Inductive Reasoning in Generative AI with Tulane University’s Nicholas Mattei

ai video series. I'm your host, Mike Ard. Today we're with Dr.

Nicholas Matay, who's professor of computer science at Tulane University, and we're talking about inductive reasoning, which is really at the core of all these gen platforms and the good, the bad, and the proverbial ugly as they say. Nicholas, welcome to the show. Hey, thanks for having me, Mike.

How's it going? Good, good. I think we all kind of sat there and when these things first came out and we were amazed in the wonder of it, but for folks who are kind of playing around with AI for a while, I mean, inductive reasoning's been at the core of it.

A a discussion has been going on for a long time, and may be all we ever really did here was throw a lot of computer hardware at the problem. But what is your sense of what's going on here with Gen ai as you kinda look at the history of AI and what are the, you know, pros and cons of inductive reasoning? For sure.

Um, so this idea of inductive reasoning is really sort of the basis of a lot of what we're doing with, you know, machine learning, and especially with the current gen AI is all built on some of these ideas from machine learning. And, and the idea of, of inductive reasoning, right, is that if I show you enough examples, then you can generalize to, to, to identify objects or, or do whatever, right? And so for machine learning, that's sort of the core way that it works, right?

Is that we're gonna use data from the past or data from, uh, the examples that we expect to see. And we're gonna use that to sort of understand and, and then classify or, or predict the world. And so for its gen ai, especially in these videos, um, and the, the images and, and the text, we're really just predicting the next word.

Um, and so I think one thing that, that folks that can get lost, I think in using these, these, these sort of modern, you know, the new AI techniques, I guess they're not really new. They've been around since the fifties and sixties, uh, always have a lot more money. And like you said, a lot more computing power now, um, is that the places that we, that, that the things that we learn are only ever gonna be as good as the data we have, right?

And so this is kind of this, uh, catch 22. So a lot of times we'll try to repurpose data. Uh, you look at some of the generative text, uh, uh, AI models, and what they're doing is they're taking data from the internet, right?

The way that folks have used language on the internet, the way that folks have used language, um, in books and things like that. And they're using the past language we've used to generate or predict new language. And that's, I think, a lot of the cases where some of these things can happen.

Where some of these sort of ethical issues can kind of crop up is because, you know, the way that we've talked about things in the past are maybe not the way we wanna talk about things in the future. And, and so that's kind of the, the joke that I've been using recently is this idea of like, sort of original sin, right? Like we, we, we've kind of like, we've used everything that we've already done in order to sort of make the new things that we want.

And sometimes we forget that that's an inherent, you know, part or possibly even flaw of the process depending on what we're doing. Um, and this is not, you know, super new, you know, um, there's a great book called, uh, automating Inequality that really talks about the same kind of ideas, right? Is using past data or using data from other domains or other distributions in order to predict the future.

And you, you're not, if your data that you're training on doesn't really match the, the problem that you have or the problem that you're trying to predict, that's, that's where a lot of these issues can come from. So when we're trying to like, guess what, you know, you or I are gonna watch on Netflix tonight, turns out that what I watched before is pretty indicative of what I'm gonna watch in the future. But if we wanna talk about something like school admissions, or we wanna talk about something about hiring practices, um, depending on what data from the past we're using, you know, it's gonna influence by definition the, the predictions that we make.

And I think that's where a lot of these, these kind of traps or these ethical issues do crop up. Are we thinking these things are smarter than they are? And maybe we're gonna get in trouble?

'cause when I look at it, they're probabilistic, so they're taking a best guess to something. But I feel like sometimes we're applying it to things that are inherently deterministic, and then we're wondering why things went wrong. And that's, I mean, that's a really good point too, right?

And so during the, um, the old AI boom, um, which predates even me, right? So there was an AI boom in the eighties all around expert systems around, you know, these decision making systems that were trying to formalize knowledge and then make, like you said, these deterministic, uh, diagnoses in some, in a lot of cases. Uh, it's a lot of medical technology in the eighties.

Um, you know, the, the kind of the gap there was that we could sort of write down all of our information and then use that to like, reason about the future. And so now we really take all this past data and we use that and as you said, to make a probabilistic or a prediction about the future. And that's the other kind of flip side of the coin for these machine learning methods is that they are just making predictions at some point.

And they're inherently noisy. They're almost always noisy. These large language models, regenerative ai, are probabilistic distributions over the words you could use.

And that implies that sometimes it's gonna guess wrong, right? It's gonna have a distribution that, that's gonna give you a word that's maybe not quite right. Um, and these machine learning models are really best suited for places where if you get one of these predictions wrong, the consequences aren't that high, right?

So again, going back to like Netflix or Amazon, if they had to recommend a product to me that maybe I'm not so interested in, or if YouTube recommends me a video that maybe I'm not so interested in, um, there's not a lot of downside to that. Um, there's a little bit, but not a ton. And the places where we can get in trouble is when we start using these things to predict things like, who should get credit or who should be eligible for housing or who should get hired, where we want to be more confident, um, in the decisions that we make.

And we gotta really investigate that, that probabilistic nature of these models. Um, and, and so it works great in some cases, but it, but sort of forgetting again, you know, that these are sampling from distributions. Like sometimes they're gonna be wrong.

Um, and, and, and try to apply that philosophy to high stakes decisions or, or really important decisions. I think just without paying attention to the underlying technology is, is really a, a recipe for, for problems. Some folks are starting to say that less might be more here 'cause we're over training the models and they're getting confused and, um, yeah, that's where all the hallucinations come from.

So maybe we should just be playing with a smaller amount of well vetted data before we just decide to throw everything in the kitchen sink at it. Yeah, I think this is, um, so my joke on this one is, you know, the, the i I work with a lot of, uh, folks in, in different disciplines and economics and, and sort of across, uh, the spectrum here at the university and kind of the, the machine learning computer science, like what we're all sort of trained for is, you know, more data is better, right? We're just gonna, we're just gonna pump everything in.

We've got all this computing power, let's just more and more and more on the data side. And when you go talk to, you know, some folks more in quantitative re you know, quantitative research methods and things like that, they're like, oh, well, but why this data? Like, why are you using that?

And I think that's a lot of times questions that we forget to ask ourselves. Um, I'm not sure if I'd call it over training. It's just this reliance on this idea that if we throw more data at the problem, we're eventually gonna solve it, right?

Like eventually with enough data, the computer will figure out what we want it to do and, and it'll work. And it turns out like this works really well, again, for certain things. Like when you have Netflix scale data and you're just trying to predict, you know, sort of small problems.

And I think there's a whole other realm of, of problems that we're trying to solve, which maybe, you know, just core machine learning techniques aren't amenable to, they work really well for a lot of stuff where we have a lot of data, um, and, and, and the things that we're predicting are not necessarily super high stakes. Um, but as we've are, as we're trying to, as we're seeing this get applied more and more places, I think some more of the problems that you're reporting to are coming up is that, you know, sometimes you can't just throw more data at the problem. Sometimes you have to go in and understand what's the process by which, you know, maybe, uh, a company is making a decision or, or, or allocating something and really try to encode that as part of your whole AI model.

And, and AI has techniques for this, right? Like as we, if we move away from just machine learning, you know, there's, there's, there's all kinds of stuff. You know, I teach the introduction to AI class here.

You know, there's all kinds of techniques that we, that we use to try to, you know, make the, make the computers do smart things, as I like to say. Um, and, and sometimes we need to use those. I think you're starting to see that with the later evolutions of these models where they're actually not just only using the output of a large language model, they're coupling it with information about the world.

They're coupling it with like data from Wikipedia that has some structured data, um, some structured understanding inside of it. And they're using that along with the language generation to try to, uh, you know, curtail some of these issues about, uh, the, the model sort of making stuff up. Um, because it's, it's, it's really a, a place where things can go wrong.

Um, at the end of the day, these language models, it's kind of fun if you think about it, right? They're, they're trained on a lot of data to sort of give you a, the next word in a sequence of words. And they're also sort of trained to make you happy as a user.

And so the model doesn't want to say, oh, I don't know the answer, right? It's always trying to make you happy, which is like, just give you more words that look reasonable. And the problem is when we use these in places where we want to have, you know, concrete verifiable, let's say citations or references, um, you know, sometimes it just, it'll just make stuff up because it's like, oh, well this looks like it could be a reference.

Um, you know, and stuff like that. So there's a lot of room, I think, to use the natural language, natural chatting part of the Gen ai along with, you know, kind of more core techniques in information retrieval, you know, which is what Google is at the end of the day. And, and, and, and, and using those to kind of combine those two things and, and sort of check what's coming out of this stuff, right?

So it's kind of gonna be a hybrid of these machine learning techniques and sort of more standard or good old fashioned AI techniques, um, from, from across the discipline, How secure are all these things? And I'm asking the question because there's been some studies recently that said, well, the LLM definitely wants to make you happy, but if you tell it that to keep a secret and then you keep asking 'em what that secret is, it will tell you the secret eventually. Yep.

Um, I think it, it, it kind of depends on what you mean by security. I think there's a lot of different, uh, levels to that question that, that, that folks are interested in right now, or right. So one is, you know, are the things that you even put into the text box on say, chat GPT or Gemini or any of those, you know, is that, is that private?

Um, and I know a lot of places, especially even us here in academia are running into issues with this, where folks are putting personal identifiable information in there, things like that, which is, I mean, if I could say one thing to everybody is like, don't do that. Like once something goes in that box, it's public, um, or has the potential to be public, um, is not a private conversation between you and a computer. Um, and so don't, don't forget that one.

Um, but in terms of secure in the, is it gonna disclose parts of the training set or is it gonna disclose, you know, can you sort of convince it to do certain things? Um, again, this is where I think it's really gonna come down to using some of these older techniques for it to understand conversation flow or understand maybe certain parts of it, uh, parts of that conversation and not just kind of respond to the context that's been built up. So it's, it's, it's kind of an interesting, interesting question.

Um, you know, back when I was first doing a lot of ai, I always thought, you know, I thought the stuff that we're doing now with, which is just sort of model free, you know, we're just gonna predict the next word and it's all gonna be okay. I never thought it was, I personally never thought it was gonna work. I'm not gonna pretend like I did.

Um, and the fact that we get so much complex behavior from these models that really don't have an understanding of grammar, right? There's no little bit inside of that LLM that's like, this is what a verb is and this is what a noun is, right? Like, it doesn't understand those concepts in the way that, like you and I understand those.

Um, not that it needs to, uh, but it, it is kind of interesting that we get such complex behavior out of something that doesn't reason in the same way that we do. Um, and I think especially the security question that you're kind of hinting at is, you know, um, how well is it understanding the context of our conversation? Like, don't, you know, to keep the secret don't, don't reveal it, you know, in later conversations.

Um, it, it, you know, it doesn't have this model of conversation, so to speak. And so it's, it's interesting. I'm not sure how to fix that problem.

It's so, it's a really interesting one and it's, it's kind of a, like a, like a tough one with the, with again, those techniques that are just kind of like looking at the past conversation and sort of generating the next bit of language. It has the capability to keep a secret of roughly your average 5-year-old. Um, we also see people talking about how the reasoning engines are gonna get smarter.

There's more parameters and more context win windows. And as part of that though, you can't help but wonder, um, how reliable is that gonna be? 'cause what they're basically saying is we're gonna have a bunch of agents that are asynchronously processing tasks on our behalf, uh, very confidently, no doubt.

But you know, all it takes is one broken handoff in a process and chaos ensues. So, um, yeah. How, how automated is automated gonna get?

It's a good question. I, you know, I think it's, I think it's impressive where we're at now. You know, if you think back to, you know, the late nineties, uh, a lot of the research that was going in that sort of eventually became Siri and some of these digital assistants, right?

Like at the end of the day, Siri or, or Google Assistant or any of these ones are, are actually handing off a lot of that processing tasks to other, to other programs, right? So when you ask it to play music on Spotify, like it's hand, it's kind of understanding the request, formulating that, and then handing it off to the Spotify app on your phone, for example. Um, and I think this is this kind of multiple agents interacting with one another.

Um, you know, it's been a, it's been a big research area in computer science for a long time. 'cause I think a lot of folks have thought that this is where it's gonna go, right? Like, we're not gonna build one, one AI model to, to do everything.

I'm not a big, you know, general AI person. Like really what it is, it's about all these little processes that hand off things to each other and, and really create a lot of that, uh, that value. And I think you see this in, in the products that we have now, right?

Like, it's something like Google Maps, right? It's, it's a map, but it's also a traffic, it's also a review, right? It's, it's, it's kind of all of these little bits sort of put together into one interface that are, that give you a lot of that value.

And I think that's where you look at where open ai, um, their new, their, their announcements for their new, their new, uh, assistant, things like that. That's where these things are kind of gonna go. But you're right in that, you know, how those handoffs work and where things get handed off to, like how does that ecosystem work, uh, in a, in a, in a, in a robust way.

It's not only an AI problem, right? It's a systems engineering problem. It's building a big thing that's connecting lots of parts that maybe don't want to talk to each other.

And those are tough problems. It brings in a lot of other issues around, you know, like engineering any big thing, um, where, where things can go wrong. Um, and so it's gonna be kind of interesting, uh, how that, how that develops and how we, we see the handoffs going between these different, uh, these different types of programs, especially across, um, providers, right?

So again, like your OpenAI agent that sort of converses with you and knows a little bit about you handing off that information to let's say Netflix or Spotify or any of these content provider things, um, you know, what's, what's, what's the, what, what are the problems that could go wrong at that, that interface, um, I think is a really interesting question, um, going forward. So is this turning into like, you know, AI in the Downton Abbey age, right? There's a head servant and a bunch of other servants that are doing all the tasks that nobody ever sees.

And um, hopefully, you know, we won't go broke one day, but, um, it seems the way that's the way to start thinking about all this stuff. Yeah, I think it really is. And I think that that's, you know, if you want to really kind of peel back the top layer that's, that is what's kind of going on under the hood.

And it's kind of interesting and that that that conversational, you know, the big, the big advance, um, has been in that conversational layer with these LLMs. Um, and they've kind of surprised us that they can do more than just have conversations they can do, you know, make predictions. They can kind of interact in, in relatively complex ways.

And then the next level is gonna be how do you sort of get that to hand off to various service and services and things. I mean, some of this already happens in the background and we just don't even notice it, right? So these generally l LMS are just a straight LLM that doesn't have other services kind of hooked into it.

You know, they're not very good at things like math. And so at the end of the day, when you ask Gemini or OpenAI or uh, chat GPT one of these math questions, it actually doesn't solve that math problem with the LLM. It actually hands it off to more of a calculator type program.

And so that's, that's does the simple manipulation, it does, does things like that. And so we've got, you know, general math software that's been around for ages, things like Wolf of Alpha or matlab, but now you kind of think of, you know, you talk to chat GPT, it sort of formalizes it into a thing that you could send to say MATLAB or something else, and that's gonna answer some complex math question. And maybe it hands off some of the search to, to other things.

I think you're starting to see this already with, um, the way that Gemini's being, um, included with Google search and like they're using the search just kind of augment the interface layer or the, the chat layer, so to speak. Um, but yeah, I think it is moving toward this Daton Abbey thing of like these different agents that do sort of different core things. And then you have your sort of assistant that's, that's talking to you.

This is the dream of Siri, right? This is what they promised us like 35 years ago and we're kind of, we're getting there, we're getting there, Getting there. I'm getting agents but no flying cars yet, but we're working on it.

Yeah. Waiting on that. So last question.

We see regulators are starting to wrap their arms around this a little bit there. In the very least, a few senators have learned how to spell ai, so that's good. Um, but you know, what's your sense of, uh, what do we need in terms of regulations?

'cause some people would argue, you know, we need to let this stuff evolve and see where it goes. And other people are saying, we need to closely watch what's going on here because it can spin outta control. And I don't know if there's a middle ground, but I only hear about the two extremes Seems right?

Um, it's interesting. So a lot of my concerns personally with it are, are more almost pro in a way. Like I, I'm more worried about, you know, what are the rules around data?

Like what are the rules around the things that I put in there? What are the rules around, you know, Netflix or not to pick on Netflix. I keep doing it just 'cause I watch a lot of Netflix.

Um, not to, uh, you know, what are the rules around how this data goes in there and how the data can then be used. And I think that's the first one that's gonna get sorted out, but it looks like it's gonna get sorted out in the courts, right? Between the New York Times and Sarah Silverman and a lot of these folks kind of asking the question like, where does it go from learning from what I've written before to generating things that are in my, you know, that are, that are replaceable or even in the New York Times case, like exactly the articles that have been in the New York Times that it's able to generate or recall.

Um, and so what are the rules around what you can train with? How close can it be in terms of output? So I think some understanding about that, understanding about who owns the data and what does that mean.

Um, I think that's a big question for like, people like content creators. Like if you are producing art or you're putting things on the internet, like what rights do you have about, you know, these things taking in and remixing, um, things like that. Those are all very unclear questions.

Um, that I think those are the biggest ones for me. I've always been a big data security, uh, not data security in the sense of, you know, what's mine? What do I own?

What are my rights over that? Um, because I think a lot of that's very unclear, um, these days. And, and it's a place where if you look at to places like Europe, um, they're a bit ahead of us in terms of guaranteeing rights around data.

Um, and I think the other one, the other big one is around these questions around like algorithmic recourse. So if you've been say denied credit, you've been denied some opportunity because you know, some computer, you know, made a decision about you, you know, what are the rules around, you know, appealing that process and things like that. So we do have some good rules around this for things like credit, um, you know, credit decisions made in traditional ways.

Um, but that was a long fight, right? That evolved for many, many, many years, um, between the eighties and today. And I think as we've moved some of this decision making into these models, you know, we can't really just let people say, oh, well the computer said this, so that's the answer, right?

Like, as you know, you, you need some kind of formalized appeals process, especially in these high stakes decisions into understanding what were the elements that went into that decision and what are your rights as a consumer or a citizen or something like that, right? So I think it's, I'm not a big fan of saying like the models themselves need to be regulated, but I think the inputs and the outputs and where, you know, we as, as, as users and as citizens and as just people have to interact with those things, what are our rights with respect to those, those end points and in the middle, you know, we'll kind of figure it out. I would tend to think All folks here earned here, hey, society is governed by rules and social contracts and norms that we've put in place for centuries.

A couple algorithms aren't gonna change that. Hey Nicholas, thanks for being on the show. Hey, No problem.

Thanks a lot, Mike. Thank you all for watching. ai.

You can find this episode, others on our website. We invite you to check them all out till then, we'll see you next time.