📰 Key Takeaways

ChatGPT left OpenAI researcher Diogo Almeida disillusioned. He helped build ChatGPT and invented RLHF (Reinforcement Learning from Human Feedback), the core training technique driving the current AI wave, yet he’s still unhappy with the results. He believes the whole industry has spent the past four years optimizing for “human language” output, when what computers actually need is a different kind of language altogether — which is why LLMs remain hard to truly use for automation.

Two years ago Almeida left OpenAI to found TypeSafe AI, trying to solve exactly this problem. This week the company unveiled its new model, Jev, which uses a transformer architecture but isn’t an LLM — it doesn’t output text, it outputs probabilities, what the company calls “calibrated decisions.” Ditching text output makes the model extremely cheap and fast, and since the output options are predefined by the user, it can’t hallucinate. Output tokens are free; input tokens are billed by the billion instead of the million.

Developer response has been intense — the company was briefly unable to keep up with API demand. Vercel engineer Pranit Sharma swapped Jev in for an instruction-safety classifier that had been running on OpenAI’s Luna 5.6, and saw a 5-to-18x speedup with better accuracy. Bryo AI CTO Nikhil Mudholkar compared Jev against Gemini for classifying commercial emails: Gemini was slightly more accurate, but cost 10-20x more — and he valued Jev’s ability to return real probability values, which are useful for setting confidence thresholds in automated workflows.

Beyond replacing LLMs in specific use cases, Jev can also serve as a cheap oversight mechanism for supervising LLM agent behavior — tracking agent trajectories, guarding against jailbreak attacks, and helping with model-routing decisions.

💬 JudyAI Lab Take

Diogo Almeida helped build ChatGPT and invented RLHF — and now he thinks the whole industry has spent the past four years headed down the wrong path, optimizing for human-language output while missing what computers actually need: a different kind of language entirely. After leaving OpenAI, he founded TypeSafe AI, and this week’s new model, Jev, ditches text output in favor of calibrated probability values.

This points to a design idea worth paying attention to if you’re building with AI: not every use case needs an LLM generating natural language. When the output options can be predefined, probability outputs eliminate hallucination entirely while cutting cost and latency dramatically. In Vercel’s real-world test, swapping an instruction-safety classifier over to Jev delivered a 5-to-18x speedup. In another case, Gemini was slightly more accurate but cost 10-20x more, and Jev’s ability to return real probability values made it a better fit for confidence-threshold decisions in automated workflows. The reminder here: beyond chasing stronger language generation, “using the right tool for the job” matters just as much.

Next time you’re designing an AI pipeline, it’s worth asking first: does this task actually need text output, or would a probability judgment do?

📅 Original Article Info

🔗 Further Reading