Transform Your Teaching

What is happening at the bleeding edge of Generative AI? Can an LLM ever be controlled? What is being done to help users get the best outputs from their LLM? Join Rob and Jared as they discuss this with our newest needle mover, Dr. Carmen Amo Alonso, Postdoctoral Fellow in AI at Stanford University.

View a transcript of this week's episode.
 
Resources
Chat with us!

What is Transform Your Teaching?

The Transform your Teaching podcast is a service of the Center for Teaching and Learning at Cedarville University in Cedarville, Ohio. Join Dr. Rob McDole and Dr. Jared Pyles as they seek to inspire higher education faculty to adopt innovative teaching and learning practices.

Carmen Amo Alonso:

Imagine that we lived in a world where when we board planes, you know, sometimes they might fail, sometimes the pilot might lose control and they crash. So that's the state where we're in with AI, unfortunately. We cannot fully control them. I think just building that awareness that AI is not a controllable system in the same way a plane is.

Narrator:

This is the Transform Your Teaching podcast. The Transform Your Teaching Podcast is a service of the Center for Teaching and Learning at Cedarville University in Cedarville, Ohio. Hello, and welcome

Ryan:

to this episode of Transform Your Teaching. In today's episode, Dr. Rob McDole and Dr. Jared Pyles conduct a needle mover episode with Dr. Carmen Amo Alonso.

Ryan:

Dr. Alonso is currently a postdoctoral fellow at Stanford University affiliated with their computer science department. Thanks for joining us.

Rob:

We're actually very excited to have you on the show. You you have been highly spoken of by Dr. Oakley. I'm very interested in, what you're doing and your work, and, look forward to this conversation.

Jared:

Would you be so kind as to give us a quick bio of who you are? And then we'll start with the questions that we have for you.

Carmen Amo Alonso:

Yeah. So my name is Carmen. I'm Alonso. I did my undergrad in Madrid in aerospace engineering. And because I was really fascinated by the question of how is it possible that planes fly if they're heavier than air?

Carmen Amo Alonso:

But then a harder problem is how can you why are they so safe if it's such a hard? Aerodynamics, it's really hard and yet planes are so safe. Why is it possible? So then I went on to do a PhD, master's and a PhD at Caltech on the reason, the math behind why they're so safe, which is called controls theory. So I did this for my PhD.

Carmen Amo Alonso:

And then for my postdoc, I pivoted to AI. I spent some time at ETH Zurich at the AI Center in Switzerland. And now I'm a postdoc at Stanford working on applying the tools of control theory that made planes so safe to AI to try to make it safe and controllable also.

Jared:

Oh, wow.

Carmen Amo Alonso:

So yeah.

Jared:

Very cool. It is

Rob:

so cool. How in the world did we have her on our podcast? I mean, seriously. Like, you have just raised the water level of our podcast. We we really appreciate your time.

Rob:

So let's just jump right into some of these questions. So you just said control theory from aerodynamics, well, I guess from flight. And now you're taking that and you're trying to move it over into controlling AI. Is that correct?

Carmen Amo Alonso:

So That's right.

Rob:

What how do our listeners even get any kind of conceptualization? What kind of analogy can you give us that would help us wrap our heads around what you're doing?

Carmen Amo Alonso:

Yeah. So control theory is is responsible for a lot of the successes in engineering. I gave you the example of a plane, but control theory is also responsible for the power network being so reliable or the Internet. And people even use controls theory for bacteria and to model bacteria in our guts. So it's really a tool that explains how the behavior of a system evolves and how can we make sure that this evolution is the way we want it to be.

Carmen Amo Alonso:

So for example, in the case of a plane, how can we make sure that we follow the trajectory that we want to follow? We land when we wanna land and all of that. So in the case of AI, when we apply these techniques to AI, what I wanna achieve is making sure that AI does what we want it to do and doesn't do what we don't want it to do. Almost like a very well trained dog that listens to us and doesn't do the things that we don't want it to do. But even though this sounds very simple, it's actually very hard to achieve.

Carmen Amo Alonso:

And probably most of you have seen how how sometimes AI doesn't do what you want it to do, and that's a big problem.

Rob:

Yeah. I think I spend more time refining system instructions or prompts. I can get a really good, you know, output one time, but the repeatability of that output, especially as your conversations go on, like, it drifts. Right? Regression to the mean is the best way I can think of it because it's a probabilistic model essentially that you're working with.

Rob:

And so it seems like everything just regresses over time. And I really want to be able to make sure it's repeatable and that it, like you said, it doesn't do what I don't want it to do. It doesn't make stuff up.

Jared:

Is that the big problem you're trying to solve? Is that making sure that it remains consistent, getting rid of the inconsistencies and the hallucinations that have popped up?

Carmen Amo Alonso:

So it's about so Robert, you you alluded at how you try to fine tune the prompt, how you try to get your model to get it to what you wanted to do. That is control. And my goal is making these models so that they listen, so that we can control them. And so Robert doesn't have to go on and on and on with refining his prompts to make sure it does what he wants it to do. So that is the goal.

Carmen Amo Alonso:

And this is hard because a lot of the times, these models just don't listen, and and it's not even clear why. Yeah.

Rob:

So you're talking about training a model to do this? In other words, in in the training of a model, are you also making sure that whatever this control these control mechanisms that you're wanting to put in place are also trained into the model itself? Is that is that how you're going about doing it?

Carmen Amo Alonso:

So that is that is one way. Another way I have done it is you take a trained model and you you can design what we call controllers for these models, which are some some other modules that we append to the existing model. And that just on its own can improve the behavior of the model. So ideally we would train them to be controllable and then together with controllers, but I've also have work showing that you can just take a model that has been trained without any control, and and you can design controllers for it and make it better. So there's a lot of ways we can go about this.

Carmen Amo Alonso:

Yeah.

Rob:

Well, let's talk about some of those.

Jared:

Yes. Please do. So one of

Rob:

the ways that I've I think I've seen some success at least in the output, and this is anecdotal. Right? I haven't done any scientific testing other than just my using of it, is the retrieval augmented generate generation rag, essentially. So having other documents that I constantly are getting put in pushed into the context and that the model has to take into account or follow. So much like just simple instructions, but doing that up of a more robust dataset of of rules and regs that I wanna do.

Rob:

Is that kind of what you're talking about, or is there something completely different?

Carmen Amo Alonso:

It's actually different than that. Yeah. So we don't have to be sophisticated as RAG. You can try I don't know if you've ever tried it, but if you haven't, I encourage you and and, like, maybe people listening. To do this this simple experiment, there's this vision language models where you can tell the model to draw something for you, and then it it draws it.

Carmen Amo Alonso:

And you can tell that something like a cat next to a desktop computer and a pink elephant is missing, and then it would draw a huge pink elephant. And that shows how this is how it just doesn't fully understand like, the the the concept of cat elephant computer, this is all in the training data. It doesn't need rag to to access this information, and yet it doesn't understand understand basic grammar to to show the picture to to to give you the picture that you're that you want. Okay. And this is the problem.

Carmen Amo Alonso:

These models are not listening to our instructions. We cannot control them with just using prompting. So this is kind of the problem that I'm trying to address. Another example is if you're trying to get your model to speak a different language, it it just it it's just really hard. I don't know if you've tried, but I have.

Carmen Amo Alonso:

And it switches back to English constantly. It's very hard to keep your model in in in another language. So these are the kinds of problems that my research tackles. Yeah.

Rob:

Wow. I'm just curious how we can avail ourselves of the goodies that you are bringing forward that we can avail ourselves of of these things now. Is that something you're willing to share, or are we gonna have to buy your book when it comes out?

Carmen Amo Alonso:

No. I can I can explain? Yeah. Yeah. So I think we're using some some of the very same ideas that we use for controlling, say airplanes, which is feedback.

Carmen Amo Alonso:

So when you are flying a plane or even driving your car, you are encountering things along the way and then you're reacting to them, but you're not breaking right before you hit a wall. You're you see the wall ahead of you and you start planning and correcting your trajectory as you go along, and you use feedback to do that. You use visual feedback, for example. This is the same that we do in LLMs and not to get I I don't wanna get super technical here, so I just wanna give the intuition. But when you typing a prompt into an LLM, it goes it it gets processed.

Carmen Amo Alonso:

And it's it's a it's a it's a series of steps that it goes through before you get to an output. So during that evolution, you can see this as the trajectory of your car. And you can start applying the control techniques that we use for, say, self driving. We can apply or or piloting a plane or and so on. We can apply those same mathematical tools in in the process as the generation goes along to control the output and get it to the to to look like we want it to look.

Rob:

Now I'm just wondering what those mathematical tools are.

Jared:

Yeah. I'm I'm curious because if, essentially, we're making it so that it costs the user less money to get the output that they want. Yeah. For the like, for if you think about token usage and the amount of time and the amount of tokens it takes for someone to get to the answer that they want. Mhmm.

Jared:

This research, these controllers that are in place are essentially going to fix that problem.

Rob:

Especially if you can depend on them.

Jared:

Correct. Yeah.

Carmen Amo Alonso:

Right.

Rob:

So what can we put into place now?

Jared:

What is something we could try?

Rob:

What is

Jared:

something that we could We could back to our offices here in a little bit.

Carmen Amo Alonso:

So for example, I have this paper. It's it's together with Apple, and it tells you with guarantees probabilistic guarantees whether so so you you choose a feature for your output. For example, the amount of pink elephant in a picture. And and you you try your prompt and it's gonna tell you whether your model you choose a model, you choose a feature, and it's gonna tell you whether your model is capable of producing anywhere in that scale from no pink elephant to the whole picture being just a big pink elephant given given your prompt. Meaning this this tool is going to tell you whether your your model is controllable or not with just the prompt.

Carmen Amo Alonso:

And if it's not controllable to with just the prompt, then you can use some of the other work we developed to make your model controllable. And this tool, people can I I can send you the link? People can download it on GitHub, and it's very easy to use and implement if you know how to use an open source model.

Jared:

Oh, boy.

Rob:

We do. Please do. Oh, and this goes right into what I've been working with here recently with OSC. Right. So for our listeners computer.

Rob:

Is the Ohio Supercomputer Center, and that was the NSF workshop that I just went to last week. And I was just blown away at what researchers were doing with open models and the return on investment they were getting, it just boggled my mind because all we've been hearing about you know, most most everyone just hears about what's going on with OpenAI, Gemini, Claude, and, you know, you have your occasional grok from

Jared:

Mhmm.

Rob:

From Elon Musk. So those are the ones that we Meta. Sorry. I forgot about Meta. Those are the ones that we hear about the most, but I think a lot of people don't know there's a whole other massive side of AI that's going on in the open source world that can be seen in models that are on Hugging Face, which is I mean, if you ever wanna go out to a website and be completely confused about everything that's possible with with models, you you should go there.

Rob:

Right. But the fact that you're actually you're creating tools that would be helpful for those of us, especially researchers who are trying to to get certain results.

Jared:

Are you seeing, maybe as the, the models have updated and changed, are you seeing more towards what you, intend to do or like, do you feel like it, that you're going to need some sort of controllers like you have talked about in order for it to truly be controlled and managed? Or do you because, you know, we've we've talked about, like I did a presentation on prompt generation at a conference last year, and some of the stuff that I presented then by this point has pretty much been outdated because of the advancing of the models. Do you see a point where the machines are programmed to a point where they're under control and managed, or do you feel like there's a need for the controllers that you're developing now in order for them to be at that point that you where they're totally controlled and there's no chance of it not answering or not listening?

Carmen Amo Alonso:

I think it's a really hard question to answer because we really don't know where these models are going. And, but my guess would be that to get to that point of a model being completely reliable and controllable, it will take a lot of data and a lot more computational effort because what happens during training is we fall into this local minima. And I think it'll be really hard to move past that and get more of an optimal solution. We don't even know how to define these cost functions so that the model is controllable. So I do think that the ideas from control that we developed are going to still be applicable like years from now in the same way that the ideas from control that were developed for early planes are still in place and the controllers grew more sophisticated, but the the basic ideas and the principles are still what makes what makes planes so safe.

Carmen Amo Alonso:

So I think those the same ideas are gonna still be there. Probably the specifics in the math needs It will need to be updated as the models get better. But I do think that it's probably the most efficient way to get us to having controllable AI rather than just hoping that by increasing the number of weights and the training data and the models are gonna just be controllable. I don't think that that's gonna happen. But it's but it's an uncertain.

Carmen Amo Alonso:

I don't know. I don't know for

Rob:

sure. Well, it it does give me hope. Yeah. Right? It gives me hope because one of the one of the things I'd like to do is I I've been having conversations with some folks again at OSC, and I'd like to do a very small model that and I don't even know if it's worthwhile doing.

Rob:

I'm talking with them. I'm thinking about it personally. I teach business ethics here on campus, and I would like to train a model from a business ethics perspective that we teach here and have it be controllable and be able to have students do simulations with it. So that's kind of that's kind of where I I wanna head, and it sounds like your tools could help me. Is that

Carmen Amo Alonso:

Yeah. That sounds very exciting. I think so. Yes. I think that you can you can probably use our controllers to sort of have it automatically tuned for your response instead of you having to go through the prompt many, many times.

Carmen Amo Alonso:

And in fact, you can use your or Apple tool to to see whether your model is controllable. And if it's not, you can develop controllers for it. Yes. So it looks like you can yeah. Hopefully hopefully that that's helpful.

Rob:

Yes. Alright.

Jared:

So Rob is very much computationally minded, if I can put it that way. Rob likes to share, all this cool stuff that he's doing and he throws these terms around all the time and I have to be someone that goes, stop. What does that mean? Because I'm more of a pewter illiterate maybe.

Rob:

I think it, yeah, it's a difference in literacy. It's difference in experience.

Jared:

Right. So, and I feel like sometimes, like you two right now are speaking over my head. And I'll just be straight up honest with you. I was with you up until you started talking about adding like, GitHub to me is awesome. I've used it twice, but at the same time, I feel like I'm out of place when I go to the website.

Jared:

So speak to me like someone who is never really gotten into GitHub or is very limited in their understanding or research of it. Speak to me like I'm not Rob and and help me. How would I get started? Because this sounds incredible. I'd love to do it, but I don't know where to start.

Jared:

That's great.

Carmen Amo Alonso:

Yeah. So I think that AI can be really helpful here because a lot of the coding platforms right now and softwares have AI integrated into them. And I I know people working on Anthropic, they say that they haven't coded in in months just because they just talk to the AI and the AI does the coding for them. So if you just know what to ask for, you can just chat to the coding interface to the to the AI interface. You don't need to do any of the coding, but you do need to have a coding program and some some sort of Python program so that you are able to so that so that the computer can run the code for you.

Carmen Amo Alonso:

So I'd say what you would do is you go you talk to the AI and you ask it, you wanna download an open source language model, for example, LAMA that was created by Meta. And and then it's gonna guide you through the process of how you download it. It's gonna write the code for you of how you run it. And you can ask it for a little example, like, okay. Set it up so that I can type in a prompt, and it produces some some generation.

Carmen Amo Alonso:

And I'd say that's the first step, and you can achieve this without writing any code at all. And then we can talk about how to how to interface that with the controllers, but I'd say that's the first step. Is that more clear?

Jared:

It is. Yeah. I'm glad that you said it because when I'm when I'm I've asked this question of people before, the answer of, we'll just ask chat GPT how to use it. That doesn't help me because I don't know what to ask, but you provided perfectly what it is I need to ask, which is a huge help because sometimes it's like, okay, I'm back to where I was. I don't know what to ask for.

Jared:

It's like, it's that I don't know what I don't know. Right. Kind of a thing. Yeah. I don't know where to start.

Jared:

So you definitely answered that. So I appreciate that. Thank you. Something I did, Carmen, this is a story that we've told so many times on this this podcast and I wanted to do one of those interactive family calendars at my house for the family to keep track of all the events and weather and do a meal meal plan and stuff like that. And I, you know, I kept seeing all these advertise, but I don't want to spend the money on it.

Jared:

So I knew there's a way of doing it because I'd seen some open source stuff. And so, Rob and I sat down, he walked me through it. And so I now have a using Codex. I've programmed a TV in our house using a Raspberry Pi to, you know, show all this stuff and it's really cool. And now I'm anxious to figure out what's next, but I don't know what's next to do.

Jared:

So, but yeah, this is this is maybe this is the direction I need to go in and start building my own LLM myself. Yeah. Which sounds fun.

Rob:

Yeah. I mean, if you have an idea,

Jared:

like As long as I don't nuke my computer

Rob:

or something by doing it,

Jared:

you know?

Rob:

Yeah. If you download a model that's larger than your memory can handle, you might you might hose it. I don't know. It's super exciting to know that there are tools that are out there that you've been working on this. I mean, I definitely can see why doctor Oakley thinks so highly of you Mhmm.

Rob:

And that you've been a a good partner with her and her work. What would you say I mean, we kinda did that. You you've kinda asked the question, but, you know, what would you say? What would you leave with our listeners, if if, you know, you only had a few minutes to talk to them? And we have mostly faculty, higher ed faculty that listen to the show.

Rob:

We've got k through 12 teachers as well. So you're speaking you know, this is definitely probably gonna be a more a higher level podcast and technicality than we normally provide, but nonetheless, I think it's still extremely valuable Mhmm. Especially for those who are looking for the next step to take in their AI journey. What would you tell them? What's their next step?

Rob:

What's the easiest step our listeners could potentially take in getting to this place where they could use your tools?

Carmen Amo Alonso:

Yeah. I think at this point, my priority is to build awareness for why controllable AI is so important. And imagine a plane that imagine that we lived in a world where we when we board planes, you know, sometimes they might fail. Sometimes the pilot might lose control and they crash. So that's the state where we're in with AI, unfortunately.

Carmen Amo Alonso:

We cannot fully control them. So to build that awareness, I often just recommend people to try out this experiment with a pink elephant. Just ask ask a model to give you whatever image and then say, and a pink elephant is missing and it's gonna draw a huge elephant. And then that shows, and like, you don't have to do this extreme. Probably people have already, when talking to AI, have seen how AI sometimes doesn't listen.

Carmen Amo Alonso:

And that is a problem, and that's why we need controllable AI. I think just building that awareness that AI is not a controllable system in the same way a plane is, and that we have to continue doing research to make it controllable is what I what I will leave listeners with.

Rob:

Okay. Interesting. Well, I mean, awareness is definitely, I mean, really good because we we have said on this podcast, don't talk to it like it's a person because it's not. It's a machine. And you're taking that, and you're you're underlining that for us even more because you're saying, we can't control it.

Jared:

Right.

Rob:

We need to control it like we do airplanes. This is think And we're still working on that, so don't go into it naively. That's that's what I'm hearing.

Jared:

Yeah. It seems like it's also one of those great cases for AI literacy

Rob:

Yes.

Jared:

And the importance of analyzing the results that you get for accuracy. And, you know, this has been a great insight for sure into that.

Carmen Amo Alonso:

Thank you.

Jared:

I, I want to have you back on because I feel like you're breaking ground. I mean, at the point where there is literally no marks in the dirt, like it's brand new ground.

Rob:

We will definitely be following you, and we look forward to seeing some of these tools, seeing if we can you know, I I personally will see if I can use them. I know you will probably now go Oh, yeah. Okay. How can I use this?

Jared:

Yes. Absolutely. Absolutely.

Rob:

So thank you very much.

Carmen Amo Alonso:

Thank you. Thank you so much for having me.

Ryan:

Thanks for listening to this episode of Transform Your Teaching. If you have any questions or comments about our interview with doctor Alonso, feel free to reach out to

Jared:

us at

Ryan:

CTLPodcastcedarville dot edu. Feel free to reach out to us on LinkedIn. And don't forget to check out our blog at cedarville.edu/focusblog. Thanks for listening.