I've found it excels at some things:
1) The broad overview of a topic
2) When I have a vague idea, it helps me narrow down the correct terminology for it
3) Providing examples of a particular category ("are there any examples of where v1 in the visual cortex develops in a disordered way?")
4) "Tell me the canonical textbooks in field X"
5) Posing math exercises
6) Free form branching--while talking about one topic, I want to shift to another that is distinct but related.
I agree they leave a lot to be desired when digging very deeply into a topic. And my biggest pet peeve is when they hallucinate fake references ("tell me papers that investigate this topic" will, for any sufficiently obscure topic, result in a bunch of very promising paper titles that are wholely invented).
These things are moving so quickly, but I teach a 2nd year combinatorics course, and about 3 months ago I tried th latest chatGPT and Deepseek -- they could answer very standard questions, but were wrong for more advanced questions, but often in quite subtle ways. I actually set a piece of homework "marking" chatGPT, which went well and students seemed to enjoy!
Super good idea!!
Luc Julia (one of the main Siri's creators) describe a very similar exercice in this interview [0](It's in french, although the au translation isn't too bad)
The gist of it, is that he describes this exercice he does with his students, where they ask chatgpt about Victor Hugo's biography, and then proceed to spot the errors made by Chatgtp.
This setup is simple, but there are very interesting mechanisms in place. The student get to learn about challenging facts, do fact checking, cross reference, etc. While also asserting the reference figure of the teacher, with the knowledge to take down chat gpt.
Well done :)
Edit: adding link
[0] https://youtube.com/shorts/SlyUvvbzRPc?si=2Fv-KIgls-uxr_3z
this is amazing strategy
forgot the link :)
Arf seems I'm one of those :).. thanks for the heads up!
That’s a great idea to both teach the subject and AI skepticism.
> I actually set a piece of homework "marking" chatGPT, which went well and students seemed to enjoy!
This. This should be done everywhere. It is the best way to let students see first hand that LLM output is useful, but can be (and often is) wrong.
If people really understands that, everything will be better.
Very clever and approachable, and I've been unintentionally giving myself that exercise for awhile now. Who knows how long it will remain viable, though.
When you say the latest chatGPT, do you mean o3?
Whatever was best on a paid account 3 months ago. I was quite disappointed to be honest, based on what I had been hearing.
I think by default ChatGPT will choose 4o for you. So unless you actually chose o3 you haven’t used the best model.
that's a cool assignment!
>When I have a vague idea, it helps me narrow down the correct terminology for it
so the opposite of Stack Overflow really, where if you have a vague idea your question gets deleted and you get reprimanded.
Maybe Stack Overflow could use AI for this, help you formulate a question in the way they want.
Maybe. But, it's been over a year since I used StackOverflow, primarily because of LLMs. Sure, I could use an LLM to formulate a question that passes SO's muster. But why bother, when the LLM can almost certainly answer the question as well; SO will be slower; and there's a decent chance that my question will be marked as a duplicate (because it pattern matches to a similar but distinct question).
>hen the LLM can almost certainly answer the question as well;
You say this in a thread specifically talking about how LLM's fall apart when digging deeper into the surface of questions.
Do people really want to learn and understand, or just feel like they are learning and understanding?
I would say that LLM might give a correct answer, in a good enough question there is more than one answer.
Furthermore the LLM might give an answer but probably not explain with the best skills available why the answer is the way it is. This is of course something that varies with StackOverflow but there it is at least possible that somebody with deep technical knowledge decides a question is worth answering deeply.
outside of 5), I concur. It's good for discovery, as is Google for discovering topics while weighing on proper profesionally resources and articles for the learning.
It's too bad people are trying to substitute the latter with the chatGPT output itself. And I absolutely cannot trust any machine that is willing to lie to me rather than admit ignorance on a subject.
I find 2 invaluable for enhancing search, and combined with 1 & 4, it's a huge boost to self-learning.
I’ve found the AI is particularly good at explaining AI, better than quite a lot of other coding tasks.