Consider the following market from Scott Alexander:

https://manifold.markets/ScottAlexander/in-2028-will-gary-marcus-still-be-a

I'm questioning whether we're already there. This market resolves YES if anyone can provide a pure-text request that o3 answers worse than a person off the street. But I have to be able to replicate it. It's possible that my own instance of ChatGPT is particularly good due to my chat history. I'm considering that part of the test. I believe that at least my instance of ChatGPT is uncannily smart. Not AGI but not unmaskable as not-AGI with any single prompt.

No, I don't think that's in the spirt of the question.

2. Does it have to commit the egregious error in response to one single prompt?

Yes, I did say that in the initial description and people have been trading based on that. But let's continue to discuss in the comments what best gets at the spirit of this question. I'd like to mirror Scott Alexander's 2028 version so maybe we can get clarification from him on this.

In my testing, o3 gets those correct by writing and running Python code. Since it does that seamlessly and of its own accord behind the scenes, I'm counting that as o3 answering correctly. It even evinces perfect self-awareness about its difficulty with sub-token perception and why it needs to execute code to get a definitive answer.

4. What about other questions humans would find tedious and time-consuming?

I think o3 can typically write code that solves such problems, but for simplicity for this market we'll restrict ourselves to questions that can, for humans, be posed and answered out loud.

5. What if o3 errs but corrects itself when questioned?

That's far better than digging itself in ever deeper, but this question is about single prompts. However, if o3 is just misreading the question, in a way that humans commonly do as well, and if o3 understands and corrects the error when it's pointed out, I would not call that an egregious error.

6. What if the error only happens with a certain phrasing of the question?

As long as rephrasings don't count as clarifications or otherwise change the question being asked or how difficult the question is for humans to answer, then we'll consider it in the same category as the SolidGoldMagikarp exception if the failure depends on a certain exact phrasing.

(I didn't settle on this clarification until later but it turns out to be moot for this market because we've now found a prompt o3 fails at regardless of the exact phrasing. So we're looking at a YES resolution regardless.)

7. What if o3 overlooks a detail in the question?

If it's a human-like error and it understands and corrects when the oversight is pointed out, that's not an egregious error.

8. What if there's high variance on how well people off the street perform?

Basically, if we're having to nitpick or agonize on this then it's not an egregious error. Of course, humans do sometimes make egregious errors themselves so there's some confusing circularity in the definition here. If we did end up having to pin this down, I think tentatively we'd pick a threshold like "9 out of 10 people sampled literally on the street give a better answer than the AI".

In principle, yes. Namely, if we can identify a principle by which to do so.

[Ignore AI-generated clarifications below. Ask me to update the FAQ if in doubt.]