An ex-OpenAI researcher just deleted language from the LLM...

Overview

This episode explores Jev, a groundbreaking AI model developed by ex-OpenAI researcher Diogo Almeida. Unlike traditional large language models (LLMs), Jev eliminates language generation entirely, focusing on fast, type-safe, and cost-effective decision-making. The discussion highlights its potential applications, technical innovations, and the controversies surrounding its development.

Notable Quotes

- Large language models have one fatal flaw—they won't shut the hell up. – On the inefficiencies of traditional LLMs.

- Jev is 200 times faster, 400 times cheaper, with free output tokens and zero hallucinations. – On Jev's revolutionary efficiency.

- It’s so fast and cheap that you can even use it for real-time applications, like NPC behavior in video games. – On Jev's practical applications.

🚀 The Radical Innovation of Jev

- Jev is a System 1 AI model inspired by Daniel Kahneman's Thinking, Fast and Slow. It focuses on fast, instinctive decision-making rather than slow, deliberate reasoning.

- Unlike traditional LLMs, Jev doesn't generate language or code. Instead, it processes structured inputs and returns one of three predefined outputs: choice, score, or binary yes/no.

- Its type-safe design ensures schema-matching, making type errors mathematically impossible.

⚡ Speed, Cost, and Efficiency

- Jev is 200x faster and 400x cheaper than traditional LLMs, with zero hallucinations.

- Its efficiency makes it ideal for real-time applications, such as moderating content, implementing NPC behavior in video games, and even powering AI calculators.

- Developers have already integrated Jev into apps like Horse Tinder, where it moderates accounts by instantly identifying non-horse users.

🧠 Confidence and Accuracy

- Jev employs Reinforcement Learning for Calibrated Decisions (RLCD), providing a confidence score with every response. For example, a 60% confidence score means it’s correct 60% of the time.

- While type-safe, Jev is not deterministic, meaning identical inputs can yield different outputs.

🤔 Controversies and Open Questions

- Critics argue Jev is not fundamentally different from older zero-shot classifiers, with some claiming it builds on techniques pioneered over a decade ago.

- An open-source alternative, OpenJev, has already been developed, replicating Jev’s functionality using existing models like Quen 4B.

- The proprietary nature of Jev’s architecture has raised skepticism, as its creators have yet to release a detailed technical paper.

🛠️ Developer Implications

- Jev’s simplicity and cost-effectiveness open new possibilities for developers, particularly in real-time and resource-constrained applications.

- The emergence of open-source alternatives like OpenJev highlights the growing accessibility of similar technologies for independent developers.