Episode Transcript
Hunter Zhao (00:02.366)
Welcome to the AI Chronicles. I'm your host, Hunter Zhao. Today, we're going to dive head first into how Carnegie Mellon is using AI inside their own institution, and we'll share the exact steps you can take in order to implement AI for yourself. Now, before I dive into that, listen closely. Are you looking to implement AI inside your own company, or are struggling to get your AI to stop hallucinating? Speak to GPT trainer.
GPT Trainer literally builds out and manages your AI for you, eliminating hallucinations for good. Go to gpt-trainer.com. I promise you, it'll be the biggest time-saving decision you've made all year. Trying to set up AI on your own is like trying to build a house from scratch. I'm sure you could do it, but the time and frustration it's going to take you, well, it just isn't worth it. It's 1,000 times faster and safer to hire professionals. Schedule a consultation today.
Once again, that's gpt-trainer.com. Today I have with me on the show John Kitchen, who is a professor at Carnegie Mellon. Welcome to the show, John.
John Kitchin (01:14.542)
Thanks for having me.
Hunter Zhao (01:17.03)
Awesome. So John, tell me, how did you end up at Carnegie Mellon? How was your journey led you there?
John Kitchin (01:25.964)
Yeah, so I started out studying chemistry as an undergrad and worked in the pharmaceutical industry for a while. I took a short break after that. I was a raft guide and climbing instructor for a short time before I got a job. And in that job, I started working with chemical engineers and decided it was time for me to go to graduate school. And so I decided to
go to chemical engineering in grad school at the University of Delaware. And there I experienced some industry and kind of got a taste for research and it was quite a journey to go from not knowing anything about chemical engineering to thinking I should be a chemical engineering professor. But that's what happened over five years in grad school. And then I interviewed for jobs and ended up here at Carnegie Mellon.
Hunter Zhao (02:19.038)
Wow. Usually when people say they interview for jobs, they don't become a professor at a university. The requirement's pretty stringent there, right?
John Kitchin (02:28.0)
Yeah, it's a challenging process and there aren't anywhere near as many professor jobs as there are industry jobs. But if you decide that's the job you want, those are the jobs you apply for. And you have an interview process, you give research talks, you give teaching talks, you meet with all the faculty, and then hope for offers.
Hunter Zhao (02:54.014)
Nice, nice. So I know being a professor usually that the, you know, they're motivated by one or two things or both, right? Either teaching or doing research. And I know Carnegie Mellon is a very big research institution. So what do you prefer or do you like both?
John Kitchin (03:10.05)
Yeah, so I mostly live my life by integrating things. So I teach to help the research, and then through research, I find new things that we start teaching. And that's the only way you can manage in a job like this, where each one is a job of its own. And so if you don't find ways to leverage it, for example, I teach classes on mathematical modeling with Python. And that's
allows me to teach students how to write Python code for scientific programming. And that sort of underpins all of machine learning. If you can't do that, you're not doing machine learning and you're definitely not doing AI down the road.
Hunter Zhao (03:49.456)
Absolutely, that makes a lot of sense. mean, knowing the most state-of-the-art tools that's used to do something that's very widespread nowadays is a good foundation for people to have. on the topic of AI machine learning, I know you're using AI over at Carnegie Mellon, and I think it's fantastic what you're doing. But before we dive into that, can you tell me how and why did you decide to use AI?
John Kitchin (03:56.398)
Thank
John Kitchin (04:10.669)
Mm-hmm.
Hunter Zhao (04:18.086)
AI in the first place, and what challenge were you trying to solve?
John Kitchin (04:19.895)
Yeah.
Yeah, so when I started as a graduate student, a lot of what we do in science and engineering is build models. And the traditional models are based on physics and chemistry understanding, know, Fick's law, the Arrhenius equation. We know how the math relates to what we observe. But as, as systems get more and more complicated, we don't know those equations anymore, but we still need models for engineering. And so I really, turned to machine learning as a way to
continue building more more sophisticated models for solving engineering problems that we have no hope of learning chemistry or physics based equations for, but we still need models to make predictions. So that's how I got into seeing machine learning as the next frontier of model development.
Hunter Zhao (05:09.894)
I see, I see. That's a very interesting topic that you brought up, John. And being an engineer myself, I've often dealt with the models you mentioned, right? Starting with the physics-based ones and then eventually computational ones. And they're usually, they have to be, to be honest, for rigorousness, to be grounded in some kind of theoretical basis. But what I'm thinking is, know, large language models are all the talk nowadays.
John Kitchin (05:22.253)
Mm-hmm.
Hunter Zhao (05:39.446)
And they're really more of a correlation-based model, I mean, a statistics model. So do you think there is any, I guess, just in general, what are your thoughts about using that kind of statistical-based model versus something that's more, I guess, and effect, that has a physics basis?
John Kitchin (06:02.721)
Yeah, think we're seeing large language models have been around maybe two and a half years, and they've been changing very, very fast. So two and a half years ago, you mentioned hallucinations before. I think of hallucinations as you're generating statistical sequences of words in a space you don't know the statistics very well. They've gotten much, much better and much, much more powerful. And so now one doesn't.
just use an LLM to say what's the most likely sequence of words. There's web searches behind it. There's tools that get involved. There's some kind of like context management, know, rag or something like that. And so it takes quite a lot of skill to use the LLMs productively right now to make sure things are going in the right way. And you have to couple it with domain knowledge so you know where it's going and you know if it's going in the right way.
So I'm actually quite bullish about how we're going to be using these in the future, but I don't think it's going to be, you know, open a desktop app and, you know, ask a 12-word sentence and expect answers, right? It's going to be much more sophisticated than that, and you will have to learn, just like working with students, you don't just ask students to do something. You provide them with prompts. You give them very specific directions, and you scaffold it to where they eventually are able to do what you want.
And I think LLMs will be, eventually we will see them this way too, that like a CNC machine in a shop, you don't turn it on and throw a piece of metal in there. You give it very specific directions on how to do it. And when you know how to tell it what to do, you get beautiful results.
Hunter Zhao (07:44.808)
That is indeed.
very true. mean, definitely alluding to the hallucinations part and explaining it, and I'm sure that's very helpful analogy for our audience. so John, language models is one thing, and people nowadays frequently equate language models with AI, but of course AI is much broader, and I've heard, and especially you being a chemical engineering professor, I'm sure you're very familiar with the scene, but you know, AI has also
John Kitchin (08:05.069)
Mm-hmm. Mm-hmm.
Hunter Zhao (08:15.872)
applications, direct applications in perhaps molecular design or especially when it comes to pharmaceuticals, right, drug discovery, right, things like that, synthesis, etc. Now, those are not language models, obviously. So how do they relate? Are they even at all related in any way from perhaps the foundational math or the statistics?
John Kitchin (08:23.575)
Mm-hmm.
John Kitchin (08:38.965)
Yeah, think when so there's a couple of styles. So if we stick start with language models and think of that more generically as a generative model where you're generating something, people are trying to generate things other than text. So now people generate images, you can generate audio, you can generate molecules. We have some very exciting work on generating solutions. So if you can figure out
There's sort of two styles of generation. One is you have learned some statistical distribution you can draw samples from. And the other is you have one distribution that you can transform into another. And those are like the flow models, if you're familiar with it. But either way, if you can generate an image, then you can think about generating a vector, which is just a flattened image. And if that vector is inputs and outputs of a complicated process, now we're generating like chemical plant data or
solutions to problems. That's just on the generative side. So to me, there's like a lot of similarities in the idea of generating text or images or engineering solutions.
Hunter Zhao (09:48.766)
I I see. And that makes a lot of sense. But with the desired output being a certain way, I imagine the input to the model in the initial training process would also have to be pretty clean. You have to have a large training data set, and then you have to extract the right features and make sure it's well represented.
John Kitchin (10:07.797)
Yeah, that's right. And so, you know, the large data is usually a problem when you need experiments. And so there's, you know, there's a whole bunch of other ways where people are trying to bootstrap, say, computational things where you can have large data with experimental data, you know, something like contrastive learning, where you can transfer something that is easy to get large data for two experimental systems, or you start looking at hybrid physics and machine learning. So
you try to get away from purely data-driven, which requires enormous data, and instead rely on approximations to the physics, which require much less data, but are approximations and less accurate.
Hunter Zhao (10:51.198)
Of mean, at some point you could also argue the use of multiple types of models to the lower fidelity ones. They get a quick estimate with uncertainty bounds, wider uncertainty bounds, and then you fine tune the model once you get. I mean, if it's an optimization problem, you always go to the close to the optimal, and then you fine tune the results using sharper models.
John Kitchin (11:00.715)
Mm-hmm. Right.
John Kitchin (11:11.158)
Right.
John Kitchin (11:16.459)
Yeah, the multimodal models are changing things a lot now too, because now we can use much richer data. Like you can use a photograph as a source of data or an audio stream as a source of data. And before, you know, 20 years ago, that was almost inconceivable.
Hunter Zhao (11:32.796)
yeah. I mean, I remember when I watched the episode of Star Trek when the time traveled back, right, and then to the inventor of the transparent aluminum and the guy was talking to a computer. Right, I was thinking, wow, would never have been, I would never have thought it was going to be this soon that we can do that.
John Kitchin (11:44.781)
Yeah. Yeah.
John Kitchin (11:53.867)
Yeah, it's changing very fast.
Hunter Zhao (11:57.288)
Yeah, yeah. So when it comes to concrete AI applications and usage in higher education, right, in universities like Carnegie Mellon, can you share some real examples and perhaps describe the actual process, aside from the chat GPT stuff, obviously?
John Kitchin (12:03.788)
Yeah.
John Kitchin (12:15.916)
Yeah, probably the biggest thing we've worked on for the last 10 years is in molecular simulation.
where the standard would be quantum chemistry. That is a physics-based framework for running calculations, but it's very, very expensive to run and limited to quite small systems. And about 10 years ago, many people, including myself, started looking at machine learning ways of building that. And so 10 years later, we now have graph neural network models trained on half a billion DFT calculations, these quantum chemical calculations.
almost as good as DFT in many ways. We still have to worry about something that is equivalent to hallucination. If you try to calculate something it's never seen before, it's not very good. And it's hard to tell in the material space what that means. When is carbon not like carbon that it's seen before? If it has a very long bond length or if you go to very short bond lengths, then it doesn't work as well because it's an
extrapolation of a nonlinear model. So that's probably the biggest thing that's happened. Much, much more recently, just in the last year, we've started integrating LLMs into all kinds of things. So we can run instruments now. We can design experiments with them. We can search the scientific literature and get summaries of articles and use them to facilitate the search of scientific literature that's not keyword limited. So we've seen lots of advances like that.
Hunter Zhao (13:24.264)
Right.
John Kitchin (13:51.952)
happening.
Hunter Zhao (13:53.598)
I see, I see. So that's very intriguing. Now in terms of the real results that you've been seeing so far, now I'm talking about the operational efficiency results in your lab, for example, or ways to be more creative in your thought process or exploratory direction in research, Have you seen anything change? Is it a significant improvement?
John Kitchin (14:06.444)
Yeah.
John Kitchin (14:12.535)
Yeah.
John Kitchin (14:19.306)
I guess I'm going to say it's different. what's tricky is, and I challenged one of my graduating students with this, my students are not writing more papers every year. They're writing the same number of papers every year. They're just on different topics. And they're on topics we couldn't have possibly done five or 10 years ago. They still work a full week in the lab. I still work a full week.
at work so like AI hasn't suddenly given me a 20-hour week for example.
we do more things. like in the past with just quantum chemistry, we might've been limited to a hundred calculations and now we can easily do a hundred thousand calculations, but it takes the same amount of time. It's just a much richer data and a much richer story to tell. And so it's hard to say that, you know, how to talk about it in terms of efficiency. If you were saying like, well, it's more data per time than yeah, it seems more efficient, but it's actually the same amount of time at the end of
day and it's just a different way we end up spending it.
Hunter Zhao (15:29.918)
I see, I see. Well, thank you for sharing. So John, what are Carnegie Mellon's upcoming AI initiatives? Is there anything, I guess, formal that's institution wide, or do you see AI playing a role in your operations next?
John Kitchin (15:44.471)
Yeah, we have a facility called the Bakery Square Lab that is a large facility for automated science that we're trying to integrate machine learning and AI across it.
In our own group, we're trying to integrate, find ways to integrate large language models into an electronic lab notebook where you'll be able to be typing in your notebook, interact with either the agents or the tools to have it do literature research and kind of write summaries in your notebook, maybe design experiments and put the design, capture it in the notebook. Maybe it will even run instruments all directly from a notebook. And that's integrated with all kinds of search tools to
to help you find things.
Hunter Zhao (16:32.99)
I I see. So I guess in terms of learning more about AI in higher education or at Carnegie Mellon in general, where do you recommend people to go?
John Kitchin (16:46.338)
Yeah, so part of what I've done in the last 10 years is developed a lot of educational content and it's you can find it at pointbreespubs.comroad.com. This is a company that I started to kind of disseminate educational materials that go from print hello world the very first day of the first year.
undergraduate education, all the way through LLMs in the cloud and everything in between. So it covers all of the foundational scientific programming needs of undergrads, grad students, transitions into data science and then machine learning, differentiable programming. And then there are two books on
Large language models one of them using Olama, so it's totally local one of them using The cloud all the cloud vendors and light LLM. So Again point breeze pubs gumroad.com is the place to find that and there's also some interesting Something interesting I call cybered fiction which is using large language models to create Long form stories that have audio images and text
in there. That's been something we've been experimenting with to learn how to use large language models effectively.
Hunter Zhao (18:09.648)
I see, I see. Thank you, John. And for our audience, once again, that's Point Breeze Pubs. All one word, no dashes in between. PointBreezePubs.Gumroad.com. So feel free to check it out. Thank you very much. And amazing, John. mean, that was great hearing your thoughts about AI usage in higher education across not just large language models, but also
for chemical engineering research and some of more technical fields. So yeah, that said, once again, remember if you're looking to implement AI into your business today, you don't have to do it yourself. The time and stress that the AI is going to cause just isn't worth it. Schedule a call with GPT trainer and let them build out and manage your AI for you. Go to gpt-trainer.com to schedule your consultation.
John Kitchin (18:55.96)
Thank
Hunter Zhao (19:10.504)
Signing off for now, have a wonderful rest of your day. And looking forward to seeing you on the next episode of AI Chronicles.
