Blending AI Innovation with Industry Demands

Sheikh Shuvo:
Hi everyone. Welcome to Humans of AI. I'm your host, Sheikh Shuvo, and today we'll be learning about the people that are building the tech that's changing our world. Today I'm excited to share with you Helen Byrne, who is VP of Solution Architecture at Graphcore. Graphcore is a UK-based company that designs and manufactures IPUs, or Intelligence Processing Units, for AI applications. Hardware solutions that rapidly increase the efficiency of machine learning tasks. Thank you so much for joining me this morning, Helen.

Helen Byrne:
Thanks for having me. Excited to be here.

Sheikh Shuvo:
Yeah. Now, a very first question I have for you is, uh, how would you describe your role to a five-year-old?

Helen Byrne:
I find that question funny because I feel like I'm constantly trying to do this with my friends and family.

Sheikh Shuvo:
Friends and...

Helen Byrne:
Families and mothers. Yes, exactly. Exactly. So, Graphcore, I mean, you just explained what Graphcore does, but Graphcore has designed a brand new processor for AI. So, instead of using your laptop to run your AI model, you can use a processor that's been specially designed to run those algorithms efficiently.

And my role within Graphcore is to interface with the users of our processors. So I help our users to port their models and optimize their models to run well on Graphcore's hardware. And with our software, I'm kind of at the intersection of the hardware, software, and the models, the applications.

Sheikh Shuvo:
Yeah, cool. I'd say that that would be a very advanced five-year-old, but oh, good start. It's really...

Helen Byrne:
Hard to describe at that level. Yeah, absolutely. I need to keep working on that.

Sheikh Shuvo:
Yeah, well, along those lines, can you walk us through what a typical day in your life would be?

Helen Byrne:
Sure. So I guess it's different day to day, but I spend quite a lot of my time meeting, um, yeah, users of Graphcore technology. So whether that's people accessing IPUs in the cloud or whether they've bought IPUs that they're running from their data center, I work with them to first of all, understand what their model is. So, if they're running LLMs, what is the application that they're trying to run? I understand how best we can get their model running efficiently.

So yeah, I have to kind of get into the depths of what is the actual model architecture that they're running and what is the software that they're using. Then I try and help them. That's kind of more at the discussion level. I meet a lot of people and try to understand what their problems are, where their compute bottlenecks are, and then in terms of the coding side, application building side, we get, so a customer, an end user might share with us a proxy model that they're interested in, or an open-source model, like Llama 2, for example, and say, I want to run this model and I've got this data. Can you help me? Or they'll say, I've tried to run this and it's a bit slow. Can you help me speed it up? So then I spend my time just helping them to run their models efficiently, whether it's training or inference, whatever their problem is.

And yeah, I spend a small part of my time traveling. Well, a large part of my time traveling, speaking at conferences and events, kind of evangelizing using IPUs.

Sheikh Shuvo:
It sounds like you're a full-time detective most of the time.

Helen Byrne:
Yeah, yeah. I love that. Nice.

Sheikh Shuvo:
Awesome. Well, one really interesting thing I saw about your background is you actually started your career as a math teacher at a London nonprofit, Teach First. From those days, what motivated you to pursue your career path?

Helen Byrne:
Yeah, so I had quite a diverse background, maybe different from some more linear projections into this role. So yes, I did mathematics undergrad. I then didn't really know what I wanted to do. And there's a teaching program called Teach First that we have here that was actually built off an American program called Teach for America. And it's a way to get graduates into teaching who might not necessarily have thought about going into teaching, and it's at schools that have a need for teachers, they might have not great results, whatever that metric is, and they might be in an area of poverty. There's a few different metrics for what the school might look like. So I did that because I was kind of not really sure what I wanted to do, and it sounded interesting. And it was a two-year commitment. So I became a math teacher at a high school, secondary school, for two, well, I actually ended up doing three years.

And then a lot of, at the time when I did this, 50 percent of teachers that did this actually stayed on in teaching. So it's a good way of getting people into teaching basically. I didn't stay. I left, it was, I'm glad I did it. It was the hardest job I've ever done. I take my hat off to teachers. It's really a challenging job. A lot of pastoral care and everything else that goes with it. Anyway, so I kind of decided I didn't want to continue doing that forever and I wasn't sure what I wanted to do. I actually then went completely the other direction and went into a kind of banking job. Um, working at a FinTech in London, doing a kind of technical, quantitative role within a sales team there. And I enjoyed bits of that, but I actually decided I didn't really like the finance industry. It wasn't really for me. And then I left, yeah, I've done a bit of everything. I then did a master in AI. Because I was very interested, I decided I was interested in tech. So I'm going to go do this master, did this master.

Then when I worked in research in AI, I was really lucky with Graphcore. So I, I actually found, I wasn't really sure what I wanted to do at this point. And I met one of the co-founders here at Graphcore and a number of the people here at Graphcore. When I came to interview here, and it was the first time I just felt like they saw that the fact that I had all these diverse experiences was a positive rather than you're not really sure what you want to do. Like I, I'd seen that some people found it was a negative and for the first time I had this really great experience here where I met a lot of the people and I really remember those conversations and they, they were really happy that I had done all these weird diverse things and thought it would be a positive for me. So yeah, then I ended up joining Graphcore. I've been here now five and a half-ish years.

Sheikh Shuvo:
Nonlinear paths are always the best if it makes you feel any better. I started my career at a baby food company, so things have come full circle.

Helen Byrne:
I look at that in my team as well. I think it really adds value that we have people coming in from different backgrounds. They've done different things. They did different undergrads. It just adds value.

Sheikh Shuvo:
Well, along those lines, talking about the team that you have, outside of technical prowess, could you talk a little bit more about the types of things you look for, like coding things on the softer skills, any types of filtering questions you'd like to ask, things like that?

Helen Byrne:
So generally, yeah, diversity actually is really important. We have people in the team, and I look for different skills within the team, and that's at the technical level and otherwise. So, at the technical, kind of software level, we have C level experts, we have AI modeling level experts, we have systems engineering, and everything in between. It's great to have people with different skills on the technical side. In addition to that, I think being able to communicate is really important in our team because we end up speaking to the customers a lot. We also speak to a lot of the different teams. It's very important for us to be able to work with the core software developers and make sure that they know what the feedback is coming in, and what we think is the priority, what is the priority on different software features. We also need to work with the marketing team and make sure that we're putting out the right kind of message that the end users are going to see. And then we have to work with our product team. So the communication side is really quite important for our job, as well as being technically sound as well.

Sheikh Shuvo:
Absolutely. And with talking with so many developers on the evangelism side, there are obviously lots of companies and products competing for people's attention there. What are some of the communication strategies that you've found working with technical audiences that work better than others?

Helen Byrne:
It's so great when you can meet the developers face to face. You just get so much more information from them rather than trying to do something virtually. So if possible, go and attend events. I love research conferences. So I really like going to those and just meeting people, speaking to them, going to the workshops, and finding out what people are working on, what they're interested in, what are their problems. The face-to-face, nothing compares to the level of information you can get from meeting people face to face. We try to go to meetups. There's a lot of great meetups over here, and even more in the U.S. So yeah, any chance you can get. There's a lot of great virtual stuff online as well, but I don't find them as valuable. I think people don't give you exactly what they think necessarily, they sometimes hold back a bit, and you don't get quite as much.

Sheikh Shuvo:
Yeah. Do you have any particular conferences coming up the rest of the year that you recommend people check out?

Helen Byrne:
So NeurIPS is the biggest AI research conference, which happens in December every year. It's in New Orleans this year.

Sheikh Shuvo:
It's in New Orleans this year too. Yeah, should be a fun one.

Helen Byrne:
Yeah, exactly. Exactly. That's always great. It's big and it's a week, basically, if you go to the workshops as well, so it's kind of a lot, but the interactions that you have there are just really valuable. Obviously, you get to see all the research, you get to actually talk to people and ask them questions about it. You just get a good, kind of, overall high-level understanding, and you get the deeper conversations as well. So yeah, NeurIPS is probably the big one that's left this year.

Sheikh Shuvo:
On the research side, it seems like there are like a thousand groundbreaking papers published every week, right now. What are some of the habits you've developed to stay on top of things and find out which ones are actually important for your day to day?

Helen Byrne:
Yeah, good question. So I used to, when I worked in research, I used to go onto the arXiv website every day and just trawl through all of the new papers and check that I've not missed anything, and you still miss stuff. Now I just don't have the time to do that, especially because I'm not working in that space directly anymore. So, I use a lot of blogs and a lot of kind of Twitter accounts, even LinkedIn. I follow some people on Twitter, LinkedIn, who are right on the pulse of research, and blogs that I've found lots and lots of. For example, there's one called Davis, or Davis Berlok from Mosaic ML. They do a roundup of top papers. There's one called Alpha Signal, which is a roundup of top papers, and I just keep an eye on what people are talking about really. You also notice now it happens so quickly that a paper will come out and within a few months, that technique is now being used in all of the new models that are being published. So you can see that really quick transition from research to actually being used and being ubiquitous, something like MQM, like multi-query attention, grouped query attention, flash attention, these set of techniques. They came out and then they were just used very quickly, kind of absorbed into all future models that are released.

Sheikh Shuvo:
Now, for someone who's just finishing school right now, who has a master's in computer science with an AI focus, who's excited about exploring the world of AI but not sure where to start, what type of role or what type of company would you suggest they start with?

Helen Byrne:
So there are lots of amazing online free courses and things that they should probably have a look at as well, like Coursera and the others, and the Kaggle competitions, just to get, if they've not done machine learning or kind of built a machine learning application of some kind, they should probably have a go at that, practice that. Play around with that, play around with the datasets that are available, but yeah, there are some, I mean, there's a huge demand for that role. And so I think if you're willing to just take on an internship at any of the startups or larger companies that are trying to find ML engineers, just if you're starting out. If you haven't got lots of experience, you can go into a team and start on as a graduate or whatever the description of the role is, and you'll learn so much in a real environment in industry. I think you pick up a lot, it's great to do the courses, and I think it's good to have the foundation and have played around with things because you probably need that to get the job in the first place. You need to be able to show something, but then the amount that you learn actually building something real in industry, or maybe it's not in industry, maybe in academia, wherever you get that first job, it's like a huge step change in terms of how much you're actually getting from it. So I would just apply for a lot of those kinds of roles, and you'll learn a lot.

Sheikh Shuvo:
Well, looking at your own background there, when you made the jump from finishing up your research days to jumping into industry again, what were some of the things that surprised you, if any at all?

Helen Byrne:
I guess the interesting thing, and it's kind of good and bad, well, I enjoy it and I miss the research side, is you actually see what people are really doing, practically doing rather than playing around with, as kind of toys, toy models, which are actually working with people that are building things that are actually in production, or whatever. And sometimes there's quite a difference. There's a delta between this kind of research mode and real industry production. So that was kind of interesting. I missed not being able to spend a long time thinking about something and maybe too long trying things that didn't necessarily work because when you work on the industry side, you just, everything, you need to get it implemented and working, and there are shorter deadlines. But I find it really interesting to see how things are actually used, you know, rather than just thinking about them.

Sheikh Shuvo:
Yeah. The veil has been lifted. Yeah. Cool. And the very last question I have for you, Helen, is if any listeners want to get in touch with you, what's the best way to connect?

Helen Byrne:
I guess LinkedIn is probably the easiest way to connect with me. I'm on there. Awesome. Yeah. I'm always at the research events, so you can probably find me there too.

Sheikh Shuvo:
We will find you at NeurIPS then.

Helen Byrne:
Yeah, find me at NeurIPS. Cool. Well,

Sheikh Shuvo:
That's it for this episode. Thank you so much for joining us. This has been lots of practical and actionable advice.

Helen Byrne:
Thanks a lot. Thank you. I hope it was useful.