Hello and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. We talk a lot on this podcast about the fun, amazing, and sometimes legitimately world-changing things people are doing with AI.
Today, we're going to go a bit meta, if you will, and talk about how people work with AI and how one company in particular is on a mission to make working with AI and deep learning more accessible to everyone.
My guest today is Sirkan Piantino. He's co-founder and CEO of Spell.
That's spell.run, if you're keeping tabs at home.
Spell is a collaborative platform that lets anyone run machine learning experiments.
The company connects clients with computing clusters built on some of the best hardware in the world, and gives them the software interface they need to do their AI work on those clusters.
Here to talk about what Spell does and how they do it, This is Sirkan Piantino.
Sirkan, thanks for coming on the NVIDIA AI podcast.
Absolutely. Thanks for having me. So at the beginning of this year, 2009, Spell closed a $15 million Series A round of funding.
I'm going to read a quote of yours from a TechCrunch article that ran just after that announcement, because I really like the way it frames your mission.
We spent decades getting to a laptop powerful enough to develop a mobile app or a website, but we're still struggling with the things we develop in AI that we haven't struggled with since the 70s.
Before PCs existed, the computers filled the whole room at a university or NASA, and people used terminals to log into a single mainframe.
It's why Unix was invented. And that's kind of what AI needs right now.
So I know it's been nine months and that's forever in the tech world, but can you kind of leap off from there Tell us what Spell is all about, how developers work with you, and then maybe from there we can get into all the work you've been doing since then.
Spell is really about treating the process of building things with machine learning and AI as fundamentally just a different type of software engineering.
We think about two different ways that you can tell a computer how to do something that you want it to do.
One way, and we've been doing this since the 70s, is programming it directly, translating what you wanted to code that the computer can understand in executing that code and you know we've got decades and decades of experience and tooling and people building things and other people building on top of those things that now make that process really easy, accessible, I think, there's something like 20 million people that do this for a living every day.
Then we think about this new way of telling computers what to do, which is well, you can't directly instruct them, but if you show them enough data, and examples of what you want them to do.
And you give them some very sophisticated algorithms that can interpret that data.
And you do this a lot. on very powerful hardware a whole bunch of times, then the computer can actually learn to do another set of tasks.
Whereas all of the novelty, all of the things that we have now rely on and all the value that we get as a society out of software.
The first sort of 40 years were based on this first paradigm of programming computers and getting them to do what we want.
And I guess a good way to think about the next 20 or 30 years is that all the novelty, all the new things that we see are going to be based on the second paradigm of innovation. learning algorithms, showing examples, and that being the sort of new way that you can create delight and novelty and value with technology.
We thought about that a lot. I thought about that a lot.
One of our key insights was that When a software engineer goes from writing software on a laptop and using a simulator or deploying a website or something to working with AI, what do they need? one of our first answers was this really flexible ability to coordinate large amounts of compute, large amounts of data, and mash them together in flexible ways and particularly doing that alongside what they're already used to, where they're already used to writing code, how they're already used to running it.
So we're really about extending software engineering, all the things that we've learned over 40 years to then... be able to efficiently be used to build things with machine learning and artificial intelligence.
So let's paint a picture for the listeners maybe.
How would an individual developer go about, you know, let's say they heard about Spell.
They go to your website. What are they going to learn?
What are they going to see that your value prop is to them?
And how is that same and different from how they might already be conducting their AI experiments?
Yeah, it's super simple. And this is something we worked really hard on is keeping it this simple.
I always like to tell my team we're in the make things easy business of our posters.
The basic tool we have is a command line where if you type spell run in front of anything you would have run on your laptop, we will then recreate your environment, we will snapshot the code, we will find the data that you need.
We will basically move that computation onto any piece of hardware that we can find in the Cloud.
You have a lot of different ways of controlling how it's run, but the basic one is you tell us what type of hardware you need, and we take your process from your local machine and run it on the Cloud.
That happens in this seamless way where you still see logs, you know you kind of tactile the tactile experience is just like running things locally but you get all these benefits of additional hardware you don't have to move data around Everything you run becomes this scientific, reproducible experiment, which later, as you've run multiple things, you're beginning to sort of build up this. data set of keeping everything organized.
How long has the company been in business?
We've been going about two and a half years.
So in those two and a half years, have you gotten... a sense, and obviously these things evolve over time, but a sense of who your core customer or maybe profiles of a few of those customers might look like?
Yeah, definitely. It's been very positive to see that we are touching companies and applications across basically every industry.
There are a couple that are quicker to adopt technologies and AI, but it really is very broad.
I mean, we can sit down with, a company in healthcare, we can sit down with a retailer, we can sit down with a hedge fund, We can sit down with a startup or somebody in research and academia, and we hear the same types of things that they need help with. running Jupyter notebooks in the cloud, managing data, keeping it versioned, organizing experiments, getting access to a wide variety of different types of compute,
All of that is pretty inherent to the craft of writing software with AI and ML.
So we found that the industry is pretty broad.
I would say the second thing that I found is my journey is going from a research institution, which is really about pushing forward the cutting edge of both technology and what we understand about the science to then industry and seeing how research and industry are different has been really sort of illuminating and they really are different different different goals so that's one thing that we're finding we get a real great window into How are things that may have only been discovered a couple of years ago showing up in products being used, for real world benefit and where is sort of the world actually using AI?
So I've got a list of questions in my head to ask you about what you just said, but let's start at the end there.
How is the world using AI and how has that maybe changed, you know, even just in this past couple of years, or we're going to get into your background a little more in a minute, But kind of in the work you've done, how have you seen that evolve?
And maybe more interestingly, how do you see it kind of spreading out more going forward?
Yeah, so I have this sort of framework I try to give people about what AI is.
Actually, for probably about the first year of the company, we banned using the word AI.
Right. just because we felt it caused a lot of confusion.
But generally, I think about three sort of trends in the world.
One is... the research world, which is quickly conquering previously open problems, things we couldn't get computers to do. we've been able to get computers to do very rapidly.
That's just over just the past 10 years.
The second is this cultural phenomenon where we're talking about AI.
We're excited about what computers might do in the future.
It's entered the public vernacular. What's good about that is it's starting these conversations that we need to have about data integrity, privacy, bias, ethics. all of these things that are going to be real problems that we have to deal with in society, that's starting to happen.
And then the third is AI actually going from not things that are being researched right now, but things that we already know how to do.
How does that then show up in products and the real world?
How do we build things with that? And That third sort of arc is just getting started.
Right. I think that would be my summary.
Companies are playing with deep learning.
They're finding some applications, I would say, particularly NLP or natural language understanding that is useful. showing up a lot across a lot of different industries, but we're definitely earlier on that third trend than we were on the first two.
Now, does your platform also allow for companies who develop products to sort of host those products or host a portion of those products to run off of the cloud compute environments that you provide, or is it sort of for building purposes only?
No, we have a model server infrastructure.
We have a hyperparameter optimizer. We have a lot of One of our strategies is to make sure that we have the full end-to-end suite of tools. on the same data and execution infrastructure and that's because it is early and so what companies mostly need is they need one tool that can do all of these different things.
Right. It sounds like in addition to just the the raw hardware itself and being able to speed up all parts of the process.
It sounds like maybe the customers that you're working with also might need a little bit of strategic guidance in terms of Well, you tell me, is that something that they're looking for, that you're offering?
Yeah. It's delicate. I mean, we try to be as helpful as we can, but we also try to avoid consulting.
It is very tempting sometimes, especially when you really like the technology.
There's problems and data sometimes sitting there.
There's a real benefit. I mean, we talked to a company recently. that is manually inspecting trucks as they come into a warehouse for defects or for little dents or dings that can really cause safety hazards down the road.
So automating that is really compelling, but it's not really in a startup's DNA to consult.
We have to pick one thing and do it well, and building the tools is what we've decided to do.
On the developer end, what kinds of feedback are you getting from the hands-on developers writing the code that they then run on on spell what kinds of features that you offer things that you offer What are they really liking?
What are they asking for more of? Where do you see the The individual, the team, the actual people writing the code, what are they looking for to make their workflow better?
I think the remote execution is definitely our most popular product.
And it's basically introducing, you know, we think about if you want to run a large-scale neural network training on eight GPUs somewhere in the Cloud, the actual mechanics of doing that can be pretty daunting.
What we think about is this magical thing a tool can do where there's some upfront cost to, in this case, running an experiment.
But if you can get that cost down to be low enough, it kind of disappears mentally. thinking about experiments you want to run, building the infrastructure to run them, running one, seeing some results and running another experiment to actually just firing off experiments. at the rate at which you can think of them and write the code.
I think that's the biggest feedback we get is that, With your tools, we're experimenting more.
We don't think about infrastructure. Everything's organized.
It's where I expect it to be even if You know, I only realized that I'm looking for it later.
And then one thing that people say, and this makes me very proud is that, you know, we feel like the the underlying product identity of spell is very similar to unix something you mentioned that we talked about before and that's what we try to do we try to build these well-understood, componentizable sort of products that you can use together and have a lot of flexibility.
Our guest today is Sirkan Piantino. He is the co-founder and CEO of Spell, spell.run. which is a tech platform for machine learning and AI developers that connects developers to computing clusters in the cloud so they can run their AI experiments more quickly, more easily, and as Sirkan was just speaking to, without having to worry about all the things you have to do before you can get to to the work, which in any line of work, including my own, I can absolutely relate to.
Let's shift gears for a second and talk a little bit about your background.
As I understand it, grew up on the East Coast, came out to the West Coast, worked at Facebook. went back to the East Coast, and was heavily involved in AI research in New York City working for Facebook, but also kind of involved in the burgeoning AI scene there.
So take us back, sir, kind of, if you will.
How did you first get interested in computer science, machine learning, AI?
Yeah, I mean, probably the most interesting thing is my first job at Facebook was building newsfeed infrastructure.
So we built this system that still powers all the ways that we render newsfeed.
And just as virtue of the fact that I had some experience working with data.
I'd worked in finance before that. I ended up inheriting and rewriting all of our ranking algorithms.
And my first experience with machine learning was actually keeping it pretty far away from our systems, which is sort of ironic because later I would go on to co-found this AI research group based heavily around very advanced machine learning.
To me, that's a story about how different what we call machine learning today is, from what machine learning was in 2008, 12 years ago.
The field is really conquered a humongous amount.
It's really come a long way. And I think it's pretty ironic that I had that experience.
Is the growth accelerating still? You know, that's something that I go back and forth on.
I think, you know, there was a time where every year forget about like an open problem, but like almost an entire field of computer science research was getting totally uh reinvented by neural networks and those were definitely heady days i feel like there's a lot of great research that's still happening Particularly, I think there was a turning point perhaps in the community, when a talk that a researcher gave at NeurIPS about how little we know about why the techniques that we use actually work on the problems we use them on.
And what I find is that there's more work being done today to understand why certain things work for certain problems, which I think is really important.
It's almost as important as demonstrating a new technique.
And what I also find is that over time, even something that okay, You demonstrate that you can solve this particular problem with this particular technique.
The types of techniques that we use to solve different problems are getting simpler and standardizing a bit.
So I was talking to somebody in the research world a couple months ago, and it's amazing how quickly we've moved out of The goal of a researcher being to find the weird, quirky architecture that works for a problem in deep learning. to much more general sort of machine learning research and much more understandable sort of work.
So I think there's a lot of a lot that we don't know and a lot of ground that's being covered every year.
But there were some there were some crazy days that we probably haven't gotten back to.
So you made the move from Facebook AI research to founding Spell in 2016.
What was the impetus for that, and was that a big dramatic leap of faith off a cliff, or did it feel like kind of a natural progression for you personally? personally?
It was something I had wanted to do for a long time.
And I think actually getting the opportunity to work with Jan And BuildFair kept me at Facebook for a lot longer than I'd ever intended to stay.
What were some of the things you were working on there that kept you?
I just thought that neural networks were really fascinating.
I was learning just an unbelievable amount, not just about the science every day, but also about how research organizations function and about how PCI Express buses work and I was just sort of surrounded by fascinating things all the time.
And that's a pretty good deal. It is. It is.
But alas, the time came. And so tell us about the beginnings of Spell.
It's actually funny. One of the first ideas wasn't really... about machine learning tools.
It was about the hardware itself. We started to see every sort of sophisticated tech company amassing compute power.
It was kind of an arms race. Right, right.
And the goal was really, well, why don't we build a giant GPU compute cluster and then lease it out to people? as what work is happening in the largest companies starts to show up in smaller businesses and federate out from there.
That morphed over time. We decided that racking our own GPUs, building our own hardware was Probably not a great idea for a small startup, but... providing access to hardware and making things easier, giving everybody the same beautiful compute cluster that giant research organizations work on was a really powerful idea.
It's too open-ended of a question to ask, you know, where is AI going?
It's permeating everything. And as you alluded to, you know, banning the term AI, it's become kind of a a cultural buzzword or at least in business circles and a little bit more broadly as well. where do you see the field headed even in the very near term, the next couple of years, and whether that's something that comes specifically out of the services that your platform offers, opening up the tools and more powerful tools and kind of that infrastructure as a service almost kind of thing to developers to make it easy for them.
So specific breakthroughs that might be coming or kind of more broadly where you think the field of ML and AI development might be headed?
There's a couple trends that I think are really interesting.
I mean, kind of the most obvious one is advances in specialized machine learning, solving problems in computer vision, NLP, have kind of given people the inspiration to try to solve problems or to try to solve general AI.
So there's a lot of work trying to build actual intelligence into machines.
And it'll be really interesting just for the species to see how that work plays out.
And I don't think it's not going to be something that happens in the next year or two.
But just the ambition of that task is one of the things that I'm inspired by.
I would point to is people in the research community actually getting more applied in their work.
So you're starting to see more of people people in the research community working on problems in healthcare, particularly in healthcare working with hospital systems and other people that have large amounts of data.
So I think the difference between research and applied machine learning are getting a little bit blurred.
The third thing I would say is the tools are advancing really quickly.
So a couple of years ago, you had some pretty rudimentary, pretty buggy frameworks that you could use to cobble together the neural network you're thinking of and maybe train it on some data.
TensorFlow, PyTorch, and there's a long list of other frameworks, but the sort of core infrastructure for building something with machine learning and with deep learning, particularly running it, getting it into production.
All that stuff is getting better really, really fast.
And I think that's bringing more people into the field.
That's sort of moving as we think about it, moving what used to be this research-y, science-y type of thing into more just an applied software engineering problem, which is, again, our focus at Spell.
Absolutely. For folks who want to find out more about the nitty gritty of what Spell does, or perhaps even to start investigating how to work with your tools, where can they go?
I mean, obviously they can come and check out our website, spell.run.
You can also head to learn.spell.run, which is a... portal we created specifically for people who are looking to do some of their first projects with machine learning or want to work on some curriculum. to get started.
And then another option is if you're sitting in a computer right now, please pip install spell and start running some things on a cluster with a bunch of other people.
It really is as simple as that. Excellent.
Sir Ken Piantino, thank you so much for taking the time to tell us about what Spell's up to, a little bit how you got there.
All the best going forward. And maybe we can have you back on the show down the line and catch us up on what you're up to then.
Yeah, that would be fun. ¶¶
