Stanford's AI Index Report 2024

0:05Welcome to Practical AI. If you work in artificial intelligence, aspire to, or are curious how AI-related tech is changing the world, this is the show for you. Thank you to our partners at Fly.io, the home of changelog.com. Fly transforms containers into micro VMs that run on their hardware in 30 plus regions on six continents. So you can launch your app near your users. Learn more at Fly.io.

0:42Welcome to another episode of Practical AI. This is Daniel Wajnack. I am the founder and CEO at Prediction Guard, and I am joined today to talk about a very interesting report that we've talked about on the show today, the AI Index Report from 2024 from the Stanford University Human-Centered AI Center. I'm joined by Nestor Masley, who is the research manager at the Stanford Institute for Human-Centered AI. Welcome, Nestor. I'm super excited to be here. It's great to have you back. As I mentioned, we've talked on the show before about the AI Index report in previous years, but for those that haven't had that background or listened to those episodes, could you just give a little bit of a sound bite about what the AI Index report is?

1:36Sure. So the AI index is an annual report that currently is in its seventh edition that aims to tell a story of really what's going on in AI from a diversity of perspectives. We look at trends in technical performance. So what can the technology do now that it wasn't necessarily able to do five years ago? We look at trends in the economy, how are businesses is integrating this tool, how much are investors investing in this tool. We look at trends in policymaking, how are policymakers responding to what's going on in the space. And we don't just look at those things. We study research and development, ethics, public opinion, diversity.

2:18And I think really we aim to be kind of a one-stop shop, kind of an encyclopedia of what's happened with AI in the last year that policymakers, business leaders, or really anybody that needs to know and understand what's going on in the space can turn to when they have questions about artificial intelligence. Interesting. Yeah. And could you tell me a little bit about the Institute for Human-Centered AI, kind of why the Institute is undertaking this and how it kind of fits into maybe the wider set of things that the Institute does? Yeah, of course. I mean, Funnily enough, it's actually very recently we celebrated the fifth anniversary of HI!

3:00Congratulations. We were founded in 2019 and we're kind of five years into this wonderful journey. And I think the Institute really exists to try to advance AI research, education, and policy in a way that will fundamentally improve the human condition. The creators of the Institute, I think, came together really because they felt that AI could be an incredibly groundbreaking technology. It could be something that could really elevate the potential of humans. But in order to do that, we have to think very carefully about how we actually want to develop some of these tools. And that's what we spend our time thinking about at the Institute.

3:37Realistically, when it comes to AI, the ways in which this tool is going to develop is not only going to depend on computer scientists and how hard they're working, but it's going to depend as well on policymakers, on business leaders, and on the regular public. So those individuals as well need to be given a tool that allows them to identify and understand how the space is evolving and developing. That's what we aim to do at the index. The AI index is the report that we feel can give these individuals the capacity to make the decisions that they need to be making about this technology. Yeah. And the index has been published for, like you say, a number of years.

4:17And of course, I'm imagining that this last year, or maybe the last couple of years have been kind of interesting in that on this show, in our conversations, in conversations across the industry, generative AI has dominated a lot of those conversations, but doesn't necessarily mean that that is kind of the quote unquote AI that is impacting humans in their everyday life. Yeah, there's more to life than generative AI. There's more AIs. Yeah, exactly. So how did you come at the report this kind of time around with the acknowledgement that, of course, this is transformative, what we're seeing in many ways, but also it's not the only, like, you know, when you talk about a report about AI, it's not necessarily just a report about large language models, I assume.

5:08Yeah, I think that was important for us. I mean, we certainly added some new data points this year on generative AI because we felt it had kind of come to the surface and it was something that we needed to chat about. But in a lot of the chapters, I think for us, it's incredibly important to exactly, as you said, draw that distinction between foundation models, generative AI and non-generative AI systems. And for example, in the research and development section, we have information not only on the number of foundation models that different countries are producing, but also how many notable machine learning models these countries produce.

5:46The kind of idea being that it is possible for a machine learning model to be notable and not necessarily be of the generative type. Similarly, when it comes to the economy section, we track total AI investments, not just generative AI investments. and we even included a completely new chapter this year on some of the ways in which AI is interfacing and engaging with science and that touches on a lot of these kinds of developments that relate to some of the ways in which AI is used in non-generative ways but is still really moving us forward and leading to a lot of really exciting advancements.

6:25Yeah that's interesting and And to give a sense of how the index is put together, I'm sure people have seen different surveys that are out there, different takes on AI and where it's at. In terms of the approach of the Institute, what is the kind of main research mechanism that goes into the report? And how is it developed? What's involved? Who's involved? That sort of thing. Yeah, great question. It's kind of a two-pronged effort in that we collect data ourselves for particular questions that we feel there isn't already good data for. So I think we try to be strategic. And we find that if there's someone else in the research community that is already collecting data that we find to be relevant and interesting, we don't try to necessarily duplicate their efforts.

7:16And as such, we partner with a lot of data vendors like Accenture, GitHub, McKinsey, LinkedIn, the study portals. They all collect data that we find to be interesting, and then we work to then include it into the AI index report. But again, for certain topics where we feel that there isn't enough data, for example, we felt that there wasn't a lot of good data on the number of AI policies and legislations that were being released on national levels. We endeavor to collect some of that data ourselves. And in terms of the research agenda, it's set by the AI index steering committee. So we're very privileged to be advised by a diverse committee of AI thought leaders, people that were really are very influential in commenting on what kinds of things are going on in the AI space.

8:04People like Jack Clark, who's one of the co-founders of Anthropic, people like Eric Brynjolfsson, who is arguably one of the world's leading economists on AI, and people like James Monika, who leads AI research efforts at Google. We work with these individuals to discuss and identify what kind of topics we want to track. And we figure out as part of the process where each report should be going and what kinds of things we need to be chatting about. Awesome. Yeah. So it's not just AI researchers that are or, you know, pulling people that are that are pouring into this. It's actual AI people from industry, academics, but also spread across other areas like economics.

8:53Yeah, definitely a lot of diverse perspectives. And I think also perspectives from a lot of people that have been in AI for a long time, because we see this with a lot of technologies where AI announced itself in 2022, ChatGPT came out, and then all of a sudden everybody and their mother became an expert on AI. I think some of the people that we really work with have been in the space for a very long time and have kind of seen it ebb and flow, so can especially offer a lot of very valuable perspective on where we are in this moment and kind of contextualize and situate that in a very nice way.

9:27Yeah. Well, I do want to get into some of the specific points raised in the report. But before I do that, I'd love to ask, especially with these sorts of things, because you've been thinking about this deeply, looking at all the data and all that. Was there any one particular thing that stood out as kind of surprising or counterintuitive or just stood out generally in the kind of year over year progress of the report? Looking at this year's, anything that stood out as surprising to you in particular? I think thinking about the cost of some of these frontier models, one of the things we did in this year's report is we partnered with EPOC AI, which is a great AI research institute, to do some estimates on how much it costs to train some of these frontier AI systems.

10:18And I think we kind of all knew in the back of our minds that these systems were going to be very expensive. And we eventually crunched the numbers. and it's one thing to kind of anticipate that. And you actually see the numbers that GBT4 is costing close to$80 million, Gemini is costing close to 190 million. And you kind of see this trend line, which is just kind of going up and almost exponentially. So it really kind of puts into perspective how far we've come. And I think it also poses a lot of interesting questions about the future. How much further can we go? I think a lot of these companies are betting very heavily that the scaling laws are going to hold, that they can continue pumping more and more data into these systems, and that these systems are going to respond with improved capabilities and even new capabilities.

11:08And I think they're all making interesting bets. And it's going to be very interesting to see if that's going to hold and what the future may look like is something that I'm going to really be watching with a lot of open eyes and an interested perspective. Yeah, it seems like it's getting to the point where, and this was something that was also highlighted that I saw in the report about industry dominating frontier AI research. And I guess that's connected in some ways to what you were just talking about, because I'm thinking back to my own PhD and the five of us in our grad student office. office, right?

11:44It's no group like that that's going to go about kind of the training of one of these new models at this scale. Certainly, there have been efforts, I guess, from efforts like Bloom and others that have brought together researchers from around the world to work on models in almost like a collaborative way, like a CERN type effort or something like that, which requires a huge budget. But at least as far as I can tell from the data and what you're talking about is, you know, there's those budgets are increasing, these models are expensive, both to create and to run at scale. And so you, I assume that that is connected then to that observation of industry kind of leading the front on the research side.

12:30Is that fair? Yeah, I think pretty directly connected. Yeah, I think it's, I think a lot of these companies know that feeding more data into the system leads to better performance, or at least it has so far. And I think these companies are betting that that trend is going to continue. So they're pouring money in and kind of hoping that we're going to see improved capabilities. And I think part of the reason these kinds of estimates were so eye-popping to me is because even coming into the report, I was kind of aware that, as you said, no grad student could build an LLM, train it on their laptop in a way that would kind of compete with some of the language models that these big industrial players release.

13:13But we're kind of getting into territory soon where, you know, you'd even have to kind of wonder of the kind of industrial giants, you know, who can really afford to be building these things, right? If we're going to get into territories where they're costing a billion, if not more, I mean, 100 million is a lot of money, but it's a lot less money than a billion dollars, right? And the scale of who can kind of get involved at that level is a lot different. and varies fairly substantially. And that has a lot of important implications for how AI research is being done and what kind of topics people focus on.

13:46I mean, I think industry actors do a lot of really valuable research and they do contribute a lot of interesting insights. But I mean, at some point you need to pay off the investors and you need to make a product that is commercializable and that could kind of shift incentives in ways that, again, coming back to this mission element that we talked about with high might not necessarily align with building AI systems that do the most to really further humanity.

14:25Hey friends, this episode of Practical AI is brought to you by our new friends over at Plum. Plum is a low-code AI pipeline builder that helps you to build complex AI pipelines super fast. You can easily create AI pipelines using their node-based editor, iterate and deploy faster and more reliably than coding by hand without sacrificing control. Deployment is easy. Pipelines are live API endpoints. They eliminate the need for constant code redeployment and debugging by deploying complex AI pipelines as API endpoints. Team collaboration is easy too. Plum's declarative node-based editor enables you to build quickly while empowering non-technical roles to iterate on what you've done without breaking it.

15:12You can build advanced AI features, get structured output every time, transform data and leverage validated JSON schema to create reliable, high-quality structured output. So Plum is built for builders. Early stage product teams are using Plum to go from idea to validation in record time. To get started, go to useplum.com. That's Plum with a B, as in plumber, to request access today. That's U-S-E-P-L-U-M-B.com. Again, useplum.com.

16:04Well, Nestor, some of the things that you mentioned in the report are related to generative AI. Some of them are not. One of the things that I think is interesting about this particular index is talking about some of the things beyond the technology that might impact, you know, not only the direction of AI development, but actually does potentially impact practitioners like some of those listening to this podcast. And one of those things being the kind of sharp increase in regulation, in particular in the United States, as related to AI. I'm wondering from your perspective, especially as this index has looked at things over a number of years, kind of every year we're talking on this show about like, oh, when and how are the regulations coming down?

17:00What from your perspective can we discern from this kind of sharp increase in regulation in the US and maybe the practicality of how that will actually influence developers? Yeah, so I would say a couple of things. I think first big takeaway for me is that we're probably going to see a lot more action on the state level, and we already are, than on the federal level. I mean, if you look at the total number of proposed AI-related bills in the United States, on the federal level and the state level, there's been a pretty massive increase in the last seven, eight years in both counts. But there are substantially more state-level bills that have been put into law.

17:41I think looking at my notes here, close to 40 on the state level in 2023 compared to just one on the federal level in 2023. And I think this, again, is not surprising. I think it takes usually a little bit longer for there to be consensus on the federal level. But it might mean that we can usefully look to states as kind of being one of the first barometers of what's going on with the regulation and what we might see down the pipe in the federal level. I think a second point is that when you actually look at the regulatory agencies, they're all passing more AI-related legislation and AI-related regulation, but the regulation itself is coming from a diversity of bodies.

18:24So it's not just like the Copyright Office is kind of, you know, hogging all the attention and passing 30 AI-related regulations. The Executive Office of the President is getting in on the fund, the Department of Commerce, the Department of Health and Human Security. There are so many of these regulatory agencies that are thinking very deeply about this tool. And I think this reflects the fact that AI is a general purpose. It doesn't surprise me. But I think it's also a note that, you know, if you think to yourself that, oh, I'm not a computer scientist, therefore, I probably don't need to worry about how this is going to affect my lives.

19:03I would kind of, you know, urge you to maybe walk back that assumption because regulators across different spaces are starting to become much more actively involved in passing AI-related regulation. And I think a lot of us are going to be affected much sooner than we would like to believe. And I mean, coming back to the state level point as well, I think you're seeing now a lot of very kind of hot and contentious debates, especially in California around SB 1047. I think you're going to see that in similar states where, you know, I think four or five years ago, the regulation that we did see was quite kind of expansive and let's say a lot lower stakes.

19:43It was very often what we would call more expansive regulation that was saying, let's explore how we could use AI or let's kind of empower AI researchers. Whereas now the regulation is starting to become a lot more restrictive. It's putting in rules about how these technologies can develop and how people can use these technologies. And obviously when you go with that approach, There tends to be, I think, much more debate about getting it right. And that leads to a lot of kind of fiery opinions on either sides of any particular regulation. I guess there was this sharp increase in that in this past year that you saw.

20:19There's probably some people out there, as you mentioned, there's a variety of mixed opinions about this, of whether that's actually kind of drives more market share for these closed providers who have already built up some of that market and maybe have some of that influence and open access or open source community that maybe doesn't have the ability to do. or the willingness to put in a lot of effort into this compliance side. Was there any thought amongst the group around how not only this would impact, so regulations would come down, they would be maybe enforced potentially in impactful ways, but how that would kind of shape the development of this technology moving forward?

21:10Yeah, I mean, I think that's an open question. We don't do a lot of, I would say, predictive work in the index. I think we try to mostly look at what we know to be true so far. But I mean, certainly, I think at least if you look in California with the debate going on with the bill that I had previously mentioned, SB 1047, I think, yeah, it's this kind of question of open versus of closed source. I think that particular bill wants to put certain requirements on models that are above of a certain compute threshold. And some of these requirements are so stringent that, you know, conceivably a lot of open source developers wouldn't necessarily be able to kind of meet all of them.

21:52And, you know, whether or not these regulations are backed for commercial reasons, as you kind of mentioned, some of these industrial players could conceivably have commercial incentives to support regulation that maybe, you know, compromises their open source players. But I mean, I do think as well, there's a lot of people in the kind of AI safety community that very ideologically believe that AI poses a serious existential risk, and that we should really be hyper cautious about how we scale. and I don't think it's necessarily the position of the index to kind of say we support one position versus another but I do think what I would say is that it's now kind of we're now getting into a moment where again policymakers are starting to think about this business leaders are starting to think about this and we really need to think carefully about what we do and what we put into law because you know once it's law it's going to shape incentives and it's going to shape how the community develops.

22:50And I think there are compelling arguments made by people on both sides. But it's difficult to know what the future holds. And I think it's hard to know whether this is the right move or that is the right move. But I think we always have to try to hold ourselves to some kind of standard of intellectual accountability and kind of acknowledge that, you know, we can't always get it right. But we try to we need to try to have as thoughtful and as well balanced of a perspective as possible. Yeah, I think that's a really good perspective. You mentioned in when you were answering that about kind of some of these regulations being tied to kind of how AI will develop moving forward and in particular as tied to the size and scale of models being developed.

23:37I was really interested that, you know, there is a portion of the report that is just titled, Will Models Run Out of Data?, which I think is really interesting. We talked a little bit about these models becoming more and more expensive. But of course, there's an element of this, which also is the scaling of the data required to train particularly these generative models, but other models as well, maybe computer vision models or other models. And that is something that I definitely get the question every once in a while from people that have heard something like, oh, you know, these models have already sucked in all of the internet of data, right?

24:18Like what's left to train on? And then you've got another perspective that I hear sometimes, which is, well, now we're just going to fill up the internet with generated data, which cycles back into training data sets. So yeah, I'd love for you to bring any perspectives that you have on this question of will models run out of data? I think it's a question that's coming up frequently for people? Yeah, it's hard to know. I mean, that's the problem. You want to kind of go on these podcasts and give kind of clear, concise answers. But in a lot of cases, it's quite nuanced. I think there's reasons to be optimistic.

24:53There's reasons to be pessimistic. I think when you kind of look at optimistic reasons for why we might not run out of data, or why data might not be the bottleneck that we anticipated to be, I think there are some papers that seem to suggest that synthetic data could meaningfully aid in training AI systems. I think as well, if you think about it, these language models are substantially less efficient than the human brain. They see millions times more text than any human would in their entire lives, and in some cases can perform better than humans at certain tasks. But it's clear that there is kind of an architecture of the mind, that being our own architecture that gets the job done with a lot less energy.

25:38And plausibly, there's going to be more research being done on algorithmic efficiency that can maybe make it easier for models to perform at higher levels with less data. And I think as well, we're always creating more data. We're kind of in an era in which more and more data is being kind of manifested and put out into the world. Some of that, of course, is going to be generative and created by AI systems. But there's also new models like meta segment anything, which make it a lot easier to get segmentation masks. So AI could be used to also take more data from the world in a way that could help these systems.

26:15On the other hand, it does seem that if you kind of just look at the amount of tech stock that we have now, it seems like we might potentially run out of that relatively soon. So I think we cited estimates from Epoch. They've since updated those. They predict that kind of in four years, we might potentially run out in terms of if we're going to kind of continue scaling models in the way that we're scaling the models. There's also other papers that are not necessarily as bullish on the potential of synthetic data kind of plugging this hole and filling in the gap. I guess for me, the kind of really interesting thing is, you know, I'm really going to be curious to see what it's going to be like when GBT5 comes out.

26:55Because, I mean, all these models have, I think, been now they're good, but they're kind of incrementally improving their capabilities at various tasks. But it still seems to me like they struggle on some tasks like planning. They struggle on some tasks like reasoning. They're still somewhat prone to hallucination. And there are still kind of limits to what they can do. And I guess I would wonder, is scaling a transformer going to be sufficient to resolve some of those problems? Or do we potentially need a new architecture, a new way of building AI systems that could resolve some of those difficulties and resolve some of those challenges?

27:33And it's hard to know, but there's a variety of perspectives here. I'd be curious to get your perspective on it as well as someone that kind of thinks about this as a lot too. Yeah, yeah, I appreciate that. From my perspective, similar to yours, I think you can look at it like a mixed bag. I think the thing that I would maybe draw in here is that at least right now, it seems like a promising route forward is not necessarily relying on these models to, you know, foundation models to have all the inbuilt knowledge that that is needed to, for example, answer every question or have every fact or, you know, be able to process any.

28:22type of input necessarily. I think though certain things, like I've seen this in the progression of like function calling, for example, there was sort of a generation of these models where people figured out that, oh, we can like use these to generate calls to APIs or to functions. But none of those prompts were in these sort of fine tuning data sets. And so like they kind of worked for that, but kind of not. And now like the ones that are coming out now that have that pre-built into the fine tuning data sets, they do that much easier and can extend to a whole variety of function calling. So I think that these sorts of not like specific facts or specific knowledge about particular APIs or these sorts of things, but the ability to figure out what those more general kind of building blocks of robust AI systems are and building the prompt data sets around those.