I find it interesting that as the models get better, the "it takes no skill crowd" grows larger, but there is no denying, in hindsight, that the "it still takes skill" group was right at each model iteration.

Opus 5.5 is amazing. It still makes mistakes all the time. If you think it doesn't, you are outing yourself.

as a power user of these products, this is what i don't understand. i work on a large web app with many many users, and it's hard for me to believe that the "it takes no skill" crowd are working on anything with any stakes. and again, i use these models all the time, and they've pretty radically changed how i do work day-to-day! but my job does not consist of telling opus to complete my sprint and then sitting around for 14 days, and i don't think it will at any point in the near future. as the models get better, expectations surrounding output increase -- i'm not saying i like this, but i'm also not fretting about job security.

that said, i'd probably view all of this through a much more helpless lens if i were earlier in my career.

Agreed. It takes skill.

3 engineers at my job were fired because all they were doing was pasting Jira ticket links on Claude without further curation and generating horrible PRs.

So I don't buy that it takes no skill.

it seems like an eye of the beholder issue, for some writing is natural and others are dreadful at it but the former don't feel any particular burden in performing, it feels seamless to them. so maybe the disconnect is when framed as a skill, it sounds like something one must practice but I don't think that's the case: if you can decompose problems, think of 2nd order effects, debug, etc you are there. in other words, I would never think to use AI as you have described above (copy & paste).

Because Clef is a larger model, larger context, can process images. Plus its open weights, so you can run it yourself.

Since it benches better than Jev, Jev is probably smaller and easier to host. They could also be losing a lot of money.

Considering Jev is named after Jevon's Paradox (making things cheaper increases spending) I'm willing to bet that it's the former. Smaller model, possibly an inference pipeline optimized for a Jev-like workload (large shared context that is used for multiple questions in parallel)

And while Jev has a lot of latency, that latency stays very flat with larger inputs. So it's probably related to their inference pipeline, not model size

It's anecdata, but I ported a Jev-based prototype i've been working on to Clef, and it ended up being significantly slower and the results were worse.

This is the answer and we should not push on it for our own protection. You click a link that takes you to a poorly secured website that leaks sensitive data, without intent protections, you could be accused of crimes.

Luna is really good model that would have been super SOTA at the beginning of the year and its basically free at these prices.

This is a step beyond baseless predictions. Tao also had a "weird" "hypothetical" comment about LLMs solving complex proofs with impossible to human verify Lean.

There are theorems like that now, like de Grey's lower bound for the Hadwiger-Nelson (unit distance graph) problem. He used a SAT solver to check that a certain graph with 1581(?) vertices is not 4-colorable. There's no way for a human to check that.

Even simpler, imagine Anthropic announces Goldbach's conjecture is false and they have a billion digit counterexample. Anyone can download it (300MB compressed), but how do you check it?

Doron Zeilberger for decades has expected incomprehensible computer proofs to eventually take over mathematics.