Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
September 21, 2026
AI Summary
5 min readSystem One Models for Code, Not Chat
Diogo Almeida, CEO of TypeSafe AI, launched Jev—and it immediately took over the AI discourse. But he describes himself as "a ragged corpse" running a company that "never been worse" emotionally, because he's a technical CEO fighting fires while also trying to explain something he believes the entire field has been getting wrong.
Jev is not a chatbot. It is not another reasoning model. It is what Almeida calls a "system one model"—a class of model where the intended consumer is code, not humans. The name comes from Jevons's Paradox: as efficiency increases, consumption increases. Jev is optimized for intelligence per dollar, not for benchmark scores or conversational polish.
The Problem with RLHF and Mode Collapse
Almeida's critique of current AI runs deep. He argues that RLHF (reinforcement learning from human feedback) creates a fundamental problem called "mode dropping." Models trained this way become hyper-conservative: they collapse toward the most common, safest outputs because errors are easy to detect while subtle correctness is not. This makes them terrible for decision-making or programmatic use.
Continue reading the full summary in the app — free to try.
Read Full Summary →Free • No credit card required
Never miss an episode of Latent Space: The AI Engineer Podcast
Get every new episode summarized in your inbox — free, ~5 minutes to read.
No spam. Unsubscribe anytime.
What you'll learn
- 1 Timestamped Outline: Diogo Almeida on Jev, System One Models, and the Future of Programmable AI
- 2 (00:03) **Launch Energy and Emotional State** - Diogo describes the chaotic, exhilarating week since Jev launched, feeling like the AI field is finally aligned with reality
- 3 (03:04) **What is Jev? The System One Model Thesis** - Diogo defines Jev as a new class of model: machine-native, system one, large programmable models designed for code as the consumer
- 4 (05:36) **The Mode Dropping Problem with RLHF** - Diogo explains why RLHF causes calibration issues and why this matters for programmatic use
- 5 (09:10) **Pragmatic ML Philosophy vs. Scaling Law Dogma** - Diogo positions himself as a practical researcher who cares about what works, not what's theoretically optimal
- 6 (11:42) **Why Jev Doesn't Refuse: Safety Alignment vs. Capability Alignment** - Diogo makes the case against safety alignment in APIs, arguing it's fundamentally anti-developer
- 7 (17:11) **Anti-Benchmarking Philosophy** - Diogo explains why TypeSafe refuses to publish public benchmarks and how they build trust instead
+ Full timestamped outline available in the app
Guests on this episode
Show Notes
Tickets for AIE NYC now open, and apply for the invite-only AIE CODE. Join us!
We have an unusual relationship with today’s guest: for years since coauthoring the InstructGPT paper, Diogo Almeida had been saying that API-available frontier models have been going down the wrong path, everything from the alignment to refusals to reliability perspectives, that we have dropped every mode other than autoregressive chat-tuned LLMs because of the overwhelming success of ChatGPT.
In a launch video now viewed ~40M times (by comparison, GPT4o was 22M, Fable 5 was 15M, Navier Stokes was 74M, and 6 Astra was 137M), Diogo introduced Jev and it immediately took over the AI timeline — we’ll skip full Jev explainers because your favorite AI influencer/educator has probably already done one. We also collected:
* the official patterns and cookbooks you should see first, from Allie
* Jev usecases
* speed based - games and computer use
* the voice + computer use example we discuss at 1h34 mins
* The must not miss Doom demo
* Driving cars in games
* “Smart Games”/smart NPCs