This might be downstream of a deliberate decision by designers.

An LLM has been trained on data through February 2025.

A user asks it a question in June 2025 about 'what happened in May?'

How should the LLM respond?

(I wish to register that I didn't miss this scenario, and intend to get around to playing it this weekend...it's just that you made the awful indie-game blunder of releasing your game the day after Silksong came out, and my free time has been somewhat spoken for for the past few days).

In some other world somewhere, the foremost Confucian scholars are debating how to endow their AI with filial piety.

You can be a moderate by believing only moderate things.  Or you can be a moderate by adopting moderate strategies.  These are not necessarily the same thing. 

This piece seems to be mostly advocating for the benefits of moderate strategies.

Your reply seems to mostly be criticizing moderate beliefs.

(My political beliefs are a ridiculous assortment of things, many of them outside the Overton window.  If someone tells me their political beliefs are all moderate, I suspect them of being a sheep.

But my political strategies are moderate: I have voted for various parties' candidates at various times, depending on who seems worse lately.  This seems...strategically correct to me?)

If you ever do it, please be sure to try to confuse archaeologists as much as possible.  Find some cave, leave all your flint tools there, and carve images of space aliens onto the wall.

This might be a cultural/region-based thing.  Stop by a bar in Alabama, or even just somewhere rural, and I think there might be more use of bars as matchmaking.

Here is a list of numbers.  Which two of these numbers are closest together?

815

187

733

812

142

312

I think the obvious approach is comparably neat until you get to the point of proving 

at which point it's a mess.  The Google approach manages to prove that part in a much nicer way as a side effect of its general result.

I looked at the Q1/4/5 answers^[1].  I think they would indeed most likely all get 7s: there's quite a bit of verbosity, and in particular OpenAI's Q4 answer spends a lot of time talking its way around in circles, but I believe there's a valid proof in all of them.

Most interesting is Q1, where OpenAI produces what I think is a very human answer (the same approach I took, and the one I'd expect most human solvers to take) while Google takes a less intuitive approach but one that ends up much neater.  This makes me a little bit suspicious about whether some functionally-identical problem showed up somewhere in Google's training, but if it didn't that is extra impressive.

1. ^{^}

   IMO Q3 and Q6 are generally much harder: the AI didn't solve Q6, and I haven't gone through the Q3 answers.  Q2 was a geometry one, which is weirder to look through and which I find very unpleasant.

(Credentials: was an IMO team reserve, did some similar competitions)

Have the actual answers AI produced been posted?   Because I could see this mattering a lot, or not at all, depending on the exact quality of the answers.

If you give a clean, accurate answer that lines up with the expected proof, grading is quite quick and very easy.  But if your proof is messy and non-standard, coordinators need to go through it and determine its validity: or if you missed out part of the proof, there needs to be a standardized answer to 'how big a gap is this, and how much partial credit do you get?'

(Also, have the exact prompts used been posted?  Because it would be very very easy to add small amounts of text or examples that make these problems much easier.  If the prompt used for Q4 contains the number '6' at any point in it, for example, I would basically just instantly call that 'cheating').