AI
The Model That Speaks Computer Instead of English
The man who taught AI to talk to humans now says the talking was the detour. Diogo Almeida helped build ChatGPT and co invented reinforcement learning from human feedback, the technique behind the chatbot era. Two years ago he left OpenAI, and this week his startup TypeSafe AI came out of stealth with 40 million dollars in funding and a model called Jev that outputs probabilities instead of prose. Where a chatbot would write you a paragraph, Jev hands back calibrated probability scores that software can act on directly.
The design starts from a simple observation Almeida shared with TechCrunch: four years of optimizing for human language produced brilliant conversationalists, and computers speak a different language. Every chatbot answer has to be parsed, checked, and translated back into something code can use, and that translation layer is where cost, latency, and hallucinations pile up. Jev skips it. Developers define the outputs in advance, a straight probability, a pick from a list, a score on a scale, and the model fills them in, hundreds at a time, in parallel, in under 100 milliseconds.
The early numbers are startling, and they arrive with the standard caveat that they are the company's own. In a demo on the TypeSafe site, Jev returned an answer in 0.114 seconds where a flagship model took 8.566 seconds. Pricing is listed at just over four cents per million input tokens with output free, against two dollars in and twelve dollars out for the comparison model. SiliconANGLE notes the company claims up to 100 times faster and cheaper operation overall. Independent benchmarks have yet to weigh in, and the figures will move with the workload, and the direction is the story.
Real developers are already putting it to work. Pranit Sharma, an engineer at Vercel, said his team replaced a ChatGPT Luna 5.6 classifier that screened commands for safety with Jev and got results five to 18 times faster with greater accuracy. Bryo AI chief technology officer Nikhil Mudholkar tested Jev against Gemini on classifying business emails and found Gemini slightly more accurate but 10 to 20 times more expensive, with Jev the only model handing back a real probability, which he called ideal for automating workflows. Demand ran so hot this week that TypeSafe briefly lost the ability to serve users from its API.
The name is the thesis. Jev honors William Stanley Jevons, the 19th century economist whose paradox says that when a resource gets cheaper, the world uses more of it. Almeida's bet is that intelligence priced like a software primitive, composable, testable, layerable, ends up everywhere, humming quietly inside programs the way the early internet spread, instead of living inside a few giant chat apps. Armin Ronacher, whose open source harness Pi watches this space closely, expects competitors to follow now that the utility is visible.
If the chatbot era was about making machines sound human, the next one may be about making them useful to machines, and the man who started the first era just placed his bet on the second. For readers who build, the takeaway is concrete: the cheapest intelligence may be the kind that skipped the small talk.
Quick answers
What is this story about?
The man who taught AI to talk to humans now says the talking was the detour. Diogo Almeida helped build ChatGPT and co invented reinforcement learning from human feedback, the technique behind the chatbot era. Two years ago he left OpenAI, and this week his startup TypeSafe AI came out of stealth with 40 million dollars in funding and a model called Jev that outputs probabilities instead of prose. Where a chatbot would write you a paragraph, Jev hands back calibrated probability scores that software can act on directly.
Why does this story matter?
If the chatbot era was about making machines sound human, the next one may be about making them useful to machines, and the man who started the first era just placed his bet on the second. For readers who build, the takeaway is concrete: the cheapest intelligence may be the kind that skipped the small talk.
Sources
- TechCrunch: TypeSafe AI's Jev model
- SiliconANGLE: TypeSafe exits stealth with $40M
- The Register: Jev demo and pricing
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.