Yep. And I'd encourage folks who don't understand this to make a personal one-time exception to any "don't ever use AI for anything ever" policies they may have. Because if you actually push on this stuff, it's quite illuminating. (Quick thread.) - ThreadSky

angus.bsky.social • 15 days ago

Yep. And I'd encourage folks who don't understand this to make a personal one-time exception to any "don't ever use AI for anything ever" policies they may have. Because if you actually push on this stuff, it's quite illuminating. (Quick thread.)

Reposted from Peter

It’s genuinely embarrassing that people keep falling for these “AI explains the hidden secrets of its programming” posts. Grok is not sentient, it cannot explain what it’s doing and why.

Comments

drgdave.bsky.social•15 days ago

I'm definitely someone who tries to avoid AI, though I can't escape it completely in my line of work.

This far, I haven't seen anything to change my opinion that it's just glorified auto-complete, and wrong just about as often as auto-complete is.

ens0.me•15 days ago

There's something to be said about how most people will use free AI tools, but you were probably using 4o-mini, which is designed to be very cheap to run and kinda sucks.

angus.bsky.social•15 days ago

I was using 4-turbo, apparently.

angus.bsky.social•15 days ago

This morning, after several people in my mentions insisted that ChatGPT was a useful research tool, I decided to see for myself. I asked it a simple but obscure historical question—who was the first president of a particular activist organization.

angus.bsky.social•15 days ago

It replied with the name of a famous activist who worked in the same field that the organization in question specialized in, but who had no connection to that organization and was in his early teens when the organization was founded.

angus.bsky.social•15 days ago

When I asked it when that person had been born, it gave me an answer that was too early by twenty years. When I said it was wrong and asked it to try again, it gave me another wrong answer. When I told it the right answer, crediting it to his NYT obit, it agreed it had been wrong before.

angus.bsky.social•15 days ago

And then I asked it how it had gotten it wrong. It gave me a non-answer, apologizing for the error. I asked again how it'd had made the mistake, and it said something about overgeneralizing. I said that didn't make any sense, and it apologized again.

angus.bsky.social•15 days ago

When I asked again how it had come up with an incorrect answer, it gave me a reply that said something about researching too hastily and not taking time to check its work. I said that sounded like an explanation of a human's failures, not a machines, and asked if it was claiming sentience.

angus.bsky.social•15 days ago

It then said that it is of course non-sentient, and agreed that its previous explanation didn't make sense. So I asked again, and it said that it had made a series of incorrect inferences from incomplete evidence, and promised to try to do better in the future. So I asked what the hell that meant.

Comments

Posting Rules

Reply