Latent Space: The AI Engineer Podcast

Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI

2h 21min

In this episode, Diogo Almeida, CEO of TypeSafe AI, introduces *Jev*, a *System One model* designed for *programmatic use* by code, not humans. He argues that current LLMs optimized for human preference via RLHF are *fractured* and unreliable for software. Jev focuses on *calibration, cost, and reliability*, outputting structured decisions (choices, scores, booleans). Diogo also shares his vision of *machine-native composable AI* as an *infrastructure layer* that will *automate the easy work* and *unlock economic growth*.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

System One Models

TypeSafe AI launched Jev, a new class of model designed for *code as the consumer*, outputting structured primitives (choice, score, boolean/noole) instead of text, making it *reliable* and *cost-efficient*.
2

RLHF is Fractured

Diogo argues RLHF-based models are *overconfident* and *mode-collapsed*, optimized for human preference over *software needs*. Jev focuses on *calibration* and *robustness*.
3

No Safety Alignment in APIs

Diogo strongly opposes safety refusals in APIs, calling it a “*type error*” that breaks dependencies. He insists on *neutral infrastructure* like a database.
4

Intelligence per Dollar Frontier

TypeSafe is obsessed with *cost-performance*, using *synthetic data* and *internal evals* to push the *Pareto frontier* of intelligence per dollar, not speed per se.
5

Decomposable Decisions

Diogo advocates breaking tasks into *small, independent yes/no questions* rather than using long prompts. This approach is *cheaper* and yields *better reliability*.

💬 Notable Quotes

"If you want to automate the easy work before the hard work, this is about making AI *reliable enough* that developers trust it like a database query."
"Safety alignment makes sense for a product like ChatGPT, but in an API it’s nuts. Software should not *stochastically break* because a user sent a weird message."
"We want to be the *infrastructure layer* of intelligence—like TCP to UDP. The code is the final consumer, not the human."