Jev: turning 'smart but useless' AI into software that acts
Key points
Key takeaways from Source interview with TypeSafe AI Co-Founder & CEO Diogo Almeida (September ️2026):
Motivation. Almeida's elevator pitch for Jev is "where is all the automation?" — AI is extraordinarily smart but useless at everything except chat and code; TypeSafe aims to make AI build real software and drive automation, starting with Jev, its first System One model.
Jev vs coding agents. Cloud Code, Codex and Cursor write the same code as a human would (Gary Tan calls it "just-in-time software"); Jev is a new primitive embedded in your code — an intelligent layer that takes natural-language descriptions of intent and a state machine, returning typed decisions with calibrated confidence, so it expands what software itself can do, rather than just how fast engineers write it.
Absolutely a classifier. Jev is built from classifiers, designed to be useful, sharing the interface of classic ML tools; Almeida suspects it outperforms hiring an MLE team from2019 — and you can program it on the fly.
Reliability. TypeSafe cares about reliability across three senses: uptime or SLAs, determinism (useful for unit tests)and robustness — similar intelligence every time; its ambition is developers programming against Jev without example queries, trusting it the way a team trusts a colleague.
Agents and architecture. Coding agents are strong at syntax but weak at semantics and "incredibly bad at architecture",which he calls the most human creative part of software; Jev is probably not in today's agents' training distribution yet, but once it is, agents writing syntax while humans own architecture is the likely split.
SaaS-palooza. The "SaaS apocalypse" thesis hasn't panned out —the average PR is about10 lines, so generating software alone adds little new capability; Almeida expects SaaS to be among the biggest winners of the AI wave, because incumbents know the workflows worth automating and have customer reach — in his words, an "inverse apocalypse".
From mathlete to OpenAI. An award-winning mathlete and Kaggle winner, Almeida moved from math to computer science, worked with Jeremy Howard, spent time at Google Brain, briefly retired, then joined OpenAI —where RLHF's generalisation genuinely surprised him,and he once believed a model he helped ship had a real chance of being AGI.
"Do what I mean". His end-stateis AI quietly running inside products — "if all technology just did what you mean"; he foreseesan era of probabilistic programming,and re-architected systems,the way mainframes gave way to client-server,andthe internet,with "intelligence per dollar" as his northstar.
Read more: YouTube (a16z) · TypeSafe AI
