Tossing a coin is type safe

Jev has been making a bit of a splash this week.

It addresses one of the major pain points of current LLMs … incessant verbal diarrhoea.

Ask Fable or Astra a yes or no question, and you’re likely to burn through tens of thousands of tokens and receive an essay response.

So Jev takes a different approach. First it’s not an LLM — it’s a classification model. Second, it’s type safe. You provide it with a state and list of possible answer types and it guarantees the output to match one of the options you provided. Third, it’s fast and cheap — it claims to be hundreds of times faster and cheaper than Sonnet 5.

Once again though, reality seems to be getting lost in the hype. Type safety is nice, but safety is not correctness.

You have guarantees about the type of output it returns. But it doesn’t guarantee correctness.

Tossing a coin is type safe. And if you believe that that’s enough to build production systems, then I have a bridge to send you.

I’m not trying to throw water on Jev here. A generalised hotdog or not hotdog can actually be a pretty useful service. But the real question yet to be answered is how correct are the results.