The Architect's Heresy: Why ChatGPT's Co-Creator Built Its Opposite
Diogo Almeida, an OpenAI co-creator of RLHF (Reinforcement Learning from Human Feedback), the post-training algorithm behind most frontier AI, now critiques the very LLMs he helped build. His team trained the first models "superhuman at instruction following," laying the groundwork for chat-focused AI like ChatGPT.
Almeida argues that models optimized for human preferences, while great for conversation, are inherently unsuitable for robust automation. He highlights critical flaws: overconfidence, 'mode dropping,' and a pervasive lack of reliability. LLMs extract intelligence through the "tiny straw of autoregression," forcing sequential generation that demands a human-in-the-loop, hindering true autonomy.
His answer is Jev, TypeSafe AI’s inaugural System One model. Jev is engineered to directly address these limitations, built for "prod, not God"—prioritizing reliable, machine-native automation over human-like conversation. It outputs decisions with probabilities and confidence, aiming to be reliable, fast, self-consistent, and type-safe, unlike hallucination-prone LLMs.
Meet Jev: AI That Thinks Like Code, Not a Poet
Meet Jev, TypeSafe AI's answer to LLM limitations: a "System One model" built for machines. It abandons the sequential, auto-regressive generation common in LLMs—great for chat, but "totally useless for computers"—for parallel computation. This architectural shift delivers "near instant" answers, making real-time AI finally practical by extracting intelligence far more efficiently.
Jev's output isn't conversational prose. Instead, it returns structured decisions with explicit probabilities and confidence scores. This makes it "type-safe" and truly incapable of hallucination, functioning more like reliable, fast, self-consistent code. This machine-native format unlocks automation previously impossible with unpredictable LLMs.
Its training method, Reinforcement Learning for Calibrated Decisions (RLCD), represents a fundamental departure from RLHF. While RLHF optimizes LLMs for human preferences, often leading to overconfidence and unreliability, RLCD instead prioritizes accuracy and honest confidence. This design ensures Jev delivers robust, verifiable decisions, built for automation where reliability trumps pleasant conversation.
Ludicrous Speed: Real-World Demos and Shocking Costs
Jev's most eye-popping claims aren't about its smarts, but its raw efficiency. TypeSafe AI says their System One models are up to 200 times faster and 400 times cheaper than traditional LLMs. Specifically, input tokens cost a mere $42 per billion, with output tokens priced as free, making them "too cheap to meter." That's a serious cost reduction.
Developers immediately put Jev to the test, building viral demos in mere hours for pennies. Examples include Jev autonomously controlling agents in Minecraft, navigating a Tesla self-driving simulator, and even mastering Subway Surfers in real-time. These aren't just parlor tricks; they showcase Jev's ability to react to dynamic environments instantly.
The key unlock for creators is Jev's capacity to inject intelligent, reactive behavior into simulations and applications without the usual burden of specialized model training. This parallel computation approach means developers can prototype complex AI-driven systems with unprecedented speed and affordability. For more technical details on this shift, visit TypeSafe AI: Home.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The New AI Stack: Where Jev Replaces Your If-Statements
Jev isn't here to replace ChatGPT; it's a complementary tool for a distinctly different job. Large language models excel at complex, open-ended reasoning – the "brain" of an AI agent, if you will. Jev, however, steps in to manage the thousands of rapid, structured decisions within that agent's workflow, where instant responses and unwavering reliability are paramount.
For developers, Jev acts like a highly intelligent frontier-intelligence function call. Consider it a "smart if-statement" that finally makes robust AI automation practical. The previous sections covered its claims of being up to 200x faster and 400x cheaper, with input tokens priced at $42 per billion. These aren't just impressive numbers; they mean developers can build AI systems that respond in real-time without breaking the bank.
This isn't another conversational AI. Jev represents a fundamental shift towards machine-native intelligence. It’s a new layer of AI designed not for human chat, but for a world increasingly populated by automated systems, robots, and intelligent agents. It enables those systems to make instant, reliable choices, moving us beyond simple human-in-the-loop processes to truly autonomous operations.
Frequently Asked Questions
What is Jev AI?
Jev is a new 'System One model' from TypeSafe AI, designed for extremely fast and cheap structured decision-making within software. It is not a conversational chatbot like ChatGPT.
How is Jev different from ChatGPT?
Jev outputs structured decisions with probabilities, not free-form text. This makes it mathematically incapable of hallucinating, much faster (up to 200x), and cheaper (up to 400x) for automation tasks.
Who created Jev AI?
Jev was created by TypeSafe AI, a company co-founded by Diogo Almeida, who was a key researcher at OpenAI and a co-creator of RLHF, the training method behind ChatGPT.
What are the main use cases for Jev?
Jev excels at real-time tasks requiring rapid, reliable choices, such as controlling game characters, navigating drones in simulations, routing data, or powering any automated system that needs to make decisions quickly.