Why the Future of AI Should be Open – The Six Five On the Road

By Patrick Moorhead - March 27, 2024

On this episode of The Six Five – On the Road, host Patrick Moorhead is joined by Dr. Dario Gil SVP and Director for IBM Research and Ion Stoica, Executive Chairman, Anyscale and Databricks and Professor at UC Berkeley. They engage in a captivating conversation about the significance of an open future for Artificial Intelligence and introduce us to the AI Alliance’s goals and initiatives.

Our discussion covers:

  • The mission and objectives of the AI Alliance within the AI industry
  • Historical context and motivations behind the formation of the AI Alliance and its unique positioning
  • The diverse makeup of the AI Alliance, including pivotal roles of IBM and Anyscale, in promoting AI accessibility
  • Strategies for fostering open and responsible AI development through the Alliance
  • Overview of current projects and future endeavors of the AI Alliance

Learn more at IBM Research and Anyscale.

Watch the video below, and be sure to subscribe to our YouTube channel, so you never miss an episode.

Or listen to the audio here:

Disclaimer: The Six Five Webcast is for information and entertainment purposes only. Over the course of this webcast, we may talk about companies that are publicly traded and we may even reference that fact and their equity share price, but please do not take anything that we say as a recommendation about what you should do with your investment dollars. We are not investment advisors and we ask that you do not treat us as such.

Transcript:

Patrick Moorhead: The Six Five is on the road here in my hometown of Austin, Texas. It is South by Southwest 2024, and we are in the IBM AI Sports Club. You may be able to hear the ping pong balls and see the golf examples here where Watson AI is being put to use. Yesterday, I attended a very provocative panel debating the merits of AI innovation. Should it be open, should it be closed? And I really appreciated the Red teaming that kind of looked at the other side of it, of the equation. It really made for a natural and learning experience, I think, for everybody in there. I have two of the panel members here. Welcome to The Six Five.

Dario Gil: Thank you so much. Good to be here.

Patrick Moorhead: Gil, Ion-

Ion Stoica: Thank you.

Patrick Moorhead: … Great to see you. No, this was good. I really appreciate. Now it’s funny, as industry analysts, we sometimes pretend we can’t learn anything. We bring all of the knowledge, but I walked away with two nuggets from that, so I want to thank you for all of that.

Dario Gil: Great. I’m glad you enjoyed it.

Patrick Moorhead: Yeah. So in the very end, which was kind of the punchline of the discussion was, “Hey, there is an open AI alliance that has been formed by multiple companies spearheaded first by IBM with Meta, with hundreds of companies and organizations out there.” Can you talk about that organization, for those who might not be aware of it?

Dario Gil: Yeah, happy to. In the context of a shared belief and a shared commitment that a healthy open innovation ecosystem is essential to the past, to the present and to the future of AI, we came together and we brought organizations from some of the world leading universities. Ion Is a good example here, represented Berkeley, but many world leading universities, startups, established players, science agencies, nonprofits as well, all coming together, all in total over 80 institutions. Collectively, they invest over $80 billion of R & D a year.

The universities, on aggregate, are educating over 400,000 students at any given time. And what we do is, we come together across a series of working groups, everything from benchmarking and the safety of AI models to enabling a diverse hardware ecosystem for AI, a whole variety of projects, but our shared commitment and our purpose is to collaborate, to create and to advance open innovation in the world of AI because we believe that that is the future. It makes it safer, delivers broader economic prosperity, if we have an open ecosystem. So that was the purpose and that’s why we came together.

Patrick Moorhead: So I have to ask you, Ion, first of all, can you talk a little bit about Anyscale and why are you interested in this alliance?

Ion Stoica: Yeah, I do wear multiple hats. I am Anyscale and Databricks and UC Berkeley primarily. And actually, in my history, I am open source proponent, advocate, build. I was part of many open source project before, like Apache Spark for big data, Ray, which actually, has been used to train ChatGPT and open AI and many others, also more recently. And so I am truly believer in open source and I do think that open source is fundamental like Dario mentioned, to accelerate and the economic progress. And why that, because the alternative, it’s obviously closed source, but closed source basically, assume that all these kind of, some of development is happening in the silos.

Each different company building a very large model and its own silo. So it’s little communication right now between the companies themselves, unless the employees will go from one company to another and certainly between academia and these companies. And there is no little network effect. Compare that with the situation in which you have an open source model, an open source infrastructure, where everyone can collaborate. So all of these great minds can come together and obviously, it’s going to accelerate the innovation which is going to accelerate the economic progress and growth.

Patrick Moorhead: So Ion described a little bit, I think, of how we got there, but why did IBM and Meta form this? What was missing? What conditions were you seeing that might be a sign that you had to pull this together? Because alliances are a lot of work, and I’ve actually shared a few, I’ve been part of a few in my history and some of them worked spectacularly and even some that I’ve used as an analyst are kind of press releases and don’t result. But what got us to this point?

Dario Gil: Look, I’ll say first, that it reflected a reality of how many people were working and what they were committed to. So if you look at, for example, what Ion was mentioning, I mean, in his lab at Berkeley, there’s a huge commitment and a long tradition of working in this fashion. So the first thing that drove this is seeing that all of these institutions, AI was a top priority for them, they were doing it in this fashion, but they also, the motivation, the action, came from the fact that we also saw lots of arguments being put forth around, oh, for some reason, “Open is dangerous, open is bad for AI.” And we’ve all heard that movie before.

Patrick Moorhead: Sure, Linux.

Dario Gil: We’ve heard it with Linux and we’ve heard it, like there’s many, many long stories around that. And all these institutions strongly disagreed with that perspective. But it wasn’t just a reaction to that. Maybe that was a catalyst, but it was also an affirmative approach to say, “You know what? This is what we do. This is how we work and this is why it makes it more innovative, safer, with broader prosperity. And actually, as you correctly point out, it’s not easy to bring 80 institutions together.

But the reason they all said it, “Yes, yes, yes,” and many more that would want to join is because they believe in it and it reflects how they want to work. Imagine the alternative, what is the alternative that only three or four institutions are going to control, are the only source of innovation and safety in the whole world? That’s not a world anybody wants. It doesn’t reflect reality. So there was a little bit of a spark of the animation, but the more important aspect is the positive agenda of, this is how we all want to work and this is how we want to drive innovation.

Patrick Moorhead: So can you talk a little bit about, what’s the makeup of the alliance? Is it all universities? Is it all for-profit companies? Is it something in the middle, hardware, software, networking? What’s the characterization of this? And Ion, you actually check three of the boxes.

Dario Gil: That’s right. The man of many talents.

Patrick Moorhead: Infrastructure, software, data and university. So can you talk about what attracted you to this?

Ion Stoica: Yeah, definitely. So I think there are very many similarities what happened in the past. And like Dario mentioned about Linux and so forth, we have the debate between the closed source and the open source. However, always analogies are good, but also sometimes are dangerous. A situation is never identical. So one thing happening here, as everyone knows, to do AI research and innovation is basically, you require a huge amount of resources. You read about everywhere about GPU shortage. Look at Nvidia stock, right? It’s a reflection of exactly that. So without this kind of resources, actually, which universities do not pose, they’re a little bit left out of this kind of process, which didn’t happen before In other areas.

When we develop Spark at Berkeley, we could do that. But here, to develop and train a new model, you need huge amounts of data. You need huge amount of computes we do not have. And I think the alliance allows all this kind of bright minds in academia and elsewhere, again, to be part of the process, take part of the process. Let’s not forget that deep learning revolution was started at universities in academia, like many other things in the past. So having everyone part of allowing them and enabling them to do now the research, I think it’s critical and extremely exciting. So we are very happy to be part of the process and help with It.

Dario Gil: But that gives you a little bit of the context of universities who are so important to this ecosystem that really are having that challenge of access to compute and data. So by creating the alliance, we also foster this way of working. So you were asking who is part of this? So you have great universities like Berkeley and Cornell and RPI and NYU and many others. And in Europe, EPFL and ETH and Imperial, and you have University of Tokyo and Japan, so many, many incredible universities are part of this. You’re seeing established companies but also diverse.

From software, like companies like Oracle and I see obviously, IBM and others, but you see AMD there and Intel there and startups that want to participate in the hardware ecosystem around that. You see startups, you see nonprofits, like the Simons Foundation for example, that supports a lot of basic science because they see how AI is going to accelerate science. But you even have entities like the Cleveland Clinic because they care about AI applied to healthcare. So actually, one of the strengths is you see tremendous diversity. You see science institutions like CERN or NASA right around that. So really, really diverse. Each one brings a different element and a different value to the equation.

Patrick Moorhead: It’s funny, everything has two edges, right? With that many people, how do things actually get done? And so how’s it organized and then what is the output? That is a key question.

Dario Gil: So look, the way it works is, that first there’s a method, we’re all going to work together and contribute in open source. So first, there’s a methodology to this. There’s no contracts, no IP, everything that we work together has to be done in the open, number one. Two, you begin by taking a survey, which we’ve done already. It says, “Okay, we have different tracks, everything from open data and models to how you benchmark to how you enable hardware ecosystem to work. So you have, “Here’s the tracks, who’s interested in what track?” People say, “I’m interested in track one and three,” whatever.

We take that in, we form working groups, which we have already, with leaders from different institutions. And then, they actually get together and they create a work scope and you’re going to see a lot of technical output that is going to happen starting a month from now. So it’s basically, we have a mechanism to work together. We have a preference mechanism to elicit who wants to be part of that, and we have an organizational structure to collaborate and to do that. And honestly, there was a lot of duplication. If you look at benchmarking or things like that, does everybody need to invent their own benchmarking or models?

Patrick Moorhead: Absolutely not.

Dario Gil: So there’s going to be a lot of also, the modern way of doing standard setting by best practices that we can all agree on.

Patrick Moorhead: Right. Again, Ion, it’s impressive, you’re representing almost three different viewpoints of this. What are you most excited about the deliverable coming out of this? How does it help what you are working on? Or is it, as an industry analyst, I do this, I have an interest, I know this industry needs this to innovate and innovation has always been disaggregated and that’s my interest in this.

Ion Stoica: Yeah, so I think, two things. First of all, like Dario mentioned, many of now, the members of the AI Alliance already, they are doing open source, is what they’re doing. That’s why they’re coming together. But they’re doing them themselves and I think that this alliance, bringing them together, providing some common infrastructure will actually enable them to collaborate closely, will enable us to avoid some duplication, to strengthen the most promising projects. And I think that should be very good for the entire open source AI ecosystem. And again, this will be another vector of accelerating the innovation. The other thing I think I’m very excited about, is that hopefully, we’ll enable the entire community to have access to high quality open source models.

And I say, open source model, is again, there are many degrees of openness. I’m referring to truly open source, which there are models which are trained on open source data sets, the algorithms to train the models are open and the models themselves are open. And therefore, everyone can take these models and then actually peer into them, see what problems they have. That’s kind of about the safety. That, we believe, that fundamentally, the open source is the best way to achieve and to lead us to safe model, to safe AI, and then iterate and then to have confidence in using them because you know what they’re trained on, you know what algorithms you are using to train. So it’s not only you have the weights which are open.

Patrick Moorhead: So let’s get tangible here. By the way, it always has to start with a vision and I like the way that you separated the work product and the deliverable and the different teams. Which projects are currently underway right now, work being done right now?

Dario Gil: The whole area around safety and benchmarking of models, that whole life cycle around that, is the most popular one. There’s many projects underway, but this whole aspect of this, a whole life cycle, as Ion was talking about, from data model validation and verification and how you do this, and there were many people proposing ways to, everything from the benchmarking to the safety aspects and the tooling associated with how you improve the safety of the models, that everybody was doing a little bit ad hoc. And this is a hugely active area where everybody’s saying, “Here’s everything that we’ve all created, and now there’s a process of rationalization, prioritization. Okay, this looks like this.”

So you are going to see in the next month around that, the first output of this very, very active working group. And I think it was really good that everybody rallied around this because in some ways, going to prove the point of what Ion was mentioning with is, “Hey, this is the best way to make safer AI.” Because then you can have many more eyes, many more talented people working on the problem, on how to make it safer as opposed to, “Hey, just trust me. Some people behind the scenes is taking care of it.”

Patrick Moorhead: Right. I’m really glad, by the way, those two seem to be very hard and like hitting a hornet’s nest. So I sat on a benchmarking board for five, six years and it was tough, I’ll admit, and I was on both sides of that, not only from a vendor side, but also, part of a chip company. And that’s a huge challenge. The bigger one is, how do you define responsible AI? Could differ by continent, could differ by company.

And I think what we’re seeing a lot of the conversation today right around certain models is, wow, this is kind of fraught with peril in a way, but to make it mainstream, it has to happen. And I have confidence that the industry is going to figure it out. Just like we figured out the safety of electricity, the safety of elevators, the safety of EAC, and things that came after that because new things are scary and have some negative consequences if not done appropriately. So I like the debate, like the conversation, I really like that you didn’t pick the easy ones to go after first.

Ion Stoica: If I may add, actually-

Patrick Moorhead: Please.

Ion Stoica: … That’s the key, right? Because you mentioned there is a large heterogeneity about what different countries may define its safety or responsible AI or different industries, different application domains. And this will provide you ultimate flexibility. Again, having access to the data, to the algorithms, to create the models because you’re going to train maybe on a subset of data, which you deem as being safer, you may use different algorithms or different alignments and so forth. So, having access to the entire process, it puts you in the best position to tune the model, to train the model for what you need.

Dario Gil: Yeah, I think that’s going to be so important because you’re absolutely right, is that different countries, different regions, different institutions will have different preferences and alignment around this. And I think it’s by sharing methodologies and transparency on this compositional of it that you say, “Look, this is what’s right for me. In your context, maybe you’re comfortable with that, but I’m not comfortable with that or I don’t need it,” around that. But today, we’re in a very different shape around that, where in the last year, it’s like, just a few examples that look like black boxes. So in opening all of these up, we’re going to end up a level of customization and preference and alignment that is going to much better serve the needs of a huge diverse set of institutions and populations all over the world, right?

Patrick Moorhead: Right.

Ion Stoica: Yeah. Think about the software industry. How did the software industry evolve? You have a set of tools, and of course, you have programming languages, but then everyone, these tools are flexible enough that everyone can take them and build their applications for their own needs, for their own domains. And the reason is, that everything, the entire process, is kind of open. Everyone knows, if you do this, you are going to get that, right? So that’s kind of the way you want to have in this kind of AI as well. And open source, it’s key to that. So it’s, again, we are not talking here about only open sourcing, the weights, the model itself, it’s the entire process, how you get those. So this will actually give you the power and flexibility to tune the model again, to get the model the best model for your own needs.

Dario Gil: Yeah, and it’s also, I mean the pipeline of the models, it’s also the tooling of how you go up all the lever to the application layer. And there’s also the support of open source project that allows you different accelerators to work. I mean, like if Intel, you want to take advantage of that or AMD or startups or others, how do we enable that to occur? So that’s why they care so much, companies like AMD and Intel and many others just to say, “Hey, I want my hardware innovations to actually work in a stack.” They could say, “Oh, I’m going to go in the proprietary way.” But they are all in and says, “Actually, I want an open stack that allows everybody to have a level playing field and everybody to benefit from an innovation ecosystem that my products can plug in into the broader AI story.” So you can just see why there’s so much business advantage that many, many, many players can benefit from if you have sort of this element of open source and open innovation, right?

Patrick Moorhead: Yeah. The hardware and the software and the entire stacks are very straightforward in terms of benefits. I mean, you can look at Linux as an example. I do get some of the feedback from some companies that might be like, “Hey, we invest $100 million to create this, shouldn’t we be the ones to get the benefit from all of this investment?” How do you respond to that? And they also say that, and you hit this in your opening, “A closed AI is safer and more secure.”

Dario Gil: Yeah. Well, the first thing I would say is, on the first part of the question, it’s an and, it’s not an or. I think, a good strategy, as a company is you say, “Here’s the elements of it in which I’m all in and open and why this is good and where I’m going to contribute.” And then obviously, it’s incredibly common and normal to also have an element of your software stack that may be proprietary and so on. So I think the key is not, is it this or that. It’s, what is it that we want in the open innovation ecosystem, and then, what are the opportunities for further differentiation that people could do in a proprietary fashion, right?

Patrick Moorhead: Right.

Dario Gil: So that’s number one. And I think, as for the safe comment, it’s just patently false. I’ll just go straight around that and we go back because back to the analogy around this, is operating systems, what is safer or not? It’s like this story has been told many times, it is not true, and why is it not true? The reason for that is, because you have many, many, many more eyeballs and many more talented people looking, understanding the thing and saying, “That’s a problem. Let me also fix the problem.” And that makes, inherently, for a safer thing of things are very complex and very distributed around that. So I think that people who say that open is less safe is just selling a kind of goods. It’s just simply untrue.

Patrick Moorhead: Sure.

Ion Stoica: Yeah. So one comment about the first point, the point about how you get the value, you invest, so what is the return to investment? I think, again, there are many examples in the past, and what the key observation here, this is not a zero sum game about expanding the pie. And the example I have in mind and the obvious one is the internet. The internet is built on open source protocols. No one kind of owns it, but creates the opportunity and the platform to build huge businesses. And it’s like let a huge growth look just around, most valuable companies or whatever you want to slice and dice it. The value it created was enormous. The same way you can think here, we are going to create this kind of open infrastructure which will drive this kind of innovation, which in turn, will drive huge economic growth.

Dario Gil: And that comparison is appropriate because AI, and the reason this matter is that, because it’s not a narrow piece of the pie, the AI piece is so horizontal. So it is indeed similar in that sense to the internet. And so things are broadly horizontal in which a huge fabric of innovation and economic opportunity is going to be built on. So it’s very hard to make the argument on something that is so broad, so horizontal that that should be controlled by just one or two players.

And even if they made that attempt, it’s just not going to succeed because too many people want to build on something that gives them their own competitive advantage on top of it, and they don’t want to be beholden to just say, “Well, from now…” It’s like saying in the internet, imagine the internet had been built, but it was only around one company or two companies. No one would accept that it wouldn’t have worked in the way it has.

Patrick Moorhead: Right. Appreciate that. It’s been a great conversation. Easy last question here, what is the next milestone or the next type of announcement that we should expect out of the alliance? I’m not asking you to pre-announce something, but what should they be looking out for in the very near future?

Dario Gil: They’re going to start seeing technical content being put out, both projects, best practices, methodologies around that, coming out of the working groups. I mean, we will see also, I mean, we announced recently an expansion of the Alliance where we added more members and so on. So you will see some more growth, but the most important people that should be on the lookout for is, Star and producing the technical content, which is the purpose of the alliance because it is an alliance of doing. It’s not about press releases.

Patrick Moorhead: No, that’s great. I appreciate that. Ion, Dario, I want to thank you so much for coming on the show here.

Dario Gil: Thank you. Thank you for having us.

Ion Stoica: Thank you so much.

Patrick Moorhead: So this is Pat Moorhead with The Six Five Media and of course, Moor Insights and Strategy. Check out all of our AI content and check all of our IBM content out. And we’ve got a lot of discussions on the AI Alliance with Daniel and I. It’s been a great South by Southwest 2024 here in the AI Sports Club. Tune in, hit that subscribe button. Take care.

Patrick Moorhead

Patrick founded the firm based on his real-world world technology experiences with the understanding of what he wasn’t getting from analysts and consultants. Ten years later, Patrick is ranked #1 among technology industry analysts in terms of “power” (ARInsights)  in “press citations” (Apollo Research). Moorhead is a contributor at Forbes and frequently appears on CNBC. He is a broad-based analyst covering a wide variety of topics including the cloud, enterprise SaaS, collaboration, client computing, and semiconductors. He has 30 years of experience including 15 years of executive experience at high tech companies (NCR, AT&T, Compaq, now HP, and AMD) leading strategy, product management, product marketing, and corporate marketing, including three industry board appointments.