Diogo Almeida, the Ateneo maths kid behind Jev’s $7.5B valuation

The bronze that ended a 13-year wait

The official IMO results for the 2008 Philippine team list him under his full name, Diogo Miguel Moitinho de Almeida. He scored 7 points on problem one, 7 on problem two and 1 on problem five, for 15 points, a bronze medal and 238th place. He was the top scorer on a three-person team.

The IMO’s Philippine country record shows the medal before his was a bronze for Bernard Chan in 1995. Nobody from the Philippines had medalled in between.

That was the story GMA News ran in August 2008, under the headline “2 Filipinos bag bronze medals in int’l Math, Physics Olympiads”. GMA reported the olympiad was held in Madrid from 10 to 22 July, with 515 contestants from 99 countries, and described Almeida, then 18, as one of “Two Filipino students” honoured that summer. The Department of Science and Technology gave him a Youth Excellence in Science medal.

He didn’t sound like someone taking a victory lap.

“Everyone’s congratulating me since the Philippine hadn’t gotten a medal in 13 years. I really thought I could have gotten silver (medal). I was one question away which is a big thing. I regret it so much.”

– Almeida, as he told GMA News in 2008

The medal wasn’t a one-off. Months earlier, the Australian Embassy in Manila had named “Diogo Moitinho de Almeida from Ateneo de Manila High School” among 23 Asia-Pacific students recognised for Australian Mathematics Competition results, in a release headlined “Australian Ambassador lauds young Filipino math wizards”.

Philippine media called him Filipino, he represented the Philippines, and he learnt his maths at Ateneo. I couldn’t find him saying where he was born or what passport he holds, and “Moitinho de Almeida” is a Portuguese surname. In the four long 2026 interviews I went through, he doesn’t mention the Philippines once. So I’ll stay with the record: Ateneo-schooled, Manila-trained and, in 2008, the top scorer on the Philippines’ IMO team and its first medallist in 13 years.

He isn’t sentimental about any of it. Asked about his background on an a16z video in September, he called himself an “award-winning mathlete”, then undercut it straight away: “I never really liked math. I never really tried. I was just like big fish in little pond.” What he liked was computer science, which he described as “basically like math, but cool and useful and fun and interesting”.

Kaggle, Google Brain and a year and a half off

After Ateneo he went to the United States. A 2016 speaker bio quoted from his Deep Learning Summit London talk says he took his undergraduate degree at Rensselaer Polytechnic Institute in New York, and joked that he’d taken more classes online than he did getting it.

His route into AI came through a competition rather than a lab. In 2013 he won the ChaLearn Cause-Effect Pairs Challenge on Kaggle, a contest to work out which of two variables causes the other. His write-up for the organisers says the pipeline finished second on the public leaderboard and first on the private one, which is the one that counts. On the a16z video he said the win came “not from sophisticated math”, but from automating as much of the problem as he could. The contest’s host, the machine-learning researcher Isabelle Guyon, pulled him into the research community, and he ended up speaking at NeurIPS.

From there came Enlitic, Jeremy Howard’s medical-imaging startup, where he co-authored the 2016 paper Resnet in Resnet under an Enlitic affiliation. By 2017 he was a research engineer at Google Brain, and his GOTO Amsterdam speaker bio that year still led with the olympiad: a medallist “ending a 13-year losing streak for the Philippines”.

Then, in 2018, he left Brain and stopped working. He’s described the next stretch in a few interviews, most vividly on the TWIML AI podcast in October.

“I was at Google and then I left. I was retired for a while just trying to like chill out and be happy. Did a lot of weird stuff like like very competitive video gaming and improv and stuff.”

– Almeida, as he told TWIML in October

“Very competitive” turns out to be literal. TypeSafe’s team page lists his fun fact as “the top 30 in League of Legends in NA and top 2 in Hearthstone”. The fan posts going round X this week have made the League ranking the headline. I find the Hearthstone line funnier, because it’s the less famous game and he ranked higher in it.

He came back because, as he told TWIML, “AI is way more fun than that”. He joined OpenAI in 2020, which he described on the same podcast as a “weirdass company that has weirdass goals”. It was, he noted, “before language models were really a thing”.

The InstructGPT years

His OpenAI paper trail is easy to follow on arXiv. In June 2021 he was first author of A Generalizable Approach to Learning Optimizers, written with Clemens Winter, Jie Tang and Wojciech Zaremba. The following March brought Training language models to follow instructions with human feedback, the InstructGPT paper, which used reinforcement learning from human feedback (RLHF) to teach GPT-3 models to follow instructions. Almeida is its fourth-listed author. The GPT-4 Technical Report names him among its contributors too.

DCVC, which later led TypeSafe’s seed, is blunter. Its investment note says he “co-invented some of the core techniques and products that launched the current boom”, naming InstructGPT, ChatGPT and GPT-4. His own Jev launch post is more modest. He says he “helped build the methods that made language models useful at following instructions and talking with people”, and that “That work ended up as the research behind ChatGPT.”

He has since built a company around a limit he sees in that work. At the AI Engineer World’s Fair in July, he opened by joking that he’s “one of the few people at OpenAI who actually hates on chat GPT”, then clarified within seconds: “I don’t hate chat GPT as a product, to be clear. I think chat GPT is a world-changing product.”

His complaint is narrower. RLHF trains a model to please the person reading its answer, and that’s the right goal for chat and coding assistants. Later in the same talk, he argued it’s the wrong goal for software that runs unattended.

“No matter how wrong the models are, they will look right because of the asymmetry within the reward model in RLHF.”

– Almeida, as he said in his AI Engineer talk

He describes the stakes in personal terms. On the a16z video he said he’d thought the instruction-following model he worked on “had like a decent chance of being AGI”, and when it didn’t, “that was like when my whole world came crashing down”. Speaking to swyx on Latent Space in September, he explained why he left: if an AI winter arrived and he hadn’t done “every possible thing I could to like avert that, I would see myself as personally responsible”. Later in the same conversation he admitted he “just really did feel disempowered” at OpenAI.

He wrote up the idea internally, he says, and recalls Sam Altman telling him “You should go work on it.” TechCrunch dates his departure to roughly two years before Jev’s launch. That fits the 2024 founding of TypeSafe with Erik Gafni, now chief technology officer, and Sasha Sheng, a former Meta research engineer who’s now chief operating officer.

Jev returns probabilities, not prose

Jev doesn’t write anything. In the launch post, Almeida says Jev is “a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out”. You hand it some state, such as a support ticket, a CV or a JSON object, plus questions with fixed answer types. It returns probabilities, not prose.

TypeSafe’s docs define three question types. Choice picks one option from a list and returns the pick, a probability for every option and a confidence score. Score rates the state against a rubric, with the same probabilities and confidence. Noul answers whether a statement is true with a number between 0 and 1.

All three can be mixed in a single call, and each question is evaluated in parallel against the same state, so adding questions “barely changes the response time”. Because the allowed answers are fixed in advance, TypeSafe says, the output can’t fall outside the schema the developer defined.

The company calls this class of model “System One”, after Daniel Kahneman’s split between fast, intuitive thinking and slow, deliberate reasoning in Thinking, Fast and Slow. Jev is named after William Stanley Jevons, the economist who noticed that more efficient steam engines led to more coal being burnt, not less. TypeSafe expects cheap machine judgement to go the same way, and Almeida’s own summary of the design goal is that Jev is “optimized for intelligence per dollar”.

Under the hood, the launch post describes a new architecture, a parallel sampler and a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions, or RLCD. In his July talk Almeida pitched it as a third option beside RLHF, which optimises for human preference, and RLVR, which optimises for verifiable correctness. RLCD, he said, is “optimized for calibrated decision-making”.

TypeSafe prices Jev to match that goal. Its launch post puts frontier LLM input prices at $0.20 to $10 per million tokens.

Jev itself costs $0.042 per million input tokens, or $42 per billion, and output is free.

TypeSafe also says Jev’s end-to-end response time is 70 to 500 milliseconds, which it puts at 40 to 200 times faster than frontier models on what it calls System One-shaped queries. The TypeSafe homepage headlines bigger numbers: 193.6 times faster and 444.6 times cheaper, based on its own workflow evaluations, which the launch post concedes are “on the higher end of real world gains”.

Twenty-four days from launch to $7.5 billion

TypeSafe came out of stealth on 15 September with a $40 million seed round led by DCVC, announced the same day as Jev. Forbes reported that the seed valued the company at $200 million, citing a person familiar with the deal.

Usage climbed quickly. Within 24 hours, Vercel said, nearly 13% of paid teams on its AI Gateway were using Jev, “2x the GPT-5.6 family and more than 6x Fable 5.1’s share”. By day three, according to Andreessen Horowitz, Jev had generated 1 trillion tokens. Demand briefly knocked out TypeSafe’s API in launch week, TechCrunch reported.

Understanding AI counted almost 40 million views of the announcement by the end of the first week, more than 100,000 Discord members and about 2,000 GitHub projects using the model. General signups opened on 20 September, paused two days later under what the company called “an immense swell of demand” and reopened on the 27th.

The valuation went from a reported $200 million at the seed to $7.5 billion at the Series A. Bloomberg reported that the deal “came together just weeks after the Jev launch video went viral”, which Almeida said brought a surge of interest from chief executives, developers and investors.

On Latent Space, he described the earlier raise very differently: “last year when we did fund raise no one believed us … they wanted just benchmarks and stuff”.

The adoption figures depend on who’s counting. TypeSafe’s announcement says “A third of the Fortune 500 are getting their Jev on” and that it has saved customers “millions of dollars in production”. Bloomberg’s report repeats “about a third”, adding that the startup declined to name any companies and that Jev “surpassed one million users in a matter of days”. The Andreessen Horowitz investment post says 25% of Fortune 500 enterprises have integrated Jev. The counts may simply have grown between drafts, but every one of them comes from the company or its lead investor.

Andreessen Horowitz says Jev is “that fastest growing model we’ve ever seen” and opens its post with this line.

“The Jev launch is the biggest narrative violation we’ve seen this year, in the best possible way.”

– Andreessen Horowitz, as it wrote in its investment post

It also says the firm has “long been fans of Diogo”, and frames TypeSafe as aligned with developers rather than with “the rhetoric around the big labs about replacing software engineers”. Martin Casado, who takes the board seat, leads a16z’s infrastructure practice and joined the firm in 2016.

For scale, a16z also led Thinking Machines Lab’s $2 billion seed at a $12 billion valuation in July 2025, and invested in Safe Superintelligence’s $1 billion first round, which Reuters’ source valued at $5 billion. Both labs were founded by better-known OpenAI alumni, and both raised before shipping a product. TypeSafe raised after launching Jev, with usage figures to show, most of them its own.

The caveats TypeSafe wrote down itself

TypeSafe’s launch post discloses that its workflow evaluations “were made by individuals on our model capabilities team, so some bias could exist”. On pricing, it admits the company “can’t prove it isn’t subsidized”. Its docs keep a page on where the current model struggles, listing literal readings, indirection and “tasks that require numeric precision”.

TechCrunch’s launch coverage reported that Vercel replaced an OpenAI safety classifier with Jev and got results five to 18 times faster, with better accuracy. Nikhil Mudholkar of Bryo AI found Gemini slightly more accurate on his email-classification test, but 10 to 20 times more expensive, and said Jev “is the only one that hands back a real probability”. Armin Ronacher, CTO of Earendil, said it “delegates the hallucination problem a little bit to the user”. The Register noted that structured probabilities can still be wrong.

TypeSafe’s own case study puts numbers on one switch. Recruiting startup Jack & Jill replaced Gemini 3.1 Flash Lite with Jev for every call in one stage of its candidate-matching pipeline, within 10 days of its first test. TypeSafe says the switch cut cost per 1,000 candidates scored from $0.755 to $0.092 and roughly halved median screening time, from 20.3 seconds to 10.3 seconds. Jev kept slightly more of the candidates hiring managers went on to pick, although the difference wasn’t statistically significant. According to the case study, Jack & Jill expects to save $265K a year at current volume.

Almeida is guarded about the architecture, which TechCrunch says outside observers suspect sits on an open-weight LLM, though he says it’s trained only on synthetic data. Other companies have since shipped similar models. Understanding AI notes that Cloudflare, Amazon and OpenAI have built models “with the same basic design”, while Jev’s performance “seems to be the strongest”. OpenAI’s version is the Decisions API, which I raced against Jev in a separate bake-off.

Still listed under PHI

The GMA report from 2008 has a second medallist in it. Tobit James Narciso of Philippine Science High School won bronze at the International Physics Olympiad that same summer.

“It shows that Filipinos can do this thing.”

– Tobit James Narciso, as he told GMA News in 2008

The report also says Almeida told GMA it had been his dream to bring that kind of recognition to the country. He hasn’t said anything like it publicly since, at least nowhere I could find, and he doesn’t owe anyone a flag in his bio. The Philippines sent him to Madrid, and the IMO lists him under PHI to this day. And the Ateneo student in that 2008 GMA report now runs a company that Andreessen Horowitz’s round values at $7.5 billion.

Narciso’s line has aged well. Almeida helped write the InstructGPT paper behind ChatGPT, left OpenAI to argue publicly that software running unattended needs a different kind of training, and now runs the company building it, backed by an $870 million Series A.

He still frames the work as a question. On TWIML he said that even at OpenAI, he’d been obsessed with one thing: “why is the AI so unbelievably good yet so useless for the easy work?” At 18, he regretted being one question away from silver. Twenty-four days after Jev’s launch, he has $870 million to work on this one.