Hacker News new | ask | show | jobs
by NitpickLawyer 307 days ago
> Not really, there are plenty of things that LLMs cannot do that a professor could make his students do.

Could you offer some examples? I'm having a hard time thinking of what could be at the intersection of "hard enough for SotA LLMs" yet "easy enough for students (who are still learning, not experts in their fields, etc)".
1 comments
c0balt 306 days ago
Present the results of your exercises (in person) in front of someone. Or really anything in person.

A big downer on the online/remote Initiatives for learning but actually an advantage for older Unis that already have existing physical facilities for students.

This does however also have some problems similar to code interviews .

link
rootusrootus 306 days ago
> Present the results of your exercises (in person) in front of someone

I would not be surprised if we start to see a shift towards this. Interviews instead of written exams. It does not take long to figure out whether someone knows the material or not.

Personally, I do not understand how students expect to succeed without learning the material these days. If anything, the prevalence of AI today only makes cheating easier in the very short term -- over the next couple years I think cheating will be harder than it ever was. I tried to leverage AI to push myself through a fairly straightforward Udacity course (in generative AI, no less), and all it did was make me feel incredibly stupid. I had to stop using it and redo the parts where I had gotten some help, so that my brain would actually learn something.

But I'm Gen X, so maybe I'm too committed to old-school learning and younger people will somehow get super good at this stuff while also not having to do the hard parts.

link
NitpickLawyer 306 days ago
Sure but that's a solution to prevent students from using LLMs, not an example of something a professor can ask students that "LLMs can't do"...
link
c0balt 306 days ago
The main challenge is that most (all?) types of submissions can be created with LLMs and multi-model solutions.

Written tasks are obvious, writing a paper, essay or answering questions is part of most LLMs advertised use-cases. The only other thing was recorded videos, effectively recorded presentations, thanks to video/audio/image generation that probably can be forged too.

So the simple solution to choose something that an "LLM can't do" is to choose something were an LLM can't be applied. So we move away from a digital solution to meatspace.

Assuming that the goal is to test your knowledge/understanding of a topic, it's the same with any other assistive technology. For example, if an examiner doesn't want you[1] to use a calculator to solve a certain equation, they could try to create an artificially hard problem or just exclude the calculator from the allowed tools. The first is vulnerable to more advanced technology (more compute etc.) the latter just takes the calculator out of the equation (pun intended).

[1]: Because it would relieve you of understanding how to evaluate the equation.

link