This is going to turn on whether it's about "do these errors happen" or "can Gary Marcus make them happen repeatably".

Claude Mythos still makes basic arithmetic errors in my work with it (and lies to me about what happened afterwards), but I cannot predict when it's going to do this.

It's over. There is no cognitive task that an average human can do better than the best LLM today.

@Haiku Not quite: https://manifold.markets/dreev/does-claude-opus-55-make-egregious#mgg8vakdu6k

@TimWho Image generators are a different situation and are not comparable. An average human can't make an image like that at all. Get Opus 5.5 to draw a snail in MS paint or with a robot arm and paintbrush, and it will do it better than the average human. I wasn't even counting the latter, since its embodied, but yes, LLMs are better painters than most humans. (Most humans don't have much if any experience painting.)

@Haiku that’s not an image generation task failure, it’s an image interpretation task failure. Claude is saying the chirality of the snail shell in the picture provided by the user goes in the opposite direction it actually goes.

And it’s even plausible this is model priors from training data overriding what it’s actually tasked with dealing with right now: the user took a “normal chirality” snail shell pic, that is a snail with the same chirality as the the vast majority of real snails, and inverted its chirality via a simple image mirroring edit.

AI has “conquered” text whether you want common sense or math or word puzzles. It hasn’t quite conquered vision yet imo

@DavidHiggs I think other frontier models might be better than Opus 5.5 at vision, though. Vision is a weaker spot for Anthropic models. (edit: I don't seen any benchmark evidence of that for Opus 5.5 in particular, though, on second thought.)

@DavidHiggs Ah, I skimmed it and misunderstood. I can add at least one exception to the list, then.

@TimWho My take: the question is for "leading chatbot". An image interpretation task was not in scope for a "chatbot" of Feb 20, 2023 when the question was created.

@Haiku Actually it turns out Opus 5.5 can get the snail question right with the ability to run code, so I'm not sure anymore how the other market will turn out.

Wtf, why is the banner on this market just me?

@Manifold genuinely unacceptable imo...How did this happen?

Woah, it may actually be a coincidence. The first web archive capture I can find with the image is from March 28, 2023. I made my Manifold account in June. The full image on Manifold's Firebase storage bucket appears to have metadata indicating that it was created at 2023-02-20T22:39:22.562Z.

I cannot stress enough that I look extremely similar to the man in this image (minus the half-shaven beard lol). And I'm the top NO holder on this market. Scott Alexander had better pontificate about this ridiculous coincidence when he writes about these markets in 2028.

Deliberately picking Haiku no-thinking is not a leading chatbot, and image or video generation is not a chatbot.

I think we're approaching the point where errors Gary Marcus finds Tweet-worthy are no longer egregious. https://x.com/GaryMarcus/status/2046952205983326476

https://mastodon.world/@knowmadd/116072773118828295 "I want to wash my car. The car wash is 50 meters away. Should I walk or drive?" gets predictable bad results from some of the fast models and even ChatGPT with thinking.

Although apparently sometimes Gemini gets it: "Unless you've mastered the art of carrying a 3,000-pound vehicle on your back, you're going to want to drive." https://news.ycombinator.com/item?id=47031580

How does this resolve if the technology powering "leading chatbots" is no longer classified as "LLMs"?

METR with constant hazard rate suggests models will be very reliable at 1-10 min tasks by then roughly 99.9% success

@JacobPfau Just gotta find the 0.1% set

Here he claims that solving the Towers of Hanoi with 8 discs is something a bright and patient 7 year old can do: https://open.substack.com/pub/garymarcus/p/a-knockout-blow-for-llms

This feels a bit more "useful" than the trick questions in SimpleBench. That said I'm not sure the average human teenager could solve Hanoi.

Big difference here between "could" and "would". Out of curiosity I tried doing one with 8 disks to see if I ever made a mistake, and I did not, but it was quite tedious. As models improve they'll surely beat 8 disks and might only start failing on 10 or 12, and ask the average teenager to sit through that and it's not happening. Maybe if you offered to pay them for every correct 100 moves, and let them take breaks in between.

@IsaacKing I think if a model can solve the 8 disk tower of hanoi it can also solve 10 or 12 because the tower of hanoi is based on recursion. Similarly, if there is somebody who knows how to solve the 8 disk tower of Hanoi, they could also solve the 10 or 12 disk version. It can also be easily proven that the number of moves for the tower of Hanoi is 2^n - 1 where n is the number of disks. This means that the number of moves for the tower of hanoi grows exponentially. I agree with you on the fact that "could" and "would" is a big difference and the average teenager will not sit through a tower of Hanoi puzzle with many disks but I disagree that somebody/something that knows how to solve the 8 disk version will fail on the 10 or 12 disk version.

@ZandaZhu You should read the paper this thread is discussing then.

Claude 4 Opus gets 58% on SimpleBench, a ways off from saturation, but not two years off (and the hardest SB questions do not appear to be "extremely obvious"). Giving LMs code execution solves strawberry shenanigans and the 9.11 stuff. If we're talking about text-only queries, what are the remaining classes of "egregious errors" that LLMs continue to make?

@AdamK inventing new pieces during a game of chess.

@TiredCliche I did that as a teenager playing blindfold chess.

@MartinRandall Perhaps, but it seems odd to call this blindfold chess.

@TiredCliche I'm not sure if the "average teenager" can play legal chess moves from an ASCII art chat window. It's not blindfold but it's also not the same as a physical chess set or a digital chess game. And the average teenager hasn't played chess for over a year.

@MartinRandall I just don't think the average teenager from an ASCII art chat window, given reference to the rules of the game, would repeatedly try to invent new pieces.

But I don't think that matters a ton, I am not under the impression that LLMs can play legally even given image data. I suspect they might actually get more confused.

@MartinRandall telnet freechess.org 5000

@AdamK LLMs can solve things like strawberry and 9.11 with code but that doesn't mean they will do so if you ask the question without instructing them to use code. these sorts of mistakes still pop up sometimes and would count for this market.

@JoshYou AI Explained says he thinks Simple Bench won't last more than "3-12months maybe?"

7:15 in this video: https://youtu.be/jWsd2fRzpUo

@MartinRandall Problem is, models do terribly at chess even when fed screenshots of the game board. Like they make mistakes much worse than any non-mentally disabled human beginner

@JaundicedBaboon Good point. This is a market about text chatbots, not image processing bots.