If a website says Jev means “goodbye to LLMs,” close the damn tab

I find it hard to take an article seriously when it announces the end of LLMs and only then gets around to explaining what Jev actually is. It might be a great tool, save you money and take work off your hands. But there’s a pretty big gap between that and declaring language models obsolete.

Jev, TypeSafe AI’s model, takes information and returns categories, scores and probabilities your software can use. You define what to evaluate and what the application does with the result. The documentation manages to explain this without declaring anything dead.

Say a customer writes, “You’ve charged me twice.” Jev can classify that message as a billing issue. You still write the application rules for routing it to that team, asking for a review or authorizing a refund. We’d normally use a cheap model for that first classification, and Jev looks like it could handle the job accurately enough for less than the LLM call you’re currently using.

But the customer is still waiting for a reply. You need to check what they’ve been charged, find out whether there really is a duplicate, and explain what you’re going to do about it. You may need a language model to write that reply using the conversation so far. Jev isn’t going to write it for you, and TypeSafe itself makes that CRYSTAL CLEAR in its explanation of System One models.

There’s room in that process for Jev, an LLM and plain old code. That hardly seems unusual. What I don’t get is looking at the first step, finding another tool that can handle it, and assuming we’ve replaced everything else.

Besides, LLMs can return structured answers too. You need to check whether switching is actually worth it. If you’re handling thousands of messages, a difference in cost or response time can matter a lot, as long as you’ve checked that the answers are accurate.

TypeSafe promises “zero hallucinations” in its announcement, saying the answer always sticks to the allowed format. If the options are billing, support or sales, it won’t invent a fourth department. But it can still pick sales for the customer you’ve charged twice.

Their published tests compare four workflows and use other models’ answers as references. Before changing my customer support system, I’d want to try it on the messages that actually come in, including the vague ones, the ones with several problems mixed together, and the ones that are a mess. Just the sort of testing we’d do with any other tool.

That’s why the “goodbye to LLMs” headline annoys me. It drags us into arguing about whether one model replaces everything and pushes a more useful discussion aside. In my experience, plenty of people use general-purpose models for small, repetitive tasks, such as classifying messages, that Jev might handle better.

If Jev does part of the job better, give it that part. It takes a lot more than that to claim LLMs have stopped being useful. Meanwhile, I’d rather hear from someone who can show me how they’re using Jev and where it gets things wrong. Then I can decide whether it’s useful to me.

Thanks to José Luis Calvo (@joselcs) for pointing out the confusion between classification and decision-making. I’d also used “making decisions” too loosely, and I’ve corrected that part of the article.