The possible uses for artificial intelligence are endless—and so are the headaches. Designing an AI application can mean running through hundreds of iterations, taking months of time.
And managing the scalability is an engineering task most smaller organizations don’t have the manpower for. As an employee at Facebook, William Falcon didn’t much need to worry about these things.
He had an army of engineers and enough computing power to simulate a universe. But what about the people who didn’t work for big tech?
The answer to this question eventually became Grid.ai, a startup William co-founded to give everyone the power to quickly run through hundreds of AI iterations while eliminating the need to manage infrastructure.
On this edition of UpTech Report, William tells us how the technology works and some fascinating use cases.
More information: https://www.grid.ai/
William Falcon is the creator of the popular open-source project PyTorch Lightning, and the recently announced Grid.ai William created Lightning while doing his PhD at NYU and as a PhD researcher at Facebook AI; Lightning allows users to scale models without the boilerplate and Grid enables large-scale training on the cloud.
Previously he co-founded the now acquired NextGenVest and spent time at Goldman Sachs.show more
His PhD (currently on leave to focus on Lightning and Grid), is funded by Google Deepmind and NSF Foundation. His research interest is in unsupervised learning and the intersection of AI and neuroscience.
William is a native of Venezuela and holds a BA from Columbia University in Computer Science and Statistics, with a minor in Math.show less
Video Transcription: AI Power to the People | William Falcon from Grid.AI
DISCLAIMER: Below is an AI generated transcript. There could be a few typos but it should be at least 90% accurate. Watch video or listen to the podcast for the full experience!
William Falcon 0:00
If you think about solving a problem with AI, you can think about like, basically, you’re gonna have to try like 50 or 100 things. And, you know, at the end of that you’re gonna find something that works. So either you do it sequentially, or you do it all at once.
Alexander Ferguson 0:18
Welcome to UpTech Report. This is our applied tech series UpTech Report is sponsored by TeraLeap. Learn how to leverage the power of video at Teraleap.io. Today, I’m joined by my guest, William Falcon, who’s based in New York, he’s the CEO and co founder at Grid AI. He’s also the creator of pytorch. Lightning, very excited to have you on. Yeah, thank
William Falcon 0:38
you for having me, Alex, very excited to be on.
Alexander Ferguson 0:40
Now, Grid AI, your focus there is enabling companies of all sizes to train, we call state of the art AI models on hundreds of cloud GPUs and CPUs just right from their laptops. But you’re also pretty well known in the community around pytorch, lightning, being the creator of that help you understand, like, what’s the journey that you’ve been on here, with first pytorch lightning community created around that, and now with grid AI, the problem that you set out to solve? You know,
William Falcon 1:08
when I, when I started doing deep learning research, it was around 2016, I want to say, and, you know, had been a software engineers for a few years, and I was coming to the AI world. And it was really interesting how there was not a lot of standards, like there weren’t best practices. And it’s still true today, right? It’s everyone’s still developing the their best way of doing something. And this is a new technology. So we’re all trying to figure it out, right? I mean, we are as well, right?
Alexander Ferguson 1:36
And it’s always changing. Yeah,
William Falcon 1:38
yeah. And so what I started doing was, I wanted to figure out how to move quickly through research right there in research, you have a lot of ideas and in the iterate on those ideas, and that process is not really that different at a company, right at a company, you have data, and you have some hypotheses as to what’s going to give you the most value, but you’re gonna have to try a bunch of them until something does give you a lot of value. So it’s the same process. It’s like iteration process, right. And I realized that I was having to copy code over and over again, and it was really not scalable. And I wasn’t focused on scaling, obviously, because I was just doing research. But, you know, if you think about solving a problem with AI, you can think about like, basically, you’re going to have to try like 50 or 100 things. And you know, at the end of that, you’re going to find something that works. So either you do it sequentially, or you do it all at once. And some of those things, you’re not going to know until you try something and then that unlocks a different thing, right? So it’s kind of this tree of ideas. So I basically wanted wanted to get a lot of this done through the lightning. And so I started working on this kind of internal project for myself. And it was really useful. I put it aside for a few years and then kind of went into the startup world and, you know, started putting deep learning to production. And then when I came back to do my PhD, I was like, let’s dusting off, and let’s see if I can take a stab at it again. And yeah, it took I mean, a second, three or four years at this point to get that right. But I think we arrived. And you know, so weakness is really a community project now and it’s, you know, overflowing or people who are working on it, we arrived at something that helps you move really fast through your research ideas, and then put it into production very quickly without doing a lot, right. Because before you had this bridge between the research group, and then the production teams, and that will take like six months or a year to put anything to production. lightnings have the light. And you know, lightning helps you do that process. But it doesn’t help you when you’re doing it at scale. So lightning can help scale up your models, but you still have to do a lot of infrastructure work on your own if you want to do that. So when I was on Facebook, you know, all of this is already done for you a lot of it right. And so the machine through there, and all these different things, so I don’t have to worry about it. But I realized that this was a problem that not you know, if you’re not a Google or Facebook, I don’t know how you’re going to solve these problems, right?
Alexander Ferguson 3:52
Like the big companies, yeah, they got the infrastructure to get the people. But outside of that, we’re getting to the point where machine learning deep learning, there’s a lot of opportunity, and we it shouldn’t be reserved just for the folks that happened to have the giant infrastructure, it should be able to accessible and as it sounds like that was your driving thought and feeling like I gotta find a way to solve this.
William Falcon 4:11
Exactly. I mean, you have to think about it. Like I didn’t start in computer science or deep learning. I started, you know, the military before. And then I, when I started research was really neuroscience. So there are a lot of scientists out there. And you know, these are people like data scientists, right. So at a company, what will happen is you hire data scientists, most of the time that person came from a PhD or master’s degree, and they were focused on math or physics or something, not not engineering. And now you’re telling him to go scale up models, and they’re like, I don’t, that’s not my thing, right?
Alexander Ferguson 4:42
It’s, we’re getting to the point where, where the use case of using machine learning deep learning you, those are two separate roles. You don’t know they don’t need to be an expert in infrastructure to be able to run their models, or they shouldn’t be.
William Falcon 4:55
Exactly and so this term ml ops came around, right. And so I’m sure you hear it in your company and I think that, you know, like I’ve said this before, but I think if we’re doing our jobs, right, that should just disappear. Like, if you use lightning today, you don’t have to worry about a lot like you just focus on what you’re trying to solve in the data. With grid when you merge with grid, now, you don’t really have to worry about infrastructure, or how you can scale things up, or how you’re, you know, making sure that your data is not going to bottleneck your training there, all these new insights and things that we know, because we spent a lot of time obsessing over these details. So we know how to get them, right. Which means that the data scientists and machine learning engineer, they’re left to just kind of focus on the data and the problems that they want to solve. So they don’t have to deal with any of this other stuff. So you don’t need to go find unicorns for like, expert engineers, and data scientists and mathematicians and all this just get someone who’s really good quantitatively, like who can, you know, look at that, and really drive a lot of insights. And, and solve that problem. Like, this is what humans are best at. Right?
Alexander Ferguson 5:53
So pipelining, that’s really focused around the the, the framework so that it can run well. But for those that need help with infrastructure, that’s where your focus of grid AI is that they can just scale up turn on and off as the as your website states to be able to train on hundreds of, of cloud GPUs and CPUs right from their, their laptop, just simplifying that process.
William Falcon 6:15
Yeah, exactly. And this is something that I think a lot of people are not used to today, right? They are, they they like don’t have this. It’s interesting, because the world of machine learning and deep learning, like they’re not really using these resources as much as they should, or, or know how to, it’s because it’s really hard, right? So before pipe, torch lightning, you know, people like even within Facebook, like we had some groups who would train things on like 3264 GPUs and like, they were very specialized at that, that ability is now democratize through lightning. So you can just do that, and you’d have to be an expert. So it’s the same thing for or grid AI, right is, you don’t no longer need to care about the cluster or care about your ml ops or any of this stuff. If you want it, it’s there, and you can access it, but you don’t need to worry about it, right? If you just need to spin up, you know, 20 models, each 132 GPUs, you can do that. If you need high access data, you can do that.
Alexander Ferguson 7:08
You mentioned that if they want you they can so it’s still kind of an open framework, so they can look at it, or is it? Is it just kind of close? Like how is that
William Falcon 7:16
every you know, we’re building things for professionals, right. And something that’s very, it’s lightning from the other frameworks is that you like it makes it very simple, like the API and the way that you interact with it is very, very simple. But if you’re an expert user, and you need a lot of control, you can always get it right. So it’s not a black box, it’s not a kind of at least auto ml things where like, you know, you don’t know what’s going on, you can actually have a lot of full control. And that’s what great AI is about as well as that’s the philosophy that we bring it. If you don’t want the control, don’t worry about it. But if you do want it, you can you can get it.
Alexander Ferguson 7:48
You guys been around just a little over a year. So it’s like a lot of happening for great AI, though. It’s been two years from pytorch. Lightning when that would that began. Right. Right. Yeah.
William Falcon 7:57
And I think, you know, the majority of people on the team have been working on deep learning or AI related things for at least four or five years, maybe longer some people so we’ve all, you know, been doing things. Well, I’ve been doing this in other lives as well.
Alexander Ferguson 8:13
It’s like all culminating and now now it’s like be able to move forward. What can you speak to that, as far as so far the use cases and folks that are on there, the types of organizations and businesses that can use it and should be using it?
William Falcon 8:26
Yeah, it really ranges. And I do want to say that, you know, grid is not just for lightning, you can use other frameworks, right? So it’s not when you get a much better experience with use lightning, but you can use things like TensorFlow and pytorch, and so on. So we’re not we’re not restricting that. And I say that because the use cases are very broad, right? So we’re seeing e commerce, we’re seeing gas and oil we’re seeing, like healthcare, finance. And it’s interesting, because the lightning community is huge, right? I mean, they’re, I don’t know, easily 50,000 people across the world using lightning 1000s of companies. And so these are just people kind of coming in and saying, hey, like, I liked how you’re solving this problem. So can we you know, if you can do the same for me labs, please go for it. Right. So you know, what’s important about grid is that there are many ways that you can, you can deploy the platform. So if you care about data, privacy and everything else, then obviously there the uptime deployments, we can orchestrate things in your own VPC. So, we’re very cautious about that, right. I mean, I came from finance. I, you know, worked in healthcare for a bit. So like, I understand all these nuances as well. And, and if you don’t, like I mean, we, you know, grid never looks at that or anything like that. But there’s a way where you can basically have us coordinated for you, like all the infrastructure, you still leave metadata coming out of the cluster, but it’s just about what machines are running and what’s what’s down. But nothing ever about code or data. So it is very, very secure. So, you know, HIPAA and all these, you know, finance things that you need to do as long So you can do it on your own cluster or on prem, like we handle that.
Alexander Ferguson 10:04
Now, you actually had some experience yourself training miles, if I understand it correctly around some of the COVID. Ai and COVID, can you can you can you share about that?
William Falcon 10:13
Yeah. So, you know, I, you know, I actually do research at NYU as my advisors. And when COVID came out, you know, we had this idea to figure out if we could basically take models and have people like, basically call in, and then you know, call for speak. And if we could take, basically record them right, and then take their sound waves and try to figure out if we could act, if there was if that COVID. Right. And so, yeah, so it’s an experiment, right. And so we collected data, basically, we built a system, where people would call in, and then we actually ended up collecting a lot of data. And what ended up happening was that we wanted to use models for that. Now, there are a lot of, we were able to detect hints of, hey, you’re sick, or you do have something, but it wasn’t foolproof. And there are I mean, with medicine, specifically, there are a lot of issues. And so I’m a little bit skeptical now that you can do this. And there are like actual tech limitations for For example, we found that you know, your your phone records things in certain bandwidths, right. So in the higher pitches may be dropped by this by the signal by the cell carrier. And that’s where that’s where you could register some of these, like, you know, wheezing, noises, and so on. So it’s not 100% foolproof. I think if we can dig into it, maybe we can get somewhere. But yeah, that was a really interesting experiment.
Alexander Ferguson 11:38
Yeah. So there’s, there’s so much opportunity, but it’s so much of an experiment, experimentation and explore, which is the, I guess, the main reason why you built this platform is there just needs to be more, there needs to be more opportunity, more exploring more testing, simultaneously, and not waiting forever to do one run and then run another one. But there’s, there’s a lot of hype around AI and what it’s capable of, and people can seem to say, well, it’s just gonna solve everything. And it’s easy and just just done, you’re good. Well, what do you what do you say to that?
William Falcon 12:10
Yeah, it’s a lot of hype. I mean, so I’m a researcher, right? So I live in the I live in brief, the academic side, the the papers, and I know where we’re headed. And, you know, a lot of these things work on very small data sets. And when you take them to the real world, they don’t scale, right. That’s not to say that they don’t work. Well, in certain cases, basically, like, if there’s anything where it’s like a natural signal, so like speed, or images, or videos, or text, AI will, will do much better than kind of more traditional methods, right. And it’s just a little bit more nuanced the way that it can kind of interpret that information and extract features from it. But I think that what happens then is the media or, you know, other companies take that and then they just like extrapolate it, of what it can do. And then then they’ll show you one specific example where works really well. But, you know, that’s the tip of the iceberg. And then you see around that, and it doesn’t work. And so I think what burns companies really is, they’ll buy the hype, and then they’ll go build a model. And then, you know, work on like, their, you know, dummy data set or something small, and then they put it into production, and then they get in trouble. Right? So that’s, yeah.
Alexander Ferguson 13:21
So what what would you then as a true data scientist and research yourself, but you’re, you’re providing a platform for more people do testing? What would you be saying to business leaders, where they’re seeing the hype? But there’s, there’s still truth underneath the hype, right? There’s still truth there. So what would you want to say to business leaders say, well, we want to use it, but how do we balance that not create too much hype around?
William Falcon 13:46
I think just temper expectations, right, I think it’s good to start with something simple and then build up, right, you have to use deep learning. So when we talk about AI, we’re talking about mostly deep learning. But you’re there other models, you can use stats, you know, if you get into the math, you can be configured, we consider deep learning or not doesn’t matter. But just get into something basic baseline that and then go from there, right, don’t jump into the most complicated models. So I think it’s, you can de risk a lot of the hype by doing that process. What What is interesting about lightning and grid in the community is that we’re all at the heart of this. So what is really helpful for everyone is that we actually can help in that way, right? Like you can join our slack group, you can join our conversations and like we will actually, you know, gut check you and be like, that’s not gonna work. And lightning and grid are built built around a lot of these best practices to keep you from making mistakes, right? Like I see people coming up and saying, okay, I want to you know, there’s there’s something called spot instances on Amazon where you can get like, you know, 90% discount on your compute, which is great. And we use that a lot of grid. But if you start training on multi nodes, for example, you could actually throw off the model, but it’s subtle, because it’s how the models learn and the gradients update. So it’s very, like mathy I guess, if you’re starting And think about it. So we kind of just kind of call it out. And we’re like, Look, we’re not gonna enable certain functionality, because we know it’s bad for you, right. And so we blocked that off. And the same for lightning. So it’s not just about structure speed, it’s also about giving you kind of best practices embedded. And it’s from hundreds of people across the world who are experts at this, right? Who are all the top companies. So our knowledge is getting embedded in there. So as a company, now, you don’t need the expert researchers, you just need people who understand the problem really well. And then they can leverage our technology and our collective expertise to kind of do a much better job.
Alexander Ferguson 15:34
And I apologize, if I am i i’m not perceiving this properly. But it’s kind of sounds like we’re getting to an age where you can build websites and HTML, and you don’t really need someone who’s amazing developer knows all bunch of code, they can just create a website because all that the contents there, we’re getting to that space now with with ml and deep learning, because the algorithms that are already proven and in different places, as long as you know which ones to use and the community around there, you can search for finding answers. You don’t need a really deep research to build new algorithms or anything. Am I getting that? Correct?
William Falcon 16:08
Yeah, I think I think of cars like, basically think about, we just invented the car AI, right. And for the first few years, every single person was either building their own car, or like learning about physics and engine mechanics, before they drove a car, you know, like today, you’re gonna need to know one need to know this, you just like buy a car, use it, right. And the car manufacturers are gonna embed all the best things in there. So you’re not going to be able to go like 2000 miles an hour, you know, 300, whatever, it is something unsafe, you’re not going to be able to do turns like crazy speeds. So they they’re going to build in a lot of these safety mechanisms as well to keep you from making mistakes. So it’s a lot of what we’re doing as well. The difference though, is that now you sometimes you do need to be an expert to dig into something. So we’ll give you an escape hatch, basically, to say, Okay, well, like a car mechanic, that’s fine, take it to the mechanic and don’t know exactly how to look in the engine
Alexander Ferguson 16:58
of machine learning mechanic, I love that car mechanic to, hey, my models not working, can you take a look that wheels coming off, it’s great. But but
William Falcon 17:06
that’s how you scale it up, right? Like, not everyone in the world can can be a car mechanic, like people just need to be able to drive it and use it. And then when they really need something special, then they can take it to those people. And that’s the community. That’s the lightning community. That’s the research community that you can tap into when you need it. But it also means you have to hire experts, because they’re amazing people out there who are coming from biology, who are coming from chemistry, who are physicists, political scientists, who know the math, they know the research process and know how to solve problems. But they’re not expert engineers. And they don’t need to be like deep learning experts either. Right?
Alexander Ferguson 17:38
So for, for a business leader to say, there’s there’s just so much energy we’re going to be left behind, we need us to, to really embed and have more opportunity for running different new machine learning algorithms, deep learning algorithms,
William Falcon 17:53
you’re also not going to, you’re not going to compete with like the big companies or like think about it like hiring, would love hiring against Facebook or Google for AI talent, right? Like, a lot of the top, you know, PhD students and researchers that are coming out, are going to these places, because that’s, you know, they’re getting those options. And so they’re kind of sucking that talent drive, but it doesn’t mean that you shouldn’t have access to it.
Alexander Ferguson 18:15
So what’s the typical role then, or title of someone that can be running these algorithms and could be using a platform like red?
William Falcon 18:24
So data scientists, machine learning engineer, research engineer, research scientists, you know, I think people call these so many different names. But I guess I’ll just say, if you have data, and you want to use a deep learning model, or machine learning models to get value out of that, then this is what you use.
Alexander Ferguson 18:42
What What would you say is for all those different roles or titles that you just gave? What qualities do they do they have, or either someone who’s wanting to play that role, something they should be building themselves, or a business leader that’s trying to hire someone to be able to run one of these, what makes up a good, good person like this?
William Falcon 19:04
I think at the end of the day, it’s having a strong quantitative background, to be able to, you know, speak intelligently about that, about insights. And about some of the math I mean, in industry, you’re not getting so deep into the math and finance, you might be in research, you prepare to spend a lot of your time in the math, but in the three, you’re just kind of the high level. So it’s really someone who can understand it at a high level and kind of read through it. Right. I think that’s the word data scientists is very abuse in the world, right? So there are people who are spending their data scientists who spend most of their time on business insights, like what’s my turn, you know, what are my users doing that kind of stuff for that you don’t really need AI. I mean, you can use models to help you but your job is more analytics, right, your analysis. So it’s this like grid, it’s less for those people. There are certain data scientists who are then sitting there and Modeling and there may be building recommender systems, maybe they’re building, maybe some of the models that the analysts will use to do something, right. So in finance, it might be stock prediction, it might be, you know, demand prediction. I mean, you can everything that I think anything that has a bunch of rules can be kind of rewritten with AI. Right? So If This Then That, then that. So predicting, you know, like fuel loads for planes, I mean, you name it right? You can, it’s basically how creative you can get. So so that’s what Grid is focus on right? Now, there, that’s a problem solving pace. And then today, there are machine learning engineers or research engineers would sit there and like, optimize models to work in near real time, right. So if you’re, for example, at a news agency, and you’re getting recommendations to have an interesting problem, because a new story will break now, and they have to retrain a model within a few seconds, and then give you a recommendation for the new story, because there’s a time element to it, right? Whereas Netflix doesn’t have that problem, because movies come out, they can return models for days, and then it’s fine, right? So there, you need to be really good at engineering. And so that’s really where Grid can help you do that. And if you’re like working with Grid and lightning, then you know, we’re going to give you basically the state of the art and how you can train the fastest, the quickest. And I think and get those solutions fast. So and then there’s the other side, I think, maybe some banks will have this, like more researchers. So like Facebook and Google have research scientists, and these people are sometimes academic focused, or literally just publishing papers. You know, in banks, it’ll be like analysts who are publishing reports about companies, and there’ll be people who are doing this for finance as well. I remember like Barclays had a strong, you know, group about this, they were like doing research into how do you do certain trading strategies and so on. And, and then there are, like the other kind of researchers who are more applied. So they’re taking the papers that are being published and trying to figure out how to apply them to do what they’re doing, right. So you could take a paper for speech generation, or let’s say, speech detection and use it for finance, because it’s just a time series at the end of the day.
Alexander Ferguson 22:03
Got it. Okay. So if I come back and try to recap again, the folks that the best can use the platform of researchers, engineers, developers of AI ml space, that’s if I got that correct as the best ones, they could use the platform.
William Falcon 22:16
Right? Yeah. And hopefully, you know, the longer that we continue to build our doing and execute on our on our vision, the less of an extra you’ll have to be
Alexander Ferguson 22:27
i get i love i love the analogy of the car concept is like, eventually, anyone be able to just run it because you built the safe guards in there. And then when you need a mechanic, you can you can run to them. Exactly the roadmap of where you guys are headed, anything you can share of features that are going to be added or something exciting that you guys are working on that you can feel comfortable sharing here.
Unknown Speaker 22:53
Yeah, of course. I
William Falcon 22:54
mean, yeah, we know the space really well, we know the problems that happened, you know, from training to inference to put some friends to production to all the challenges in between, right? I mean, we have people who have done this for many years. So we’re well aware of the challenges. If we’re not tackling them right now. It’s because we want to get one part, right, and then we’ll kind of move on to the other areas. Right. But I think what’s important is that, you know, I don’t know, I see, there are a lot of companies out there that are focused on machine learning and AI, and you know, a lot of amazing platforms. And I really think that the problems are so hard that, you know, we’re all going to have to collaborate at some point. So actually, you know, My take is really like, I want to make sure that we’re bringing in the best in class around to do all work together to resolving this, right? Because it is a massive problem that that is that can help a lot of people who can figure it out. So today, we’re focused on training, right? How do we scale up your training? You know, there’s also how do we help you do production and inference much faster, right. So that’s kind of the next phase where we’re starting to get into as well. And, and that’s lightning, as well. So lightning for the first year was focused on? How do we make it really easy for researchers and creators to build fast? And now we’re focused on for the last year Really? Well, since September, I guess, we’ve been focused on Okay, now you created how to bring that into production as fast as possible and scalable. And and so you’d have to worry about it. Right? I actually think that all of the production stuff, it can be automated, right? I think that’s the research process is the part that counts. Like, this is where humans come in. And this is what they’re best at. Right? Creating ideas and going through that. But we’ve been deploying things forever, like we’ve been, you know, how long have you been deploying websites for like 1020 years ago? So it’s not it’s a much harder, you just need to know a little bit more about machine learning. So if we do our jobs, right, like, you shouldn’t have to worry about that stuff.
Alexander Ferguson 24:40
People can focus on where people are best utilized, and automation technology can, can, can run the best I can can run the rest. For for those out there that whether the researchers engineers or developers in this in AI and machine learning. Are there any resources that you would recommend for continuing growing and learning themselves and what they’re they’re trying to go in, whether it’s books or podcasts or places that you read sites, anything that you would recommend.
William Falcon 25:08
Yeah, so there are a few classes. I mean, if you’re getting into deep learning, I think her path this class, I mean, it’s a very popular class from Stanford. It’s a really good one. I don’t remember the name of it, but it’s it’s pretty on YouTube. So one of my advisors Yon, it’s a steep learning class at NYU with Alfredo as well. So it’s, it’s actually I think, the first year that’s made been made publicly available. And you know, he’s one of the creators of deep learning and founder. So I, it’s really cool to see how he thinks about stuff. So I would check that class out as well. And then I do think that there are a few Coursera classes that are interesting, but I would also work on the books, right. So there’s, you know, depends on Matthew. And again, I think the deep learning book is a good start by yoshua bengio. And Aaron kohrville, and Ian Goodfellow. So it’s a great, great place. And then you know, how deep you want to get into the journey. I think there’s, you know, Kevin Murphy’s book, learning as well. And what are other good resources, I mean, we have a lot of tutorials, as well. So we do a lot of teaching, you know, the documentation and lightning specifically, like teachers, a lot of these concepts anyways. And through Grid, we do we do a lot of these ourselves. So I think we are very focused on helping people, you know, 10, next, their, their skill set, and something that like I’m personally passionate about, because I’ve, you know, I’ve been a data scientist, I’ve been researcher, and I’ve done this thing says, like, I don’t ever want to replace these folks, because I think that my goal is really to empower them. So I, and we see this today with with lightning or CSS Grid is, I want to make people 10x better, right. So if you’re a data scientist, or machine learning engineer, like using the stuff that we’re building, is going to kind of help you become much better at your job, right. So I think that’s where we’re gonna crush it as humanity is where you can take, you know, people who are brilliant, and augment them with machine learning or these tools and let them do their jobs better, right? I’m not a big fan of these kind of like, auto ml blackbox things because they just, they remove what humans are best at. and solving problems. It’s not just a function of looping over data. And then like getting the best results. There’s more to it than that.
Alexander Ferguson 27:21
I appreciate your distinction of where humans play and play best plus a role here. I want to kind of close here on a question of just looking ahead. predictions, tech predictions that you make, would make in the near term, what do you see happening next year, two and longer the 510 years from now.
William Falcon 27:39
So a few years ago, we had bird was like a big model and deep learning. And you know, that introduced transformers, this type of neural network. And then, you know, you saw open and GPT, GPT, two and C three, and so on. And under the hood, for bird and former, you have this kind of like ability to learn about labels where you know, blocks parts of a sentence out and then uses that as a straining signal. And a lot of what I’ve been researching, or last year or so, and kind of what a lot of the lab at NYU and Facebook are focused on is something called self supervised learning. And obviously, I’m I’m doing that research, which I think is interesting. But I think it’s also probably the most promising today because that’s going to enable a lot of different modalities. So like computer vision, images, videos, those kind of things, to work without labels, right. And so I think that the golden thing that will happen over the next few years is that we’ll be able to train models without having to label the data. Now, I’m not saying you won’t have to label some of it, but it won’t be the huge efforts that we have today. And you already seen the proof of that in the GPT models and the transformer models, right. So now think about how that like what will happen when we bring that into computer vision, audio videos, and so on. So I think that’ll be probably the biggest change for company and her, you know, everyone out there. And it also means that you’ll need fewer that annotation, you’ll be able to spend more time on just getting more data and then it’ll tap into that other you may not be using today because it’s not labeled. Right. So that’s one and then I guess outside of AI I mean, I think
Unknown Speaker 29:14
Yeah, I don’t know, you know, I
William Falcon 29:15
always love FinTech. So I think there’s gonna be I like Bitcoin and crypto and all these things, I think we need to do a lot of work there on making, you know, making sure that we’re not like damaging systems and platforms with minors. So you know, we’ve seen this, you know, GitHub, I’ve seen this a lot of other companies where miners exploit your computer to basically do something so they’re kind of making a lot of the free offerings out there, you know, painful for people to support.
Unknown Speaker 29:43
William Falcon 29:44
I think I think crypto if we can find a way to make it a little bit more, I don’t know less resource intensive, that’ll
Unknown Speaker 29:51
be super helpful.
Alexander Ferguson 29:54
Now making that happen, that’s, that’s, that’s a whole nother thing. But that’s ideally the direction I I appreciate the the future that you’re that you’re painting. And the platform that you guys have built to help with research itself, in some ways, and democratizing. And so even though you don’t have to be the big giant tech giants to be able to run more, more algorithms, thank you so much for your time. For those that want to learn more, you can go over to grid.ai. And it looks like you can even give a an individual level with the free credits right, a free, free tier and then a next year in the enterprise tier. Is that how you guys work?
William Falcon 30:31
Yeah. So you can come on the free tier and there, you’re basically going to get some credits. And then, you know, that’s really meant for you to go wild and do what you want to do there. That for us, we’re not making money. And that’s here, we just want you to like try the platform and do what you need to do. And yeah, so please, you know, check it out.
Alexander Ferguson 30:50
And get started. Thank you again, so much when everyone go to UpTech report.com. again to see these full interview and a lot more, and we’ll see you on the next episode at UpTech report.com. That concludes the audio version of this episode. To see the original and more visit our UpTech Report YouTube channel. If you know a tech company, we should interview you can nominate them at UpTech report.com. Or if you just prefer to listen, make sure you’re subscribed to this series on Apple podcasts, Spotify or your favorite podcasting app.