One of the most interesting AI announcements I’ve seen in a while came this week from Diogo Almeida (co-inventor of ChatGPT and RLHF) and TypeSafe AI.
They’ve released Jev, a new kind of model designed not to generate language, but to make fast, structured decisions.
That distinction could be much more important than it sounds.
Think about a very simple task: classifying an incoming customer support ticket.
Today, we might send the entire ticket to an LLM and ask:
“Is this billing, technical support, cancellation or something else?”
The model reads the input, then autoregressively generates tokens, perhaps producing something like “This appears to be a billing-related request…”, which we then parse back into the thing the computer actually wanted in the first place:
Billing: 94%
There is an extraordinary amount of machinery involved in generating language that nobody actually needed.
Jev takes a very different approach. You give it unstructured state and a set of typed questions with a list of possible answers (e.g. {billing, technical_support, cancellation, other}), and it returns probabilistic decisions directly. Crucially, those decisions can be generated in parallel rather than token by token.
TypeSafe claims this makes Jev 20–200× faster and 40–400× cheaper than frontier conversational models for suitable tasks, with latency measured in tens or hundreds of milliseconds.
Perhaps even more interesting to me is how Jev is trained.
Diogo calls the approach RLCD: Reinforcement Learning for Calibrated Decisions. Rather than optimising a model to produce outputs humans prefer, as in RLHF, the objective is to produce decisions with useful, calibrated probabilities, so that software can reason about the model’s uncertainty and decide when to act automatically and when to escalate.
There are important caveats. Jev isn’t an LLM replacement and it can’t generate arbitrary text. And “hallucination-free” shouldn’t be confused with “incapable of being wrong”: restricting a model to typed outputs prevents it inventing an invalid answer, but it can still select the wrong valid answer.
But that’s precisely why I find this so interesting.
We may be entering a world where we stop asking giant generative models to do every kind of intelligence task, and instead have specialised models performing millions or billions of cheap, fast decisions inside software.
And there’s another broader trend here that I think is worth paying attention to:
reinforcement learning is escaping the box we put it in.
Not so long ago, RL was still widely associated with Atari, Go, simulated environments and robotics research. Increasingly, it is becoming a practical mechanism for training and optimising real-world AI systems.
https://lnkd.in/eCRBncrz
![]()
17
1 Comment
