NPTEL-NOC IITM · Transcript · · Updated
2. Role of AI in Research - 2 – Transcript
Video by NPTEL-NOC IITM · Watch the original on YouTube
Use this video with Pro
Reading, copying and transcript downloads are free. Create saved AI results for this video with a signed-in Pro account ($9/month).
Transcript
Search ignores accents. All transcript text stays visible.
>> Hello everyone. Welcome to another episode of research methodology. Today we have our guest Professor K.C. Sivaramakrishnan from computer science department IIT Madras. Welcome K.C.
>> Yeah, thanks thanks to be here. I'm very glad to be here. >> Yeah, so this as I mentioned this is primarily for young researchers. So they are just about to do PhD. So they enroll [clears throat] for this research methodology course.
So to begin our conversation I would like to ask you just to comment for them that those who are just starting research and they want to use AI as a tool. What are your recommendations or what are your suggestions of using AI?
>> Wonderful. So first thing I want to say is AI is not a tool. You would typically think of a tool as I have some process I'm doing today. And I want to accelerate it.
Right? So that is what I would call a tool. But AI is sort of bringing up new ways of doing things. It is sort of changing how the economics of doing an action >> Okay. >> which was not possible previously are becoming possible.
>> Okay. >> Right? So if you are a if you are a researcher one thing we would do before Google it predates my time. But sometimes what I hear is oh you have to talk to people who know this area to find out what is the prior art.
Then we sort of shifted to Google and it was a sort of a big change where you can just Google whatever you have. Now with AI what you can do is you can of course use it as a search engine to identify what open research problems are there.
But the biggest thing that you can do with AI is have a conversation with it. >> Okay. >> Right? As a not just a literature review, but you can further ask questions around Look, okay, I understand this is the problem >> Mhm. >> that A or B has solved.
>> Mhm. >> Now, between A and B, >> Mhm. >> what falls in between? >> Right? Okay. >> And is that a interesting problem to solve? If I solve this problem, >> Huh. >> what will I >> So, this conversation is the one that you suggest that's beyond the tool feature or >> Yeah, so so this is one example.
>> Okay. >> But the other example is if you think about producing ideas. >> Okay. >> Right? One of the biggest challenges is coming up with good ideas. >> Right. Right. >> AI is a very creative machine. >> Mhm.
>> Right? You ask for 100 different ideas, it'll give you 100. >> Right. Right. >> But whether the 100 is good or not, >> Mhm. >> that is what the student should do. >> Okay. >> Right? And that is where young researchers have to develop a taste for good ideas. >> Right.
>> You can Now, there is no dearth of ideas that you can generate. >> Right. >> But now, I finding out whether a idea is worth pursuing, >> Mhm. >> that's where the taste comes in and I mean, that is something that comes with practice. >> Mhm. >> But it's very cheap now.
>> Yeah, but the confusing part is that also we can feed to the system, right? How how good is the idea? >> Yeah, so so >> So, so So, for the for the beginner, this is a puzzling thing. What is my role as a researcher and what AI can assist us. So, this is a very confusing the line that separates, right? So, >> Yeah. So, one of the big pitfalls of these large language models is that um
I mean, I guess the biggest pitfall is they are very creative, but they are also very um subservient. >> Okay. >> ask uh, to do something, they will try very hard to please you. >> Mhm. Mhm. >> They are built to please you. >> Okay. Okay.
>> So, why is AI writing, if you look at AI writing, it's sort of, uh, always in the sense of, "Oh, this is a great idea." >> Yeah, yeah. It always reply like that, yeah. >> Because they are trained to please you. Because the companies that are producing it are not, uh, producing it for, uh, um, the factual accuracy or whatever, right?
>> They are producing it because you you want to get, uh, dopamine high of, "Oh, I'm using it. I feel smart." >> Uh-huh. >> So, the biggest mistake that you can do is also give the verification back to it. >> Mhm.
>> So, you get some ideas and then asking it back again, "Give me the best ideas out of this." >> Mhm. >> That is the downfall of, uh, going towards a good research because it will it will say, if you say, "Oh, among these five ideas that I've narrowed down, is there a good idea?" It will say, "These five are great."
>> Okay. Okay. >> "You should go down this path." >> Uh-huh. >> Right? That is the, I mean, this this, uh, generation is great. >> Okay. >> But offloading verification, >> Mhm. >> offloading taste to AI, >> Okay. >> is a definite downfall. >> Okay.
>> That is definitely one. The second one is, um, think about what is happening, uh, um, in, uh, contemporary, um, AI use, right? So, there is Anthropic, there is OpenAI, there is some open source models, but everyone is asking the same model similar sort of questions.
>> Uh-huh. >> So, the biggest risk, the other big risk is that, um, everyone ends up homogenizing on same ideas in the same way. >> Right. Right. Right. >> Plurality is lost.
>> The plurality is lost. I think we, in research, we value ideas that are, sort of, out of, uh, the, uh, common understanding, right? Now, if you offload everything to AI, the same question you ask, 10 other people will ask. >> Mhm.
>> And AI will repeat the same thing like a slight variant of it. >> Mhm. >> And I think if we end up offloading all of it, we end up with the homogenization of thought. Right? And [clears throat] someone who stands out, who is thinking without an AI will really make the big world-changing thing, right? Because the whole world is going to homogenize.
>> Exactly. So, when we, you know, see all these videos or general readings, they say AI can think. So, do you validate that statement? AI can actually think. So, how much comparison with our brain thinking process and AI thinking process?
>> It's a It's a very difficult question to answer because I think for all practical purposes, the level of higher-order thinking that we do, I don't think we are using AI to that extent. But surprisingly, they will make a connections which you are not aware of.
But I don't think AI is thinking, right? It's definitely making connections to because it has a very wide body of knowledge, you can ask A and B and it sort of makes the connections between ideas.
And sometimes it can use external tools for validation. So, or you ask it to write a program, what it will do is it won't just regurgitate the program from its mind. What it will do is it will write a program. It will use your compiler to compile it.
And then it finds an error. Then it knows, okay, from this error I'm going to make a next step. But really, it is not thinking as in if you ask a person like an expert programmer to go to a whiteboard and write a program, they can work through it by just thinking about it.
AI doesn't AI isn't very good at doing that. AI is great at doing this when you give it tools. Okay. Okay. So, it is not uh I mean, for the sort of things that I do, which is mostly programming and programming language development. It's the tool used that AI is great at using.
So, it doesn't matter if if the feedback is unstructured or you give it a lot of data, you feed it hundreds of files, it will still be able to power through. But, what is really happening is it is making connections. But, I don't think it is original thought thinking. I I I refuse to believe that.
>> >> That's interesting. So, so in your area, can you just describe the area where you are working and the role of AI to what extent you're using it? >> So, so I work in this general area of programming languages.