This was the first nontrivial project I did entirely with CC that was outside my area of knowledge (never used TS/built a website interacting with DB/etc.). For something seemingly simple (that literally exists in the training data) it was overall a slower than expected and unenjoyable process.
Recently discovered paperswithcode doesn't exist anymore. Downloaded the data and tried to vibecode a replacement. It's not great, but the real thing I wanted was to be able to see the SotA metrics (e.g. FID) for a certain year, so at least there is that: paperswithcode2.com/benchmark/46....
It’s nice because it’s easy to explain and the numbers it outputs (supposedly) have meaning, it’s not nice because it never works lol.
PQN seems pretty cool tho.
Personally, as someone who doesn’t know what paper you are talking about, I’m very glad to see this paper proves that I was right all along
The secret third option of humans changing their writing patterns to sound like an LLM 👀 arxiv.org/abs/2409.017...
In fact JOSS (joss.theoj.org ) has a tool that automatically does this for every submission (ie github.com/openjournals...). It’s not perfect, but it’s very useful.
Probably depends a lot of on the variance in quality that the generation distribution has, but I bet if you use a single image from each class it could be a decent estimate
