So I was reading a post that was supposed to be about testing different LLMs at chess…and the author keeps saying things like “I asked it for the next move, and if the first 20 responses were all illegal, I chose a legal move at random.”

My dude (gender-neutral), this means the model cannot play chess.

Just imagine applying this logic to any other kind of tech. “If I run the vacuum cleaner over the same cat hair 20 times and it still doesn’t get sucked up, I pick up the cat hair by hand and keep going. And look, I end up with a clean carpet! This proves how well the vacuum works!”

I mentioned all this on Mastodon/Bluesky, and added that what I really wanted to see was a breakdown of the kind of illegal moves LLMs try to make. Someone replied with a rec for GothamChess on Youtube. (I’ve watched a bunch of his LLM game videos now, they’re exactly what I was looking for, more on those later.)

The thing is, though: I was out at the time, I couldn’t stop to watch videos, so I just googled the guy on my phone. When I leave a tab open, it’s a reminder to check this out once I get home.

…And one of the top search results was a Reddit post with the summary, quote, “American Internatiol Master Levy Rozman, AKA “GothamChess” has just been charged with one count of first-degree murder.”

I was, uh, pretty alarmed by this. I clicked through, hoping to find out more about what happened.

…The actual Reddit conversation is all about what makes GothamChess’s Youtube channel engaging. No charges. Not even allegations. No mention of murder at all!

So…what gives? Is Reddit putting AI-hallucinated summaries in the metadata of its own posts, or is Google using AI-hallucinated summaries to replace what the site gives it? Which executive signed off on this?

Edit: The line is apparently the title of a completely different (and joking!) Reddit post. Thanks to Gwen for spotting it! It’s not linked in the post above, or in any of the comments — apparently Reddit was showing it as a “Related Post” for Gwen, and for me, it isn’t even doing that.

So it’s not completely hallucinated text…it’s just pulled from a completely inappropriate part of the page. And then put at almost the top of the Google search results. Without linking back to the context that would show it’s a joke. Either it’s a normal algorithm, but for some reason it was programmed to pull summary text from random parts of the page…or it’s still an LLM, having the “you should eat several small rocks per day” problem.

Wish I knew which it was. And I’m still curious which of the companies is falling on their face, here.