The overlords of AI, realizing the embarrassment of such a simple question being answered so stupidly, went in and made specific corrections so this would not happen again.

What we actually have is this: it was wrong in February, it's right now. The rest is you filling in the gap with people in a room.

And your claim isn't even the most obvious one. The plainest explanation is that the question got famous, so the answer is now everywhere in what these things are trained on. Nobody had to do anything. It got handed the answer by the internet arguing about it.

You're also assuming a change to the model happened at all. Models get replaced constantly. The thing answering today may not be the thing that was answering in February. If a new version handles it better, that's not a correction to the old one — it's a different model.

The bigger point in the post is what gets lost here. They put the same question to 10,000 people. Almost 3 out of 10 said "walk". Nobody's going to go fix those people. That's just what happens sometimes — a quick answer shows up, sounds fine, and you never stop to ask whether it actually solves the problem.

And that's the interesting part. It's not that these things are broken in some way people aren't. They trip over the same stuff we do. The guy in the podcast has a bunch of examples — the same kinds of sentences that tangle us up tangle them up, the same parts of a list get remembered and forgotten. Nobody built that in. It came out that way.

The work going on now isn't fixing answers one at a time. It's trying to get these things to catch themselves — notice when the easy answer doesn't add up, stop and check before committing. Same thing we humans use second opinions and checklists for. Nobody's solved it yet.