One other way if you're using a local model is you can train a custom VAE decoder to output them at a lower resolution and get an exact pixel grid, but that's not doable with remote models like ChatGPT, MJ, etc.

Pixel art styles are the easiest to tell. They can't maintain a consistent pixel grid, so you just have to zoom in on one pixel, measure its size and find one more that doesn't match.

I saw someone attempt to fix the grid recently www.reddit.com/r/MachineLea...

I'm getting some deserved push back on focusing on the enumeration property, since the old NALU paper alluded to it in a footnote on one page. Perhaps i should have focused on the catastrophic error reduction or “universal” training distribution more.

Despite not computing math, they are excellent at actually selecting it (again obvious that the error should be so low with hindsight) idk, this is exhausting and I need to get back to billing hours. I just hope it's useful for someone.

Its more useful in that it answers whether NALU style units do math. They just select from discrete options, not compute anything. Its all very obvious with hindsight bias. But the paper was from 2018 and there were follow ups trying to make it work. So now it works, but the emperor has no clothes?

Yeah, you found it. I had to host it on Github because I can't get a DOI so far. I'm just a dev out here trying to share what feels like an important finding for neural arithmetic, for free, on my own time.

Update: TechRxiv finally responded minutes after I posted this by... sending a rejection for ORCID privacy settings, and lacking an affiliation. Neither of these are listed on the submissions guidelines page they direct me to in their response.

@iamtrask.bsky.social I hope your shoulders are feeling alright, I'm not as light as I used to be.

I'm still waiting for #TechRxiv review - #IEEE support told me somebody would review it by yesterday.

Starting to think I should release this work without a DOI - the code, website, and paper are all beginning to gather dust.