AI talk: Introducing a new Inside Angle AI expert Nathan Brake

This week I chatted with artificial intelligence (AI) and machine learning engineer and new Inside Angle blogger Nathan Brake to find out how he got started in the industry and his take on the latest generative AI models.

Welcome to Inside Angle, Nathan! I hear your passion for machine learning goes back from early days. Tell us more.

My journey into machine learning goes back to my childhood. My dad was always getting the latest computers, and my brother and I would get his hand-me-downs. I grew up with Windows XP then Vista. I would basically destroy the laptops by messing up the software – installing all kinds of open-source software widgets and things that cluttered up the computer memory and storage space. I was always super interested in customizing things and remember the first time I overwrote the windows operating system with Linux Ubuntu to try to figure out what was going on. I was more fascinated by the process of building and understanding systems rather than just getting the final product.

My parents, seeing that I was good at math and science, urged me take up electrical engineering. My approach to studies was focusing on efficiency. You might say I was already looking at learning curves and optimizing based on my time to figure out when the payoff starts to get low enough that it's not worth trying harder.

I got my degree and started working. I wasn't doing much software in that job, so I began learning it on the side. I first went through some free online courses and taught myself JavaScript and HTML and CSS to be like a web developer. When I started working at 3M, I worked with our speech recognition infrastructure, programming in C++ and JavaScript. I started my graduate program at Georgia Tech for computer science in machine learning. That’s when I transitioned to working with machine learning frameworks and mostly Python now. I love what I do, and it's been it's been a ton of fun to work on our team at the forefront of the computer science community.

You are indeed a research scientist! Since we're working on generative AI, I know you've been playing around a lot with GitGub CoPilot, and I also use it and it's awesome. Can you share some thoughts on why you find it so useful?

It’s been so truly like life changing in terms of efficiency. When you are writing code, you're using certain design patterns that you've been taught, like object-oriented programming. So, just being able to use a system that knows how to write good software and can figure out what design pattern you're trying to use and then can help you fill out the rest of that design pattern, that saves you time because you have to type less. But it also helps you write things in the architecture that everyone else is writing, things which makes it more readable for everyone else.

You’ve been a recent contributor to Inside Angle specifically about AI and LLMs. What got you interested in all this blogging?

I think I realized we do a lot of work internally that I sometimes don't document as much as I wish I did. It's not a research paper, but writing something down has really helped me understand it more and then have something to look back on. I've so far really been enjoying blogging. I think it will help some people and hopefully it does, but a lot of it is for my own sake to make sure I understand the ideas I’m working with.

I know that you've been playing with some of the open-source solutions out there like Llama2 and Mistral AI. Do we take an open-source solution and productize it or do we go with the closed solutions from Google and Open AI? What are your thoughts?

It is a tricky question, and I see good points on both sides of the debate.

I think it's important for there to be some open-source community to be able to provide further innovation. If we go close source, then we really risk stalling innovation because our field would have been so much slower if Google hadn't released its “Attention is All You Need” paper about the transformer design that basically revolutionized all of this. Then you have the debate about what's more secure. Security through obscurity or security through complete transparency. The most secure piece of software we have is arguably Linux because it's open source. It's used everywhere and it’s one of the most secure pieces of software because everything about it is out in the open. So, if there are any vulnerabilities, everybody has a chance to fix it. I think those are the big pros for open source.

But, closed source is an easier way of controlling what happens in the application layer. So, I heard a podcast that was talking about the idea controlling who has access to these huge models to train them is kind of like controlling who can refine uranium or enrich uranium. It's kind of like they're trying to get one step ahead by saying we could try to regulate on the application layer and say it's about what you use it for, but it's going to be so much easier if you can just prevent how many people are allowed to have it.

It’s been great to chat with you, and I’m looking forward to seeing more great AI thoughts and ideas in your upcoming blogs. We’re glad to have you on the Inside Angle team, Nathan!

 “Juggy” Jagannathan, PhD, is an AI evangelist with four decades of experience in AI and computer science research. 

Nathan Brake is a machine learning engineer and researcher at 3M Health Information Systems.