scikit-learn & data science you own

0:03Welcome to Practical AI, the podcast that makes artificial intelligence practical, productive, productive, and accessible to all. If you like this show, you will love The Change Log. It's news on Mondays, deep technical interviews on Wednesdays, and on Fridays, an awesome talk show for your weekend enjoyment. Find us by searching for The Change Log wherever you get your podcasts. Thanks to our partners at Fly.io. Launch your AI apps in five minutes or less. Learn how at Fly.io. Okay, friends. I'm here with a new friend of ours over at Timescale Avthar. Suothen. So Afthar, help me understand what exactly is Timescale.

0:45So Timescale is a Postgres company. We build tools in the cloud and in the open source ecosystem that allow developers to do more with Postgres. So using it for things like time series, analytics, and more recently, AI applications like RAG and search and agents. Okay, if our listeners were trying to get started with Postgres, Timescale, AI application development, what would you tell them? What's a good roadmap? If you're a developer out there, you're either getting tasked with building an AI application, or you're interested in seeing all the innovation going on in the space and want to get involved yourself.

1:17And the good news is that any developer today can become an AI engineer using tools that they already know and love. And so the work that we've been doing at Timescale with the PGAI project is allowing developers to build AI applications with the tools and with the database that they already know, and that being Postgres. Yes. What this means is that you can actually level up your career. You can build new interesting projects. You can add more skills without learning a whole new set of technologies. And the best part is it's all open source. Both PGAI and PG Vector Scale are open source. You can go and spin it up on your local machine via Docker, follow one of the tutorials on the Timescale blog, build these cutting edge applications like RAG and search without having to learn 10 different new technologies and just using Postgres and the SQL query language that you will probably already know and are familiar with.

2:08So yeah, that's it. Get started today. It's a PGAI project. And just go to any of the timescale GitHub repos, either the PGAI one or the PG vector scale one, and follow one of the tutorials to get started with becoming an AI engineer just using Postgres. Okay, just use Postgres and just use Postgres to get started with AI development, build RAG, search, AI agents, and it's all open source. Go to timescale.com slash AI. Play with PGAI. Play with PG Vector Scale. All locally on your desktop. It's open source. Once again, timescale.com slash AI.

3:04Welcome to another episode of Practical AI. This is Daniel Whitenack. I am the CEO at Prediction Guard, and I'm joined as always by my co-host, Chris Benson, who is a principal AI research engineer at Lockheed Martin. How are you doing, Chris? Doing very well today, Daniel. How's it going? It's going great. I was saying that I'm really pumped to be talking about something that's near and dear to my heart over many, many years. Because today we have with us Jan, who's the CEO at Probable, and Guillaume, who's an open source engineer at Probable. Welcome. Thanks for having us. Well, Jan and Guillaume are working on data science that you own, including projects like Scikit-Learn, which is, of course, very near and dear to me, along with other data scientists all around the world.

3:59So Jan, if you could, since you're coming from the CEO perspective, help us understand a little bit, maybe for those that have heard of Scikit-Learn or some of the other projects that you're involved with, but they haven't heard of Probable. if you could give us a sense of what is Probable. As you mentioned kind of in the lead up to this conversation, it's a slightly different kind of company that came about in different sorts of ways than other types of startups. So yeah, if you could give us a little bit of context, that would be great. Well, very glad to be on the show with you to date. And Probable is a company that is typically known as a spinoff from a research center in France called Inria.

4:45And INRIA is the place where this technology, PsychicLearn, has been developed over the past 10, 15 years. Not many people know that. And the project has been somewhat protected and sort of incubated within that research center. And after all that time, as you know, PsychicLearn has been adopted or even probably participated in creating the field of data science. because it is applied math and essentially has created a sort of paradigm for how data scientists approach data science, typically through two functions, fit and predict. And the French government has a national strategy for AI, like many, many countries.

5:31And the government decided to double down on psychic learn. And they came up with a budget. They entrusted the research center with that budget. But then they also asked for the project to be break-even at some point. And the team said, okay, break-even is fine, but we don't do that in the research center. We don't break-even. So why don't we call an entrepreneur to try and help us figure it out? And they called me. So I have a track record as a software engineer and an entrepreneur in tech for the past 25 plus years. but I'm not a data scientist. So I did my due diligence and I sort of dug deep to find out what this project was all about under the hood.

6:16Is it any good? Is the community any good? And of course, Psychic Current is this quite amazing jewel of a technology that every data scientist on the planet uses. I discovered that it was downloaded 1.5 billion times cumulatively 80 million times a month, 22 % in the US, only 3 % in France. So this is a project that is used all over the world. And Probable is essentially the spinoff that takes all of the team, including Guillaume here, from the research center and turns it into an open source company that inherited the mission that was initially given to the research center. And the mission is to build a suite of open source technologies, including Psychic Learn, but above and beyond Psychic Learn as well, for data science.

7:12So the scope is large. The mission is noble. And this is what we're building, essentially. So Probable is a one-year-old company that has already started doing many, many things. And Guillaume is the representative here for Psychic Learn, this technology that is used, again, by every data scientist on the planet. Well, this has brought up a lot of interesting questions on my end. And I really love the part of your pitch and at least how you framed it on your website and in your materials online about data science that you own and the open source side of this. which I know from experience, there can be some interesting challenges around finding business models that really work with open source technologies.

7:59And we've seen technologies where companies start with the posture towards open source and then gradually become more closed over time. So I'm wondering from the leadership perspective, it sounds even in the way that this company was formed that there is a posture towards stewarding the, you know, PsychitLearn and these types of projects. But from your perspective, what is your posture towards stewarding these projects in the open source side? And how do you view the business element of this to make it sustainable in the longer term? So that is the hard question, but it is the one that is important here.

8:41Psychic learn is a technology that is, again, applied math. It's not rocket science, but it's applied math, and it's quite intricate. The thing is, the scientific community uses it day in, day out, and everyone depends on this. So typically, when I discovered the scope of the project and the mission that was entrusted to the research center, I realized that this project is bigger than me, number one. Number two, the mission is to actually create more open source. In other words, in 2024, it's even more acute. Typically, big tech keeps on amassing so much power, concentration. And we could argue that they do not distribute as much as they should.

9:28So that's not a judgment, but it is a fact. And psychic learn is precisely the contrary. It actually enables so many companies to do data science. So with that in mind, before creating the company, we decided to craft a sort of architecture for the company that would respect that. And so before we created the company, before Guillaume joined as a co-founder, before he even incorporated the company, we had a template that actually created the governance, the shareholding structure, and also leveraging a new law in France that allows us to do a sort of B Corp. So a company with a mission, where the mission is clearly stated in the bylaws, and that mission is to create open source for data science.

10:17So in a way, we've created a sort of constrained environment that is unlike many companies because it's by design. This company by design has created guardrails so that the governance cannot take this company too far on the right, let's say, proprietary technology or even changing the license. That's not in the cards. And we've created a sort of mechanism where, you know, if we do not uphold to the mission, then we can actually lose some of the assets, such as the brand. We are the official brand operator, but the brand belongs to the research institute, right? Still. So there are many mechanisms, trigger mechanisms that force us, including shareholders that we would bring in, to actually bind with the mission long term.

11:08Gotcha. You've raised so many questions for me that I want to ask. I actually want to take just a moment and kind of go back because it occurred to me as we're talking about this, for some folks listening who may have never even used Scikit-Learn, they might have heard the name and stuff, and you talked about it being applied math. Could you guys expand on that a little bit for somebody who hasn't had a chance to ever actually utilize it themselves in terms of what it's doing and kind of catch them up to us in the conversation a little bit? And then I'm going to pepper you with a few more questions because you've got me really interested.

11:42You hit so many topics on that last. Yeah, so maybe I can give a bit of background. So, basically, the tagline is machine learning in Python. So, it goes back to the, let's say, statistical route. So, the simple answer is we try to make predictive modeling. So try to use mathematics to form like data doable in the future to like give answer to a specific question, to a specific paradigm. The big difference with generative AI or deep learning is just that all the statistics that you have there are simple stuff. So they are fundamentals and deep learning build on those, but are just like much more, let's say, costly to train, costly to in the inference states or not in the same scope as well.

12:35And is like the de facto choice when you want to have like tabular data. So Excel spreadsheets data structuring this way. So that's a de facto way of training to be able to hit those spreadsheets and give back some labels or some regression, let's say. And whatever is like image or NLP or like this is like, let's say, deep learning and Transformers is more in that area. So we are more like back to what was machine learning like a few days ago, but that have many, many, many applications. Well, considering how incredibly popular and foundational in the data science world, could you kind of give me a little bit of a landscape view?

13:21And I'm not sure which of you would be the right one to answer, so you guys pick between yourselves. But a little bit about kind of how that fits into the data science landscape with AI coming in, just so that with people listening, they can kind of go, Ah, I see how it fits into the many organizations and tools that are out there. How do you think about that for that? And then I'll get back a little bit more to the organizational stuff that was talking about a few minutes ago. So maybe I can answer like partly, which is by giving use cases and to see that, for instance, with a partner as well that we worked over the years to give like, where do you find machine learning?

14:02And for instance, machine learning can be found in healthcare, where you want to know if the drug works or not. Then if you want to find diseases as well in some type of data, it could be as well like fraud detections in banks, in insurance, predictive maintenance, and those type of like all applications that you have since like many years. so let's say the use cases are very very large and what brings like cyclone on that is that this is not I mean it's not for one of the use case I mean it was fought from the beginning to be general enough such that you can apply to any of those use cases and to come back to let's say classification and regression programs let's say or unsupervised learning as well but like that you can apply the tool anywhere in that field So maybe, Jan, you have something more to add?

14:55Perhaps also at the macro level is to say, you know, Psychic Learn does a lot of things, including deep learning. But to be frank, when you want to do deep learning, typically you'd go to PyTorch or TensorFlow. But for everything else, you know, Psychic Learn. In other words, in the great AI family of algorithms, there is machine learning. And within machine learning, you have deep learning. Within deep learning, you have other categories of algorithms such as transformer-based models that lead to LLMs. So it's basically Russian dolls of sorts. And scikit-learn is the biggest provider of algorithms in the machine learning space.

15:40And in fact, if you look at the downloads, typically, Scikit-Learn is downloaded as many times as PyTorch and TensorFlow combined, which is crazy because now everyone is talking about LLMs, of course, but also deep learning because deep learning is currently in a spring state, not quite a winter yet. So, of course, deep learning and Gen.AI is a wonderful breakthrough. That being said, I like to simplify sometimes the 80-20 Pareto distribution. So I had the intuition that 80 % of the use cases out there use scikit-learn when it comes to machine learning. And people actually tell me, no, Jan, you're wrong.

16:26It's more like 90, 95%, right? Because in terms of, you know, technology that is robust, that is tried and true, that is used to actually, you know, turn a profit or return on investment. Banks and insurance companies, right? Guillermo was mentioning fraud detection. Fraud detection typically uses scikit-learn. And that actually saves money. That actually, you know, banks would be losing money without that. So it is actually quite essential. But again, it's applied math, right? So scikit-learn is only a facilitator to this category of problems.

17:16what's up friends i'm here with a friend of mine a good friend of mine michael greenwich ceo and founder of work os work os is the all-in-one enterprise sso and a whole lot more solution for everyone from a brand new startup to a enterprise and all the ai apps in between. So Michael, when is too early or too late to begin to think about being enterprise ready? It's not just a single point in time where people make this transition. It occurs at many steps of the business. Enterprise single sign on like SAML auth, you usually don't need that until you have users. You're not going to need that when you're getting started.

17:54And we call it an enterprise feature. But I think what you'll find is there's companies when you sell to like a 50 person company, they might want this. They actually, especially if they care about security, they might want that capability in it. So it's more of like SMB features even if they're tech forward. At WorkOS, we provide a ton of other stuff that we give away for free for people earlier in their lifecycle. We just don't charge you for it. So that AuthKit stuff I mentioned, that identity service, we give that away for free up to a million users, one million users. And this competes with Auth0 and other platforms that have much, much lower free plans.

18:28I'm talking like 10 ,000, 50 ,000, like we give you a million free because we really want to give developers the best tools and capabilities to build their products faster, you know, and to go to market much, much faster. And where we charge people money for the service is on these enterprise things. If you end up being successful and grow and scale up market, that's where we monetize. And that's also when you're making money as a business. So we really like to align, you know, our incentives across that. So we have people using AuthKit that are brand new apps, just getting started companies in Y Combinator, side projects, hackathon things, you know, things that are not necessarily commercial focus, but could be someday.

19:03They're kind of future-proofing their tech stack by using WorkOS. On the other side, we have companies much, much later that are really big who typically don't like us talking about them. They're logos, you know, because they're big, big customers. But they say, hey, we tried to build this stuff or we have some existing technology, but we're sort of unhappy with it. The developer that built it maybe has left. I was talking last week with a company that does over a billion in revenue each year. And their skim connection, the user provisioning, was written last summer by an intern who's no longer obviously at the company and the thing doesn't really work.

19:35And so they're looking for a solution for that. So there's a really wide spectrum. We'll serve companies that are in a, you know, their office is in a coffee shop or their living room all the way through. They have a, you know, their own building in downtown San Francisco or New York or something. And it's the same platform, same technology, same tools on both sides. The volume is obviously different. And sometimes the way we support them from a kind of customer support perspective is a little bit different. Their needs are different, but same technology, same platform, just like AWS, right? You can use AWS and pay them$10 a month.

20:03You can also pay them$10 million a month. Same product. Or more, for sure. Or more. Well, no matter where you're at on your enterprise ready journey, WorkOS has a solution for you. They're trusted by Perplexity, Copy.ai, Loom, Vercel, Indeed, and so many more. You can learn more and check them out at WorkOS.com. That's W-O-R-K-O-S.com. Again, WorkOS.com.

20:50so Jan you were kind of already going there and I I love the direction that you're going with this but I think maybe I could tee up a softball for you here because I'm personally passionate about the answer to this question and you probably have a better view on it. But there might be people out there maybe listening to this podcast who are thinking, well, now that we have Gen AI, we have large language models, I could put in a prompt to one of these models to do fraud detection or to find entities in text or to make some prediction of a classification. And sometimes that works. And so maybe there's people thinking, well, there's these general purpose large models out there.

21:38How does that change the way that something like Scikit-learn plays in industry? And I personally would argue and think that this actually makes Scikit-learn more valuable, if anything, rather than less valuable in terms of the ways that it can be combined even as a tool that's orchestrated with Gen.AI models. But I'm curious your perspective on this from the business side, and maybe Guillaume has some ideas on the technical side. Yes. So, Psychiclaren typically is this one technology that is patrimonial. In other words, it belongs to everybody. In fact, there's another stat when you look at the figures that are public, by the way, the number of dependencies.

22:24So Psychiclern is actually used by nearly 900 ,000 projects on GitHub. So there's nearly a million projects that depend on Psychiclern. And there's a new law that I discovered recently, someone mentioned that, Lindy's effect, which means that something that's been used long enough will remain important for long enough. So not saying that PsychicLearn will go the way of COBOL, but PsychicLearn is here to stay and we are with the community, the guardians of that. So we're going to make sure that PsychicLearn remains there forever for companies that actually need it in a stable version. And of course, Guillaume and the team are building up new features as we go.

23:10So there's a dedicated effort, and I should say that we have carved out nearly 10 people in a team are doing only that, contributing to Psychiclern and the other associated libraries. Now, your question, Daniel, is whether Psychiclern will be obsolete in, say, a number of years because general purpose technology has made it irrelevant in some ways. Number one, Psychiclern is extremely frugal. It actually works on CPUs. And it is well-controlled, well-understood. It's actually quite predictable in some ways, whereas deep learning is usually known as a black box where it's really, really hard to introspect.

23:56And so Psychiclern does produce for certain categories of problems things that are actually working quite well, More so than large language models, for sure, today, and more so than any sort of deep learning-based technology that we understand today. Now, it is possible that with additional data, additional training and techniques, and even evolutions on the transformer-based model, we could improve and probably render obsolete psychic learning. But to us, and then Guillaume and I, we talk about that with the team. We also experiment with LLMs. And we are also trying to figure out how we can use these new technologies to actually help our first persona.

24:45And that is the data scientist. So we are a technology provider to help data scientists and increasingly so the data scientists in enterprises, because we will be creating value-adding services and solutions so that we can generate revenue to sustain our mission. So the goal for us is to actually project ourselves while contributing to open source, but also create a sort of business value proposition, not dissimilar to Red Hat, because that is the closest type of company that we identify with in terms of spirit. To that point that you're making right there, I'd like to get back to something that you said earlier that feels like you're kind of tying back to it anyway there.

25:31And that's that you talked about the mission to create more open source and the mission that you're trying to create this environment that you're describing by design, you said. And that with Scikit-Learn here to stay for the long haul, it's going to be something that is not going away soon. It's solving such a high percentage of the problems. Could you describe a little bit about kind of what you're thinking around that in terms of further developing this particular set of software and the ecosystem around it so that we have the benefit for many years to come? How are you approaching that? So the company is built with multiple business units, if you wish.

26:12That's a big word for startup, right? But we have multiple revenue lines and multiple activities, even within the open source team, which is dedicated. So Guillaume perhaps can elaborate on some of the other libraries that we support that complement Scikit-Line. So that's one way to answer the question. But also we are building a new product, which I call reversible SaaS. So we are building a product that will provide additional value to data scientists. and the goal is to create a sort of, I don't want to use the term copilot because that is too close to LLMs, but it is the spirit. We are building a companion to augment the work of data scientists all the way to teams.

26:58So that is an additional product on top of PsychiCon because PsychiCon just works. And so we don't want to change that. And contrary to a company that would build a SaaS solution with a proprietary approach, we want to say, okay, whatever you guys use is fine. We need to find a way to add new value. And some of it will be open source, fairly modular. But for those companies that have more money than time, that need more service than be on their own, we'll have a solution for you and we'll make your life easier. And data scientists are a new breed. It's a new type of job. It's not been around for very long.

27:41And in a way, when I talk to people, so I've been in code forever, and you know this, right? The developers, when they get hired, they are turnkey in some ways, right? They have their Git environment, and they know how to peer code, and that's all pretty standard. But when you talk about data scientists, it's actually quite artisanal. It's an art and a science at the same time. And you're manipulating two objects, actual code, but data scientists are not coders. And you're manipulating actual data. It's not code. It's patterns. And so data scientists have a difficult task, which is to combine these two things and create value for the enterprise.

28:21And then they talk to business units and they're like, what do I do with this model? How do I put it in production, right? So there is a huge conundrum to solve. And that's what we're going to do, additionally to building open source that are modules that people can use. Maybe, Guillaume, you can elaborate on some of the other libraries that are key to actually help. Yeah, so within Probable, we have the open source team. And so we worked for many years on PsychicLine already. But we see the importance, as a community, we see the importance of putting models into productions and as well getting closer to the data sources.

28:57So we are just like working on libraries that should like make those come together. So for instance, we have a library that is more on the MLB side that is called SCOPS. We work a bit to make like the persisting more secure in some way. But we look as well on how to bring databases like SQL words into like closer to the machine learning models. so like how can you transform data with states with different tables and how you can be in your python words we were caring so much about like sql for instance and how you can bring this into psychic learn and within psychic learn as well we want to improve like whatever is visualization evaluations inspection of models which is on the top of just like training an algorithm because So we want to augment all those aspects beyond those.