Why reject public benchmarks? That looks like a vendor convenience.

The source records the reasoning, not just the decision. It says he argues you get what you optimize for, that TypeSafe sees itself as a data lab rather than a model lab, and that it optimizes for intelligence per dollar instead. It is still the founder's position.

Is RLCD described in enough detail for me to build on it?

No. The source calls it a novel, unpublished technique that optimizes for epistemically honest probabilities on System One tasks. There is no objective, data recipe, or algorithm in the interview you could implement.

How does he separate RLCD from RLHF and RLVR?

By the north star each optimizes. The source says RLHF on human preference causes hallucination, sycophancy and permanent reliance on humans, while RLVR on verifiable rubrics solves Navier Stokes but worsens jagged intelligence and does not integrate well with other software.

Should I read the post or the YouTube cut?

They are the same interview in two forms. The caveats say the written post and the video cover the same conversation, and the entry lists the video separately. Take one unless you want both the timestamped and the written version.

Who is the interview guest and what did he coauthor?

Diogo Almeida, CEO of TypeSafe AI, who coauthored the InstructGPT paper and spent two years in stealth building RLCD and Jev.

What is the interview's framing of Jev?

System One models for prod, not god; the episode is described as the definitive Jev podcast with its lead creator, running about two hours and twenty minutes.

What resources does the interview round up?

Official patterns and cookbooks, use cases such as games and computer use, the Doom demo and Excalidraw virtual try-ons, Jev's official guide for coding agents, and Diogo's note on the Tyranny of the KV Cache.

CommunityJev Guides & Articles

System One models for prod, not god

swyx's Latent Space interview with Diogo Almeida, posted on Hacker News as the first long technical conversation with TypeSafe's CEO.

See highlights

Open latent.space

Format: Article
Published by: Community
Author: swyx
Last verified: Oct 2026

Tagscommunity article explainer

Works forJev for Routing

Used inJev for Software & DevTools

Highlights

Hacker News title: System One Models for Prod, Not God. 4 points on 21 September 2026.
swyx's comment calls it the first in-depth technical interview with Diogo he knows of.
The YouTube cut of the same conversation is a separate entry.
Distinct from Latent Space's AINews launch roundup.

Watch out

The written post and the YouTube video cover the same interview. Point totals on the thread will move.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Hacker News: swyx posts the Latent Space interview with Diogo Almeida.

swyx4 pointsJev: System One Models for Prod, Not God – With Diogo Almeida, CEO, TypeSafe AIRead the thread on Hacker News

Hacker News: Simon Willison's overview of Noul, Choice, and Score, and where a decision model fits.

benwerd53 pointsJev introduces a new shape of LLMRead the thread on Hacker News

X: Diogo Almeida in five minutes

X: LLMs vs. Jev, clearly explained

More Jev resources like this

OpenRouter: what Jev returns

Kenny Rogers explains Jev as a decision model and shows Choice, Score, and Noul responses from calls he placed through OpenRouter on 21 September 2026.