"The goal isn’t to rank models, but to understand how they fail."

The goal isn't to write an informative blog post describing what you learned, but to generate slop and expect other folks to read it.

I really wish people would stop doing this. I love reading about your side projects and all of the cool things you're doing. But, it just feels insulting to open up something that's so obviously completely AI generated. If you aren't willing to write it in your own voice, why would it be worth reading?

You know the meme where a concise sentence is translated by an LLM into a loquacious formal email which is then again summarized to a concise statement by another LLM on the receiving end?

I believe that's what we need to do here. People have some interesting information to share, but they don't care about penmanship and that's not just being lazy. It takes a lot of time to produce a nice post. I cannot guarantee the author used an LLM but there sure is a suspicious amount of em-dashes.

Anyway, there are still some interesting data points so I'd recommend to run the website through an LLM to get a nice summary if the prominent TL;DR is too short for you. Times are a-changing.

I agree somewhat. My issue is primarily that, without the author actually penning the post themselves, we have little to no evidence that they've actually done anything. Maybe the data is all AI generated or hallucinated, maybe the validations weren't thorough. I could determine all of these things myself, via rigorous review of the blog post. But at that point, I'm just doing the research myself, of what use is the post?

For work communications, I agree with you. There's an inherent accountability there. If you send me AI slop, and something goes terribly wrong, you'll be held accountable for the slop. Here, the slop is just noise that prevents us from finding the truly interesting posts.

It's a very interesting issue you raise here. Notice that even if he typed it all out himself we wouldn't be any wiser. Literally nothing would have changed. Using the "I can verify that he performed a lot of work" as a quality signal always was a questionable - albeit understandable - choice but in the LLM-age it's useless.

<super_weird_rant>

I don't think I like it, but I think we are heading towards a situation where all information is filtered, reviewed and validated before it even becomes available to you. We need to do a lot of work to define what "reviewed" and "validated" mean here, but I don't see many ways around it. This would, however, require a vast attitude shift whereby we have some way of proclaiming "facts" and "arguments" tied to "proofs" that can be automated in some fashion, not just for code, but for all communication in general.

Stuff like "X is true in 50% of cases" need to be automatically and transparently tied to some part of a system that supports your claim which itself can be tied to some greater system, etc. If we have UIs that support this cleanly we can inspect the veracity of claims ourselves as so far the validation is feasible/practical/economical. Perhaps some sort of "this claim is true under X,Y conditions"-fingerprint made by some trusted VerificationAgency, a chain of trust so to speak, like our certificate systems. Or perhaps a P2P network of open-source "ClaimVerifiers". If everything is by default written with verification in mind, not just code, but literally everything that needs to be correct, I think that would be quite interesting.

OK, this is super weird so I'll let myself out now.

</super_weird_rant>