality.chalearn.org/experimental-design
17. US Patent 10,340,044 B2, "Medical scan image analysis system", inventors Li Yao, Devon Bernard, Kevin Lyman, Diogo Almeida, Jeremy Howard, assignee Enlitic Inc. https://patents.google.com/patent/US10340044B2/en
18. Diogo Almeida, Google Scholar profile, accessed 16 September 2026. https://scholar.google.com/citations?user=0T4y07QAAAAJ
19. "Diogo Almeida: Bio, Work & Ideas", AI Engineer, and talk page "What's next after RLHF?", AI Engineer World's Fair 2026. https://ai.engineer/speakers/diogo-almeida and https://ai.engineer/talks/what-s-next-after-rlhf (recording posted 31 July 2026: https://www.youtube.com/watch?v=cJ0EOzey--o; the tool-use post is https://x.com/completeskeptic/status/2073506868098662783, 4 July 2026)
20. "Diogo Almeida", Data + AI Summit speaker page, Databricks, accessed 16 September 2026. https://www.databricks.com/dataaisummit/speaker/diogo-almeida
21. Diogo Almeida, "AI: too good to be true, too bad to be useful", AI Council, 2026 (talk page with transcript). https://www.aicouncil.com/talks/ai-too-good-to-be-true-too-bad-to-be-useful
22. James Hardiman, "TypeSafe emerges from stealth with a new way of doing AI", DCVC, 15 September 2026. https://www.dcvc.com/news-insights/typesafe-emerges-from-stealth-with-a-new-way-of-doing-ai
23. Diogo Almeida (@CompleteSkeptic), post on X, 15 September 2026, and profile description. https://x.com/CompleteSkeptic/status/2099925682726002904
24. Thomas Claburn, "TypeSafe AI debuts model for machines that plays Doom", The Register, 16 September 2026. https://www.theregister.com/ai-and-ml/2026/09/16/typesafe-ai-debuts-model-for-machines-that-plays-doom/5296711
25. Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, Dario Amodei, "Deep reinforcement learning from human preferences", arXiv:1706.03741, June 2017. https://arxiv.org/abs/1706.03741
26. "The Bitterest Lesson", TypeSafe AI blog, 10 September 2026. https://typesafe.ai/blog/bitterest-lesson
27. "Lies, Damned Lies, and Benchmarks", TypeSafe AI blog, 11 September 2026. https://typesafe.ai/blog/antibenchmaxxing
28. "Composable AI: Build Prod, Not God", TypeSafe AI manifesto, accessed 16 September 2026. https://typesafe.ai/manifesto
29. Founders You Should Know, "Diogo Almeida - TypeSafe AI", YouTube, 31 March 2026 (embedded on the TypeSafe blog). https://www.youtube.com/watch?v=LE3bGTaAgOE
30. GOTO Conferences, "Deep Learning - What is it and What It Can Do For You - Diogo Moitinho de Almeida - GOTO 2016", YouTube, posted 26 October 2016. https://www.youtube.com/watch?v=fvNsz7XMM9w