From TechCrunch · Tim Fernholz · · 2m
ChatGPT inventor's new AI model skips language entirely
Diogo Almeida left OpenAI to build Jev, a model that outputs probabilities instead of text, making AI cheaper, faster, and hallucination-free for software automation.
In brief
Diogo Almeida left OpenAI to build Jev, a model that outputs probabilities instead of text, making AI cheaper, faster, and hallucination-free for software automation. By abandoning natural language output in favor of calibrated probabilities, Jev achieves 5-18x faster performance and lower costs than LLMs for specific…
ChatGPT inventor's disappointment
Diogo Almeida helped build ChatGPT and RLHF, but felt the technology wasn't useful for automation because AI was optimized for human language, not computer tasks.
“We have lightning in a bottle, and yet it is not useful.”
TypeSafe AI launches Jev
Two years after leaving OpenAI, Almeida founded TypeSafe AI and released Jev, a transformer model that outputs probabilities instead of text, eliminating hallucinations entirely.
No text output means no hallucinations
Because Jev outputs predefined probabilities rather than generated text, it cannot hallucinate. Users define outputs in advance, making the model reliable for automation.
Dramatically cheaper and faster
Jev's output tokens are free and input tokens cost per billion instead of per million. Demand was so high the company briefly lost API availability.
Vercel sees 5-18x speed gains
Vercel replaced OpenAI's ChatGPT Luna 5.6 with Jev for safety classification and got results 5 to 18 times faster with greater accuracy.
Confidence scores enable automation
Jev returns real probability scores that developers can use to decide whether to act on results. Bryo AI found Jev 10-20x cheaper than Gemini for email classification.
“it is the only one that hands back a real probability which makes it ideal for automating workflows!!”
Monitoring AI agents affordably
Jev can monitor other AI agents to prevent jailbreaks and misbehavior. Using expensive LLMs for this task is costly, but Jev makes agent-monitoring economically sensible.
Model routing becomes practical
Jev can predict which workload requires which model in real-time. This routing task would be expensive with LLMs but is now feasible with Jev's low cost and speed.
Named after an economic paradox
Jev is named after William Stanley Jevons, whose paradox describes how falling costs lead to increased use. Almeida hopes cheap AI will be deployed everywhere.
Trained on synthetic data exclusively
Jev uses reinforcement learning from calibrated decisions on synthetic data. Almeida calls this bet better than RLHF and says half the company is a lab focused on this technique.
“Half of [our company] is a lab that basically owns this entire subfield of statistically well-understood synthetic data”
Competitors will follow soon
Jev currently stands alone, but observers expect competitors to emerge now that its utility is proven. LLM subsidies have discouraged this kind of innovation until now.
“We should have seen this earlier in many ways, but presumably because the LLMs are so cheap and subsidized, you often don't have to be creative yet”
The essence
By abandoning natural language output in favor of calibrated probabilities, Jev achieves 5-18x faster performance and lower costs than LLMs for specific automation tasks.
