I prefer using a book on the subject with chatgpt and forums/chatrooms as supplements.

I find that prompting chatgpt with book excerpts keeps it grounded in reality.

Books are obviously better if you're trying to learn about a subject.

If you just have one very specific question, it's often easier to ask ChatGPT and then fact-check its output.

Yes, sometimes, depending on the subject. Other times the fact checking needed is so rigorous that you might as well have skipped the AI.

> but human teachers are not always right, either.

It also seems to me that it's gotten a lot better about hallucinating. When I first started using it it would often be wildly wrong, but not it seems like it hallucinates about as much as the average person I talk to.

But with a teacher, I can get a feel of their qualifications after a while (or cross-reference whatever they tell me) and if they're wrong, I can go somewhere else. Also, a teacher is less likely to gaslight me and hopefully will research the topic in case they're not sure about something. With AI, it's always a coin toss.

I don't know, maybe I'm just a old person yelling at clouds, but I just find it crazy to use AI for learning anything. Maybe one day.

With an AI you also get a feel for what they typically know and what they'll probably fail at. The limitations are mostly consistent, at least on a per-model basis. Even more so than humans, since they never grow or learn anything new.

I feel like I've learned a ton of new things from GPT 4 in terms of software development recently, while also feeling I've been stagnating in my job with nobody to learn from. It's just so easy to have it explore alternative approaches or have it structure things in ways you may not be comfortable with yet (and would thus otherwise avoid) but would be better in the long run. It has the breadth to show you all that's possible while reducing the entry effort for learning to near zero.

It's just another useful tool, just like books, the internet, and real teachers are useful tools. They all have their pros and cons. The advantage of chat bots is that they're always there and scale to an arbitrarily large number of students. The advantage of teachers is that they're, well, people.

I'm not sure they scale so well. They probably need so many GPUs that teachers are more efficient.

That seems very unlikely to me. Certainly LLMs are very compute intensive relative to past applications, but I would find it shocking if they are not nonetheless far more scalable than human labor.

It's easier to make a teacher than to make a GPU though.

Again, no it isn't, what are you talking about?

Human beings are incredibly incredibly expensive to "make" into any useful adult thing. I have spent more money just this week just on a single one of the humans who are dependent on me than it would cost me to buy a fully packaged GPU in a box at best buy (and I think that box would actually have multiple GPUs packaged together in it). And I'll do that week after week for many more years. And then it will still be like four more years after that before they're capable of being a teacher.

It was harder to make the first GPU than to make one marginal unit of human teacher, but it's vastly easier to make a marginal unit of GPU.

I'm honestly curious what your mental model is, where you seem to think human labor is cheaper than computer hardware...

With GPT it's more like rolling a d20 than a coin toss.

19-20=You receive the enlightenment you were asking for, more or less instantly.

3-18=You get what amounts to a Socratic conversation with Wikipedia. Which, unless you're fortunate enough to have access to expert personal mentorship, is better than what you had before.

1-2=You get a load of highly-plausible bullshit. You end up worse off intellectually, possibly much worse.

All in all, I'm OK with these odds, especially at this early stage of the LLM game.

The Khan Academy has made some good strides in this area with AI.

And just as I switched to a teaching career. FML... :-D

A teacher just teaches what he or she has been taught. And history is written by the winners. None of us can really know what is true.

Let me introduce you to the concept of theorems.

I’ll second this. It really shines when learning new technologies in a field in which you already have some expertise. Being able to apply the smell test, or quickly validate the output is important.

I haven’t tested this, but I would imagine the type of information you’re requesting dictates the quality of response you’ll get. In my case I use it mainly as interactive documentation for different tooling. Asking it to explain the process for synthesizing new hypothetical polymers likely will not yield useful results.

It’s worth noting that my experiences are with GPT 4.

I've been using it a lot while working through two very math-y books right now - Princeton's Companion to Mathematics and Data-Driven Science and Engineering. To your point, I have some background in undergraduate level mostly applied math, but there are a lot of things where I don't understand the argument or immediately have an intuition for what's going on, and I've found gpt-4 to be actually really helpful with that. Sometimes it is wrong, but it can't totally BS me, because I'm sitting there looking at the correct equations and proofs, so if it's wrong I always know what's right. But a lot of times it really helps me with intuitions on things that if I was purely just self-learning, outside a class with a professor and TA, I'd probably just accept the poorer level of understanding and move on, instead of investing the time in trying to find what I'm looking for on the internet.

> I think in the future, the main value of going to college will be socializing, networking, and learning "life" skills through interactions with others, dorm mates, etc. The teaching part will be already taken care of by AI.

Based on a recent experience, I agree.

I'm not a professional software engineer or architect, but I have a web app used by a few thousand people a month and recently decided to rewrite it using modern web technologies — Vue 3, Vite, Pinia, TypeScript, and serverless functions. I was familiar with these, but hadn't used them for a real project. During the rewrite, I leaned hard on GPT-4/Phind/Copilot as teachers/pair programmers. It's hard to overstate how helpful these tools were.

For the questions I had, search engines were hit-or-miss. Within a few days it became clear that the pre-2023 "search engine" is dead, and that post-2023 search engines will look like these new LLM-powered tools. The future of topic-specific search engines is well represented by Phind, which leverages GPT-4 but is also trained with current technical documentation and includes references in answers.

This is really something that has to be experienced to be understood. I have a friend who's a professional software engineer, and he had to watch me for a couple hours before it really clicked. I recommend trying this if you haven't already.

I think something that would be awesome would be human teachers scaling themselves via AIs. Like students having an escalation path to say "wait, I'm confused and the AI isn't helping!" and have the teacher step in to try to help. Or even better, step in to help proactively after the AI says "hey, this person is asking questions or answering things in a way that makes me think they are confused, maybe take a look?".

If someone is already building something like this, I'd love to hear about it :)