Exploring Practical LLM Research In Class At MIT

To many who are looking closely at where technology is going, it’s a bewildering landscape – there’s a lot of complexity, and quite a lot of uncertainty, as we move forward. One thing that many people can agree on, though, is that academia should have a seat at the table. In other words, we should consider what’s happening in classrooms, as we’re also considering what’s happening on the stock market or in corporate boardrooms.

So, partly because I recently focused on keynote remarks by Jensen Huang at CES (and other industry tidbits), I’d like to move into some of our own exploration at MIT, in the classroom, where young people start to learn about AI, sometimes for the first time.

In an MIT Deep Learning class, Ava and Alexander Amini manage a syllabus for those who are going to go out and be that next generation of applied knowledge experts. In a recent session, we had input from two seasoned researchers.

Peter Grabowski runs a Gemini Applied Research group at Google – and that’s no small thing, as the company prepares to deliver a version of Gemini on consumer smartphones. Grabowski talked to the class about various elements of LLM research and why they are important. Then Maxime Labonne a Machine Learning Scientist, Author, Blogger, and LLM Developer added some insights on what’s happening with systems now.

I’ll go over some of the duo’s remarks, and what we covered in this class.

Envisioning Powerful LLM Systems

Peter Grabowski has a lot of experience with LLMs. He talked to the class about how systems are evolving. For example, he explained what parameters are, and why having systems with billions of parameters is beneficial, and advances the science behind applied AI.

Using an example from Dickens, (It was the best of times, it was the worst of times) he talked about how a context window should be big enough to enable the LLM to work properly. I thought this was a pretty good example: because you have a very limited input set, with a word repeated, the system can get stuck in a loop. That alone demonstrates the need for data set diversity.

“If you're thinking about the number of parameters as a mechanism for understanding and representing information about the world, the more parameters, the more you're able to do that,” he said. “The other thing that's changed is that the context window … the context length has changed.”

Grabowski also talked about varying and diversifying prompt information. For example, he gave the example of a zero shot prompt, where you don’t have any context for a question, and then suggested that by changing up these systems, you can empower the LLM to work better in many ways.

“If we can save researchers time in chewing through 1000 papers … by providing meaningful summaries of those papers, does that accelerate the basic research?” Grabowski asked. “I think the answer to that is yes, too. I saw a really interesting … foundation model case where people took a language model framing, and applied it to atomic movements, like very low metal chemical interactions, and without ever having seen it before, when given sodium and chlorine atoms, the model correctly predicted the structure of salt crystals.”

Grabowski also talked about the dangers of jailbreaking, where hackers and bad actors may be able to take consumer products and make them do what their creators never intended them to do, bypassing key safety programming.

He also talked a bit about agents in AI, and what these agents will look like, what they will do, and how they will work together to solve problems. I’ve been covering contributions by various experts on the state of “agentic” AI and what that means for the advances that will happen in 2025 and beyond. It’s big news.

Liquid Network Research

When Maxime Labonne took the stage, he talked about different stages of LLM development with new liquid networks that can do more with fewer parameters. This is changing the math in the markets, and enabling companies to do a lot more in enterprise IT. Disclaimer: I have been consulting the team at the MIT CSAIL lab working on liquid networks, and the Liquid AI company that Labonne is working in.

In going over applied science here, Labonne talked about the questioning stage, in which engineering teams may apply their own processes to model use, and preference alignment, as well as the eventual fine-tuning of functions. By adding knowledge, he suggested, you can fine-tune these functions to give the system better power for accuracy with a diversity of samples.

“During supervised engineering,” he explained, “what we do is … we give instructions and answers to the model and input, and we ask the question, and we teach the model to answer the question, and we do it a lot. We teach the model a structure, a chat template that we're going to talk about later, … Then it's a model that is able to follow instructions, it's a model that is able to answer questions.”

As for preference alignment:

“During preference alignment, we give preferences to the model,” Labonne explained. “We give not only one question and one answer, we give two answers. One is the chosen answer. This is how we want the model to behave, and the other one is the rejected answer.”

He also talked about formats of inputs, like real life conversation as a data set, and why that’s so valuable and important. Speaking about tools for post-training, and automated benchmarks, eh also had some things to say about evaluation, and the role of human bias.

“Humans are incredibly biased, too,” he said. “We like to think that we are kind of the ultimate (resource for) evaluation, but we're really not.”

I appreciated both of these pros visiting the class: all of this illuminates what researchers are doing to make their work on LLM systems more meaningful, and to add value to those processes of bringing products and services to market.

So if you’re in enterprise, trying to figure out the context of AI right now, you can probably benefit from these ideas that are coming out of the classroom. I’ll bring you more as we move through the semester.