The AI Daily Brief: Artificial Intelligence News and Analysis

Why a New Class of AI “Judgment Models” Could Have Big Business Implications

September 16, 2026

AI Summary

5 min read

A New Class of AI Models Built for Decisions, Not Text

Diogo Almeida, who helped invent ChatGPT, spent two years in stealth building something that is not another large language model. His company TypeSafe released JEV, what they call a "judgment model" that outputs probabilities rather than text. Instead of generating a paragraph explaining whether a customer sounds angry, JEV returns a single number like 0.9—meaning there is a 90% probability the answer is yes. The model is 20 to 200 times faster and 40 to 400 times cheaper than comparable LLM calls, with output tokens free. Those numbers sound implausible for another LLM, but JEV is a fundamentally different architecture.

How Judgment Models Work

The training method is called reinforcement learning for calibrated decisions (RLCD). Where LLMs optimize for human preference—responses that human raters find fluent and helpful—JEV optimizes for "epistemically honest probabilities." It answers narrowly defined questions with scores, categories, or probabilities that software can use directly in if-then logic.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of The AI Daily Brief: Artificial Intelligence News and Analysis

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 (00:11) **Introducing AI Judgment Models** - The host introduces a new class of AI models that output probabilities for specific questions rather than generating text, with today's example being TypeSafe's model JEV.
2 (00:43) **Headlines: Zuckerberg's AI Safety Response** - Mark Zuckerberg posts on X arguing that AI safety pacing should be handled by individual labs rather than collective government action.
3 (01:24) **Headlines: Bannon-Sanders AI Alliance** - Bernie Sanders and Steve Bannon jointly call for human-centric AI regulation at the Future of Life Institute's Pro Human Assembly.
4 (02:08) **Headlines: Salesforce Dreamforce AI Moves** - Salesforce unveils a new in-house model (COA, fine-tuned on Nvidia's Nemo Tron) and the AIFORS connector initiative.
5 (02:52) **Main Episode: The JEV Model's Core Mechanism** - The host introduces JEV from TypeSafe, a judgment model trained via Reinforcement Learning for Calibrated Decisions (RLCD) that outputs epistemically honest probabilities rather than text.
6 (03:28) **What Judgment Models Replace** - JEV is not an LLM replacement but a tool for a category of work where LLMs are a "square peg mashed into a round hole."
7 (03:40) **Business Use Cases for Judgment Models** - Concrete applications in customer support, sales, marketing, and editorial workflows.

+ Full timestamped outline available in the app

Guests on this episode

Diogo Almeida

Show Notes

A new AI model called Jev is built to make fast, inexpensive judgments rather than generate text. NLW explores how this approach could reshape business automation, help agents check their work, and coordinate decisions across teams. In the headlines: Zuckerberg pushes back on a collective AI slowdown, Bernie Sanders and Steve Bannon find common ground on AI regulation, and Salesforce announces a new model and tools for third-party agents.

Multiplayer AI Sprint - ⁠⁠⁠⁠⁠⁠⁠⁠https://multiplayerai.ai/⁠⁠⁠⁠⁠⁠⁠⁠

Brought to you by:

KPMG – Research from KPMG and the University of Texas at Austin shows the highest-impact AI users treat AI like a reasoning partner — and those skills can be taught at scale. Learn more at ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://kpmg.com/us/Sophisticated⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠

Harbor - Invest in the AI ecosystem. ⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠https://www.harborcapital.com/aidaily⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠⁠