Because Jev generates all its outputs in a single query rather than one token at a time, it operates at blistering speeds and radically different economics.
If you’re feeling a little fatigued by every new AI model promising to write better emails, craft perfect essays, or chat with you like a best friend, you aren’t alone. For years, the AI industry has been obsessed with building the ultimate digital conversationalist.
But Diogo Almeida, a former OpenAI researcher who helped build the underlying tech for ChatGPT, recently posed a fascinating question: “Models have been superhuman at chat for years, so where is all the automation?”
To answer that, his startup TypeSafe AI just emerged from two years in stealth with a completely different premise: Software doesn’t need a chatbot. It just needs a really fast decision-maker.
Enter Jev, an AI that has zero interest in talking to humans. Instead, it wants to talk to your code—and it wants to do it so cheaply that you stop thinking twice about using it.
Here is why Jev might be a bigger deal than the next shiny chatbot update.
Reinforcement Learning from Human Feedback (RLHF) has led to LLMs that are optimized for human preferences. This has led to models that are superhuman at instruction following, and are what we now call “chat.” Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.
a new model
We built a new class of models, System One Models, to be natively used by machines. We’re building with a new architecture, a new sampler, and a new training algorithm: Reinforcement Learning for Calibrated Decisions (RLCD). –Typeface.ai
The “System One” AI
Think of how you make decisions. If someone asks you to solve a complex math problem, you slow down, think it through step-by-step, and explain your work (System Two). But if someone throws a baseball at your head, you duck instantly without thinking (System One).
Right now, we are using System Two AI for everything.
If you ask a frontier model if a customer email sounds angry, it generates a beautifully written paragraph one word at a time: “Yes, the customer appears quite frustrated, and I highly recommend…” That’s great for a human reading a screen. But software doesn’t want a paragraph. Software wants this:
Urgent: 0.94
TypeSafe describes Jev as a “frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.” Instead of being trained to produce text that humans like to read (known as RLHF), Jev was trained using a new method called RLCD (Reinforcement Learning for Calibrated Decisions). It skips the conversational pleasantries and parallel-processes data to spit out structured, machine-readable values.
The Math is Mind-Blowing
Because Jev generates all its outputs in a single query rather than one token at a time, it operates at blistering speeds and radically different economics.
- Lightning Fast: While frontier models can take anywhere from 3 to 329 seconds to fully respond, TypeSafe claims Jev’s end-to-end response times range from 70ms to 500ms (up to 200x faster for decision tasks).
- Practically Free: Input tokens cost just $42 per billion. Output tokens? Completely free.
To prove how this changes the game, TypeSafe’s team hooked Jev up to play the video game Doom. The AI was making 10 real-time queries every single second to navigate the game based on unstructured data. The cost for running an AI that fast and that continuously? About $7 an hour.
The Ultimate AI “Babysitter”
If you’ve ever built software with an LLM, you know the fear of “hallucinations”—when the AI suddenly apologizes, invents a new category, or spits out malformed code that crashes your app.
TypeSafe claims Jev has “zero type errors.” This isn’t just a marketing boast; according to their engineering team, because the schema matching is strictly guaranteed at the architecture level, it is mathematically impossible for Jev to return the wrong data structure.
Even better, it gives a calibrated confidence score for every single answer. As Almeida notes: “If a model can do a task 95% of the time but doesn’t say when it’s in the 5%, it can’t automate that task.”
This makes Jev the perfect “linter” or babysitter for bigger, more expensive AI agents. Imagine an autonomous AI drafting a contract. Before it sends the document, Jev can run instantly in the background, checking a dozen things at once:
- Did it follow brand guidelines? (99% confident: Yes)
- Is this action destructive? (98% confident: No)
- Does this need human review? (75% confident: Yes -> Send to a human)
The Jevons Paradox
If you’re wondering where the name “Jev” comes from, it’s a nod to the Jevons Paradox, a famous 19th-century economic theory. William Stanley Jevons noticed that when steam engines became more efficient, people didn’t burn less coal to save money. They just built way more steam engines, and coal consumption skyrocketed.
TypeSafe is making the exact same bet on AI.
If you make machine intelligence 100 times cheaper, businesses won’t just slash their computing bills. They will start putting AI in places where it previously would have been completely absurd—like checking every single line of data in real-time, or map-reducing petabytes of raw text into structured insights.
We might be entering an era where the most transformative AI isn’t the one you chat with—it’s the one quietly making billions of tiny, invisible decisions in the background.