100% agree, have found through tests a big discrepancy between models optimised for SWE benchmarks and usage in prod - there is a somewhat subjective element to it as well in terms of output that can't be quantified effectively (less so with decision models naturally)

Or, you know, it's an easy excuse because they're already behind on many of them.

They are? In all my evals at work, Jev seems to win, despite me wanting the open weight versions to.

I imagine if I fine tuned one of the open weight models for my use case it would be possible to beat it, though, but as a general drop in decision model, I’ve not really seen much that beats it especially for the cost

Same here, and I have a project 100% hosted in Cloudflare ecosystem (AI Gateway, Workers, Queues, R2, D1, DO, etc) I was so happy when I saw clef. Tested, and it's bad. Slower and wronger than Jev.

Still waiting on Decisions API using Luna, as this what I was using before.

Run your own evals, folks.

We evaluated the Decisions API at work and it was worse in all dimensions (price, speed, and quality of the result).

Or they use it as marketing wording to get liked by devs.., which is most likely given we know that’s typesafes usp

we've been public about this before the category was even created

I like to test the decision models with coin flips. This has a well known answer and I haven't seen one give it back. It undermines my trust in every other decision.

How do you test that? Are you looking for 50% confidence in an answer? I imagine it’s more than that, I’m curious.

A very simple setup, you can try here for Jev: https://jevplayground.com/

It's essentially, "predict the outcome of a coin flip", "what is the result?", and "heads" and "tails" as the answer options. With the Jev website, they do not show the "this is not a real probability" disclaimer until after you get the answer (misleading imo, should be upfront and more prominent, eg. by the input form like the gemini app)

When I give this to an LLM, it will generally respond with something like 50/50, 100% confidence.

It's a fun question but the models aren't designed for questions like that.

iirc, Jev has a "jaggedness" page that lists known problem areas with the model, and plain mathematics (which I'd argue this is) is one of them.

This is common knowledge though, and not some deep math problem, small children understand this concept.

If they cannot get something so simple right, while at the same time offered as a decision with likelihoods machine, why should I trust their other answers with "probability" they produce?

I think a better way to view these numbers is that they are "confidence" levels, not probabilities in the sense you are interpreting.

You should never trust the machine. ;) But you can have a level of confidence with it, and you can know where it's strong and weak.

This is a language association machine, not a math and probability machine. It is specifically weak with any real math.

As simple as a coin flip is -- as simple as a child may be able to understand it -- the machine is specifically not designed to produce an accurate mathematical probability.

That's just not how LLMs work. It's not how they're designed. Nor these decision models which use identical principals.

All that aside... agreed regardless. Irreversible choices should never be finalized by such a volatile technology, regardless of how well we benchmark it. If there's even a chance it could be wrong, that is unacceptable.

> not probabilities in the sense you are interpreting

I am not interpreting them that way, that's how they are being sold by the companies behind them. I do not think "confidence" is accurate either. As far as we can tell, and in the open source/model versions, it's derived logit values.

> You should never trust the machine

Rule #2 https://susam.net/inverse-laws-of-robotics.html

> coin flip...

LLM's get this one exactly right in my limited testing "50/50, 100% confident"

That’s true, for example here, this benchmark chooses tasks and datasets that Jeffy is really good at, better than Jev or Laya: https://jeffyclassify.com/blog/when-llms-fail-at-classificat...

But the Jeffy models are explicitly trained on those datasets (to be fair, classifiers are just really good, even a decision model finetuned on the same dataset, doesn’t beat a classifier - in classification tasks)

I generally trust benchmarks but my personal experience in the last 6-9 months is they have little relation to real LLM performance, unfortunately.

I've had better luck just running a few of my own simple tests.

It's almost turning into something as good as standardized testing for humans...helpful indicator, but can be gamed.

I think the point is we need better and open benchmark players which can bring and monetise on domain specific benchmarks for end users and not just selling them to the big labs.

They also link to another post they made about AI benchmarks that I really like: https://typesafe.ai/blog/antibenchmaxxing