# Diogo Almeida Diogo Almeida is the co-founder and CEO of [[TypeSafe AI]], the lab behind [[Jev]] and the [[System One Models]] idea. Before that, he was a researcher at Google Brain and then at [[OpenAI]], where he worked on post-training and is one of the primary authors of the [[InstructGPT]] paper (Ouyang et al., arXiv:2203.02155, 2022), the RLHF recipe that led to [[ChatGPT]]. He's also credited on the [[GPT4|GPT-4]] technical report and was an acknowledged contributor to ChatGPT. What makes him interesting to me: he helped build the recipe that made chatbots work, then spent years arguing it was the wrong recipe for automation. His launch post for Jev (15 September 2026) opens with a question he says he'd been asking for four years: "Models have been superhuman at chat for years, so where is all the automation?" ## Background - **Google Brain**, then **OpenAI**. At OpenAI, he co-authored *A Generalizable Approach to Learning Optimizers* (Almeida, Winter, Tang and Zaremba, arXiv:2106.00958, 2021), a system that learns to tune optimizer hyperparameters and reported 2x speedups on ImageNet and 2.5x on a language modeling task - **[[InstructGPT]]** (2022): one of the starred primary authors, fourth in the author list. TypeSafe markets him as having "co-invented RLHF and InstructGPT". Worth knowing: RLHF as an idea goes back to Christiano et al. in 2017 (arXiv:1706.03741); InstructGPT is where it became the standard recipe for assistants - **[[TypeSafe AI]]** (founded around mid-2024 according to explainx.ai [UNVERIFIED]; other coverage says two years in stealth): co-founded with Erik Gafni (CTO, repeat founder, early employee at Invitae and Freenome) and Sasha Sheng (COO, ex-Meta/FAIR research engineer). Out of stealth on 15 September 2026 with a $40M seed round led by DCVC. The team works in person in San Francisco ## Ideas he's known for - **"Build Prod, Not God."** TypeSafe's tagline. Stop chasing a model that does everything; build narrow, reliable components software can call. See [[Machine Native Intelligence]] - **RLHF as "a weird detour".** "Why do all LLMs require a human in the loop? The simple answer is we literally put them in the loop." In his view, preference training optimizes for what reads well to humans, which rewards [[AI Sycophancy]], overconfidence and [[Mode Collapse|mode dropping]]. "Overpromising is a feature... by design" - **Name the objective, not the algorithm.** "It's not about the PPO. That part doesn't matter." RLHF, [[Reinforcement Learning with Verifiable Rewards (RLVR)|RLVR]] and [[Reinforcement Learning for Calibrated Decisions (RLCD)|RLCD]] are defined by what they reward. "You get what you optimize for and the bitterest lesson in ML is that the most important part of it isn't ML at all" - **Data over compute.** TypeSafe calls itself "a data lab, not a model lab". Jev's training data is 100% synthetic. "Data is so unbelievably complicated, and that is what gets nines" - **Calibration over eloquence.** When a model says 80%, it should be right 80% of the time (see [[AI Model Calibration]]) - **Refusals are type errors.** "Refusal is just, like, obviously a type error." If a background dependency refuses, the software "stochastically breaks" - **Against public benchmarks.** "Public benchmarks are extremely gameable... even if they try not to, they still will." He prefers evaluating inside real workflows - **Jevons paradox.** He named Jev after William Stanley Jevons: "Jev is optimized for intelligence per dollar, hence the name." See [[Jevons Paradox]] ## Criticism His positions draw fair pushback. Critics point out that TypeSafe publishes no paper, no calibration curve and no ablation for RLCD, that its own evals use model-generated reference answers, and that the terms forbid public benchmarking (which sits awkwardly next to "trust us, it's calibrated"). The "co-inventor of RLHF" label oversells what a co-authorship on InstructGPT means. And plenty of people argue the underlying pieces (zero-shot classifiers, logit reading, prefix caching) existed long before Jev. ## My take I find his critique of RLHF convincing, mostly because the evidence comes from the big labs themselves (the GPT-4 report's calibration numbers, Anthropic's work on sycophancy). Whether TypeSafe's execution lives up to it is a separate question, and the secrecy doesn't help. But "optimize for what software needs, not for what humans like to read" is an idea I expect to outlive any single model. ## Quotes <!-- QueryToSerialize: LIST FROM #type/quote AND [[Diogo Almeida]] WHERE public_note = true SORT file.name ASC --> ## Books <!-- QueryToSerialize: LIST FROM #type/book AND [[Diogo Almeida]] WHERE public_note = true SORT file.name ASC --> ## References - https://typesafe.ai/team - https://typesafe.ai/blog/introducing-system-one-models-and-jev - https://typesafe.ai/manifesto - https://www.latent.space/p/jev - https://ai.engineer/speakers/diogo-almeida - https://www.explainx.ai/blog/diogo-almeida-typesafe-ai-rlhf-detour-profile-2026 - https://arxiv.org/abs/2203.02155 - https://arxiv.org/abs/2106.00958 - https://www.morningstar.com/news/business-wire/20260915525333/typesafe-ai-emerges-from-stealth-with-40m-in-funding-with-new-model-for-composable-ai ## Related - [[TypeSafe AI]] - [[Jev]] - [[InstructGPT]] - [[Reinforcement Learning From Human Feedback (RLHF)]] - [[Reinforcement Learning for Calibrated Decisions (RLCD)]] - [[System One Models]] - [[OpenAI]]