He Helped Build ChatGPT. Now He Says Chat Was the Wrong Bet. | Edition 333

In partnership with

Wispr Flow turns your voice into clean, ready-to-send writing — speak naturally, it strips the filler and fixes the punctuation. I've used it daily since February 2026 to build this newsletter. Read the full review →

The man who helped build ChatGPT just bet against it.

Diogo Almeida worked on the instruction-following research at OpenAI that became the foundation of ChatGPT — the work that taught language models to actually follow what you asked instead of just predicting the next word. He thought it might lead to AGI. Two years ago, in stealth, he decided something important was missing.

On September 15, his new company, TypeSafe AI, launched Jev — a model that refuses to talk. Ask it a question and you get no prose, no code, no explanation. You get a typed answer — a choice, a score, a yes/no — with a confidence number attached, in well under a second. TypeSafe calls it a "System One Model," after Daniel Kahneman's term for fast, intuitive thinking, built to sit alongside slower "System Two" models like ChatGPT or Claude rather than replace them.

This guide distills 10 AI strategies from industry leaders that are transforming marketing.

Outpace your competitors by mastering AI.

Get Your Free Guide

What's actually new here isn't the speed. It's how it was trained. Every major model gets tuned against some reward signal — RLHF rewards answers human raters like, RLVR rewards answers that pass a test. TypeSafe trained Jev with something new: RLCD, Reinforcement Learning for Calibrated Decisions. The goal is a model whose stated confidence is actually trustworthy — if it says 80%, it should be right about 80% of the time, not just sound sure of itself. That's a real, previously underserved research gap: most labs optimize for "reads as correct," not "knows when it doesn't know."

The numbers TypeSafe is leading with — 40 to 200 times faster, up to roughly 444 times cheaper, $0.042 per million input tokens with free output, 70-500 millisecond response times — are all the company's own, run on its own benchmark suite, compared against itself. To TypeSafe's credit, its launch post is unusually candid about that: it admits it "can't prove it isn't subsidized" and that its reference answers come from averaging GPT-6 Astra and Claude Fable 5.1, which it says biases the comparison in Jev's favor.

The catch critics flagged fastest on Hacker News: "can't hallucinate" only means Jev can't return something outside the categories you gave it. It can still confidently pick the wrong one. And calling something that can't code, chat, or write a sentence a "frontier model" is borrowing credibility from a category it hasn't actually competed in.

The realistic use isn't replacing your LLM — it's putting something fast in front of it. Triage support tickets, flag a risky action before an agent takes it, score which leads to call first, sanity-check another model's output — then hand anything genuinely uncertain to a slower, smarter model, or a person. A fast filter in front of the model doing your real thinking, not instead of it.

Can you use it? Only through the waitlist. Jev is closed-weight and hosted-only — no self-hosting and no published architecture paper, and TypeSafe's own GitHub adapter repo held a single commit and a README on launch day. In the week since, access has widened and independent developer guides have started to appear.

Almeida is making a specific, testable bet: that a lot of AI's real value was sitting in cheap, bounded decisions inside software nobody talks to, not in a better conversation. The size of the seed round and the size of that Hacker News thread both say plenty of people think it's worth testing. Worth testing and proven aren't the same thing yet — every number above is TypeSafe grading its own homework.

You don't have to scroll every AI thread, track every new tool, or watch every demo. 

The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning. 

Trusted by 2M+ professionals at Apple, Google, and NASA.

Sign up to start learning.

— Jerry