[MUSIC]
>> LARS KOTTHOFF: I'm Lars Kotthoff. I'm a professor at the University of St. Andrews in Computer Science and I hold the Johan and Gaynor Rupert Chair in Artificial Intelligence there.
"Is there any point in studying computer science if AI can do it all?"
Well, if you look at the popular press, then you might get the impression that AI can do it all, but if you take a closer look, that is not actually the case. So, that would be the first part of the answer to that.
Let me give you a concrete example here. I was working with a friend, so we're both researchers in AI and we've been doing this for a number of years, and we did do the vibe-coded LLM thing where we just had, I can't remember which one of the particular LLMs it was, write some code for a particular AI system that we were to put into practice.
And we both had a look at it and we both find, yes, that looks good and we are going to use this. And then a couple of weeks later, it turned out that actually this code was broken. And it wasn't even broken in very subtle ways, it was broken in very obvious ways which once we had a proper look at it without relying on the AI to do it, we realized pretty much immediately and it was very easy to fix for humans who know what they're doing, but the takeaway for us from this was that we really cannot trust AI to produce even relatively simple pieces of code.
And that's really the reason why now is a great time to study computer science because what's happening right now when you listen to companies like Microsoft or Amazon who are saying that X% of the code that they write is generated by AI is that they're amassing huge technical debt.
And this is really just a fancy way of saying that if you build a system that you don't understand and you intend to use this system for a prolonged period of time, you will have to understand it at some point.
All this code that is currently being written by some kind of AI system and put into production, somebody, some human will have to understand it at some point.
Now, what will that human need to be able to do? They will need to be able to, first of all, program themselves. They will also need to be able to understand the larger implications of smaller changes with respect to an overall system.
Because when we talk about implementing software, it's not just programming. It's not just sitting down and writing your Python code or whatever your language of choice is. It's about architecting an entire system. It's about understanding how different components of that system will work together.
It's about understanding what the implications are of implementing something in a certain way, and these implications might not manifest themselves in that part, but they might manifest themselves in a completely different part. And if you have never done that before, then it will be really, really hard to do it.
So, what's happening right now, where people are deciding that, "Oh, actually, I shouldn't study computer science anymore. Those jobs are going away, right?" Well, it might look like that at the moment, but in a couple of years, I am sure that companies will realize that, "Oh, hang on, we have all this vibe-coded code that nobody really understands, and now we need to understand it."
So, this is when it will become really, really important to have a background in computer science, to have studied computer science, to know how to program, to know how to architect systems, to be able to understand code that somebody or something else has written, and to understand how a particular system is architected, how to make changes to it so that you can change it in a meaningful way to do something new or to do something differently.
And for that reason, I think that now is actually a fantastic time to study computer science, because in just a couple of years, and let's keep this secret, so just you and me, we will know how to program, we will make millions. We'll be the only ones who are left by then who know how to program and we can retire after that because our skills will be in high demand.
"AI has developed at an unprecedented rate. Where do you think it will be in 5 to 10 years time?"
Well, predictions are notoriously hard, especially of the future. I would like to see AI go in a direction where we have a better understanding of how these large systems work because here's one weird thing about large language models that you might not be aware of. Nobody really understands why they work as well as they do. Not even the companies.
Of course, they might claim that they understand why it works and oh yeah, we can make this change here and then it will stop confabulating and hallucinating answers. But the truth is that really nobody knows why they work as well as they do.
And this is something where I would like to see some progress. That first of all, we get a better understanding of why these systems work as well as they do. And then we will hopefully be able to harness that understanding to build better systems. And I mean better in many different things here. So ideally, we would want to get models that don't make up things.
We would want to get models that give us reliable answers to important questions that we have where we are relying on getting something that is not completely made up or at least some kind of fact-checking has taken place. Um but we would ideally also be able to run these systems in a much, much less resource-intensive manners.
Because at the moment, we're building massive AI data centers to run things like ChatGPT and Claude and all these large language models that exist because they're consuming a large amount of resources. Anecdotally, we know that this is not really necessary because humans do it all the time with much fewer resources.
I know I'm not using nearly as much resources as an LLM. Am I giving you better answers than an LLM? Well, that's maybe for you to judge, but I'm certainly doing this in a much, much more efficient manner.
So, I hope that in 5 to 10 years, we will be in a place where we can use AI much, much more efficiently, but also in a way where we can actually use it in practice without having to double-check everything, but being able to rely on the answers a little bit more.
[MUSIC]
