Diogo Almeida joins swyx on Latent Space days after Jev's launch, which we covered when TypeSafe put the decision model in front of developers. Almeida helped build instruction-following models at OpenAI, and the episode opens on the question he spent years on afterwards: why AI can solve extraordinarily hard problems yet still automates so little real work.
His answer is that the next wave will not look like chat. Almeida calls TypeSafe's approach System One Models — machine-native AI designed for code, calibrated decisions, reliability and intelligence per dollar, rather than for conversation. The episode separates that from the reasoning-model lane, discussing System One against System Two intelligence and where reasoning models stop being the right tool.
TypeSafe does not optimise around public benchmarks. Almeida argues the measure that matters is intelligence per dollar, and that the discipline of picking the right task and the right data beats simply scaling compute — what he calls the "bitterest lesson." It is also why he describes TypeSafe as a data lab rather than a model lab, and why he says he would not pre-train from scratch even if handed a billion dollars.
On method, he sets RLCD against RLHF and RLVR as fundamentally different north stars, and is direct about RLHF's failure modes: mode collapse, weak calibration, and the costs hidden inside optimising for human preference.
The safety discussion turns on placement. Once a model is buried inside a software dependency rather than sitting behind a chat box, refusals and alignment become a different kind of problem — a refusal is no longer a polite decline to a user but a failure in someone's control flow.
That shapes the engineering advice. Jev's programming primitives map intelligence into software control flow, pushing developers to decompose AI workflows into many small, measurable decisions, with structured state replacing giant prompts and system messages. Almeida also makes the case that Jev changes the architecture of coding agents built around a single model, and sketches an "inverse SaaS-pocalypse" in which AI supercharges existing software instead of replacing it — with AI eventually disappearing into the background.
Sources: Watch on YouTube · Latent Space
