[davydm]

Pure delusion.

There are no "digital gods" only the super-powered autocorrect people call "ai". They can't make new stuff. They can't solve novel problems no human has solved before, though they _can_, with the correct setup, brute-force solutions to understandable problems by throwing everything at it until something sticks.

They don't learn. They don't teach. They are not the deities that are presented here. This article is fantasy, projected from real circumstances, by an over-active imagination.

[trehalose]

I'm curious to know what the author suggests we do. Elect Claude, ChatGPT, and Gemini as our leaders? Put these future cancer cure discoverers in a hospital and get them to work curing individual cancer patients?

[AlexCoventry]

I would suggest paying the $20/month rent, and trying to use ChatGPT o3/o4-mini-high/o1-pro as a tutor, to help you understand something you're curious about but never really had time or energy to dig into before. It's pretty glorious, and a straight-up pedagogical revolution, IMO.

[wizzwizz4]

I would suggest getting a library card, and trying to read a good book, writing down any questions you might have for later review. While it doesn't feel as though you're learning as much, or as quickly, most people will find that they actually know a lot more about the subject afterwards.

A chat log that takes me 2 hours to produce, I can read in 5 minutes. There's no world in which that's efficient pedagogy, even disregarding ChatGPT's truthfulness issues.

[AlexCoventry]

Try it with a subject where you'd have to pick up text from multiple references, like a modern deep learning paper if that's not your field. I've tried it both ways, and I know what works for me, FWIW.

The higher ChatGPT services hallucinate much less, and you can tell them to give confidence estimates on their claims. The confidence estimates are pretty reliable, in my experience.

ChatGPT is improving very swiftly. A couple of months ago I was equally dismissive.

[verbify]

How do you know it isn't hallucinating? My experience of AI is that it can be very convincing and produce information that "looks right".

[AlexCoventry]

I mostly use it in fields where its claims are easily verifiable.

[aprilthird2021]

So then it's no different from reading a Wikipedia page about a topic? I mean, if you only use it to teach you stuff you can verify, then couldn't you skip one step in this process?

[AlexCoventry]

I learned Stochastic Differential Equations this way, for instance. I wasn't going to learn that so quickly from Wikipedia.

https://chatgpt.com/share/6834ed85-4dac-800e-b940-3b9a5f13d6...

[aprilthird2021]

It's definitely more involved than the Wikipedia article. How do you confirm what it's saying is true?