The AI nerd world is in uproar again. This time the star is Jev from Typesafe AI. The premise is seductive: take a large language model, take away its ability to talk, strip out the expensive, slow text generation and train it to do nothing but make decisions. A pure-bred “System 1” model that delivers classifications and probabilities in 70 to 500 milliseconds. Output tokens cost nothing, because there aren't any.
Clever, no doubt about it. The ambiguity of language is radically reduced, the output trimmed down to usable values. You could say it's an attempt to stamp out the probabilistic nature of AI and make it appear more deterministic. And that word, appear, is exactly where your ears should prick up.
The analysis: what is Jev, really?
Looked at soberly, Jev is an AI that has been trained out of talking. Instead of paragraphs, it delivers categories, scores and probabilities. For software automation, that's a perfectly sensible interface: software doesn't need poetry, it needs values an if statement can actually work with.
But, and this is where the trouble starts, the vendor claims the model cannot hallucinate. I think that's nonsense. Technically it's true insofar as the model can't break out of the predefined structure: no invalid data, no made-up categories. In terms of content, though, it remains a stochastic system that can take a wrong turn. Just silently and very fast. Jev doesn't solve the problem of unpredictability; it merely hides it behind a deterministic interface. It confuses output format (what gets returned) with process logic (how the decision comes about).
And that may well be more dangerous than classic hallucination:
- The chatty LLM writes: “The customer wants to cancel because his hamster was turned down for a loan.” Any human spots the nonsense instantly and steps in. The hallucination comes wearing its own hi-vis vest.
- The silent Jev returns:
Category: account closure, confidence: 0.98. No reasoning, no context, no visible nonsense. The automated process keeps running, and nobody notices until the train comes off the rails.
A wrong decision in deterministic packaging is harder to spot than wrong text. The probabilistic doesn't disappear. It just hides behind a mask of clean numbers.
On top of that:
For many of the advertised use cases (spam detection, ticket routing, classification), classic IT has long had proven solutions built for determinism.
Rule engines, fine-tuned specialist models, good old heuristics. Whether Jev actually beats them or just looks more elegant is a question nobody has yet answered with solid data.
Before we inject the complexity of a stochastic decision unit into our core processes, we have to ask ourselves: are we solving a problem here, or are we just buying a more modern aesthetic for something we could have solved better, cheaper and more safely with a simple if-else statement?
Anatomy of a hype: from POC to open question
Jev follows a pattern we know well by now, one already dissected in “The POC Trap”: “An interesting proof of concept, certainly useful in some scenarios. The only question: how do you actually put something like this into production?”
Jev is the latest bandwagon rolling through the AI village. A stealth startup, an impressive demo, cheering timelines, and zero solid answers to the questions of production readiness. We're right back in the proof-of-concept loop: the tech industry tosses a prototype over the fence, and we get to play unpaid guinea pigs figuring out where it falls apart in practice. This time with a spicy twist, though: with a chatty chatbot, we at least notice when it falls apart. With a silent decision machine, we only notice from the consequences.
The question of production use goes unanswered amid all the cheering. And on closer inspection it breaks down into several uncomfortable sub-questions:
- Accountability: Who is liable when the silent decision was wrong? With text output there's at least a traceable “justification”. With a naked confidence score of 0.98 there's nothing to audit.
- Calibration: Typesafe AI advertises “RLCD” for “calibrated decisions” (Reinforcement Learning for Calibrated Decisions. Top-notch marketing gibberish!). But calibrated on whose data? Whether the confidence scores hold up on your support tickets, your logs, your domain, nobody knows. Unless you test it yourself, thoroughly and continuously.
- Dependency: Building a proprietary model from a stealth startup, with an architecture that isn't public, deep into your production processes as a decision component is vendor lock-in in its purest form. What happens if Typesafe AI gets acquired, repriced or shut down in 18 months?
- Regulation: A system that makes automated decisions about people (ticket prioritisation, PII detection, risk scores) quickly lands in the higher risk categories of the EU AI Act. How do you document a model whose inner workings the vendor deliberately keeps secret?
These aren't rhetorical questions I'm using to give Jev a slap. They're the questions that have to be answered before any production deployment, and that the hype cycle systematically skips.
Back to the moving target
Which brings us back to my thesis from “AI as a Moving Target”: yet another interesting development, yet another uproar in the nerd world. Anyone who started a project six months ago to train a BERT model for exactly these kinds of classification tasks is rubbing their eyes today. And anyone who bets everything on Jev today will be rubbing theirs in a few months (weeks?), when the open-source wave of “System 1” clones rolls in, which is entirely foreseeable.
More developments are coming; that's the only certainty. When and how to embed tools like this into a sound AI strategy is another story entirely, and that one hasn't been written yet.
The bottom line: vigilance, and the eternal leitmotif
Once again, I'm deliberately not making a recommendation here, neither for nor against Jev. The point is a different one: Stay vigilant about every AI hype.
When a vendor claims its model “cannot hallucinate” (or could wipe out humanity), that's not a technical fact, it's a marketing line. When an impressive POC sweeps through the timelines, the right reaction isn't “build it in!” but the sober question: In which scenario does this solve a problem we actually have, and what does the mistake cost us when the silent decision is wrong? Or put another way: don't let the tech industry's POCs push you around. Build your own, on your data, in your processes.
Because one thing applies to Jev just as it does to every tool: an AI that decides faster makes mistakes faster. And an AI that stays silent while doing so makes them more quietly. Anyone who plugs Jev into a complex, undocumented workflow doesn't have a smarter process. They just have a process that now produces garbage in 70 milliseconds, with a confidence score of 0.98 slapped on top as a seal of quality.
Chatty or silent AI, System 1 or System 2: the tool is never the strategy. Let the hypes be hypes, go back to the whiteboard and take a hard look at your processes. Because in the end, what always applies still applies:
It's always the process, stupid!
Live long and prosper 😉🖖
