 Meta platforms and IBM have launched a coalition of more than 50 artificial intelligence companies and research organizations that are advocating an open model of AI. Dubbed the AI Alliance, other members include Intel, Oracle, Cornell, AMD, the National Science Foundation, Dell, Hug & Face, Red Hat, Stability AI, ServiceNow and many others. We see this move as a way for the industry at large to collaborate with each other to ensure responsible growth and address key challenges around governance, chip supply, privacy, and to really level the playing field relative to Microsoft and open AI and the advantages they've gained. I'm Dave Vellante and I'm here with Rob Streche to unpack this announcement and give you our first take. Rob, what's your impression? I think it's a good thing and I think that it really does put a line in the sand about how are we going to get to open governance, how are we going to get to being able to have actual confidence in AI and I think there are definitely people who are starting to look at it and go from a privacy governance and data protection perspective. How do we really make this more transparent? So let's bring up Ken, the money slide here, the AI Alliance members. I mean, look at this, Rob. It's worldwide. I mean, they are really sending a message, $80 billion an annual R&D spend, 400,000 students supported by academic institutions and talking about a million staff members. I mean, California obviously has a big piece of this pie, but it's really global. Yeah. I mean, a lot of the supercomputing facilities, you have UC Berkeley on there, you had Boston University with the big centers for open compute cloud here in Massachusetts. One that you had, like you said, AMD, Intel, Dell, Hugging Face, stability AI out of the UK, a number of others, all worldwide and I think that to me was really the number of people who've jumped on board with this in addition to IBM, Meta and Red Hat really made a lot of sense because these are people who are building out the platforms, building out the infrastructure that AI is being built on, especially when you think about our power law and how as you get out to that long tail and things get more specific, it's going to be on that type of compute and that kind of more open infrastructure versus all being used in open AI or what have you. What do you make of who's not in the coalition? Bring up the slide again if you would, Ken, who's the obvious company that's missing? I think it's pretty enough Nvidia's not there and I think that Microsoft's not there, open AI is not there, Google's not there, I think that those four are kind of, I would say, very absent in the fact that this is where, if this is really gets steam and this becomes a center of where regulation is taken from and how governments start to look at and lean into, this could be a problem for those organizations that don't want to at least participate in there. So I mean, it's no surprise that they're not in there, right? In particular, a lot of the industry is saying, hey, Nvidia's gouging the charge of $250,000 for these massive supercomputers that weigh 70 or 80 pounds and you can't get them from anywhere else. So of course, they're driving prices up. At the same time, open AI and Microsoft have kind of taken off in the lead. Now, of course, they did fumble open AI in particular, the board with the whole meltdown of the governance structure and that's in part where I think this resonates certainly with me and I think it will with the industry. I know we talked to with our friends at ETR probably about 10 or more open AI and Microsoft customers right after that weekend, that disastrous weekend, and many of them, I would say over half were turning off co-pilots. Yeah, I think it becomes a, if you're not part of the club and you're left behind, it becomes a big issue for those who are trying to figure out where do I invest, where who do I believe, who do I, where do I push in on? And I think that becomes a big advantage to the openness. And I think, again, when you look at, you know, Matt Hicks is the CEO of Red Hat and his comments on this as part of IBM and everything like that. But still, they have their OpenShift, what used to be data science but now called OpenShift AI platform, which really competes against some of these. I think one of the ones that, another one that was absent, that was kind of surprising was Amazon, given that they're pushing in on bring your own model, a model openness and things of that nature, you know, I think that one was also particularly an interesting one, maybe it's just not yet kind of thing. I asked Matt Wood last week at the, at Ray W's re-invent about, you know, have they thought about consortia like this and, you know, could they potentially lead it? I would have thought they would be one that would be a natural, you know, leader for this. So it is somewhat surprising that they're not, maybe they've got some other designs on their own. Let's look at some of the key initiatives and the projects that the alliance is attacking if you bring up the next graphic. You know, Rob, this is, you know, sort of an ambitious effort by these guys, the first one right off the bat, fostering a vibrant AI hardware accelerator ecosystem. You know, what does that tell you? Yeah, I mean, that told me that there's a reason why NVIDIA doesn't want to be a part of this. If they're going to actually do something like open compute has done to servers and really focus on open GPU and open accelerators, I think this is a key to why certain people are involved. Some of the other ones that really stand out to me, creating, testing and benchmarking tools and methods for safe and trusted AI deployment, I think that's very much needed. And I think certainly with IBM being one of the leaders here, I think there's a trust factor that they've built up over the years. Number five, supporting AI skills, building education. That's of course very important. They've got a lot of universities involved in this advocating for policies that create a healthy open AI ecosystem. I mean, I think in general, the tech industry is not very good at lobbying, you know, governments. That said, IBM actually has a lot of experience in that regard. So there's another, you know, possibility. And so does Meta, by the way. And that is, you know, relatively new to that game and under a lot more fire. IBM's used to it from the, the, the DOJ in the 80s and 70s actually in 80s. But yeah, Meta more recently. Yeah. Let's talk a little bit about some of the other factors around large language models. I mean, IBM's got Granite, AWS has Titan and even, you know, some of the later entrants into the market. You know, last week at AWS re-invent, there was a lot of talk about this thing called Q, which is essentially their co-pilot, what was it trained on? We suspect it was trained on Titan, but Titan, you know, doesn't follow those scaling laws that, that we've come to know that are associated with these more advanced large language models. So we asked IBM, is Granite characteristic of those scaling laws that we've come to know. And they said, yes, Granite is the same as this is in the same algorithmic class of transformer based LLMs as others. And so far, obey, so far is the operative phrase there. Obey similar scaling trends as these other models, notably the Chinchilla scaling laws. So what are the Chinchilla scaling laws? The Chinchilla scaling law for training transformer language models. Robbiske just says that when given an increased budget in terms of floating point operations per second to achieve the optimal compute the number of model parameters and the number of tokens for training the model should scale approximately in equal proportions. But as we wrote last week, I want to share this with you and get your opinion. George Gilbert and I in breaking analysis, we talked about Titan not being of that class because it didn't adhere to the scaling attributes around compute power, model size, parameters and training data. Maybe I'm inferring a little different than the Chinchilla laws of scaling. What do you make of that? I think it goes what we're looking at goes beyond what the Chinchilla laws are. And I think that really is one of the keys is that you got to look at the other costs associated with because not everybody's going to build. In fact, most people will not or organizations will not be building their chat GPT. They will be using small more segmented models or SLMs, those smaller segmented language models so that they can really be focused and what they're going to do is look for the right foundational model that's been trained on the right data that they believe in and they understand and is transparent. Then they're going to go and actually tune that foundational model to their own data and then they will deploy that for inference. So I think when you start to follow this string through, you can't have everything go up at the same time at the same level. If you start to say, OK, you have my tokens and my floating point, you know, as the floating point and as the floating points go up, so do the tokens and the number of parameters. Well, that makes sense. But at the same time, at what cost and at what expense do you have that happening? And I think it doesn't take that into account, which is why I think what we have is goes well beyond that. So granted, it's 20 billion parameters. So the question is, does it scale to 200 billion parameters or a trillion parameters? How do those scaling laws, you know, get affected? You know, we'll see. But in general, IBM's posture, they've got a track record and open source. Similar to to Amazon, they're talking about LLM optionality. So whether it's anthropics, stability, AI, of course, hugging face gives you a lot of options for LLM's llama too. So that's that's smart. Customers want choice. But at the same time, if you don't have that core foundation model, like an open AI where Microsoft has locked it up, or in the case of Google, then you really do want to diversify. By the way, those companies can diversify as well. So, you know, it seems to me that granted's differentiation is you got data sources, they're indemnifying. In production, which is, I think, unique, you know, so they're really focusing on this enterprise ready piece of it, especially when you start using these co-pilots and, you know, they're what they call Watson code assistant, what Q is from Amazon, what, you know, Microsoft's co-pilots. Those are those are all critical. And I think, you know, the real focus on governance here and model and domain specificity in a moment will bring up the power law, but your your take first. Yeah, I think it is. It gets into that how many parameters do you really need for your use case to be accurate and how much is that going to really cost you what it's more about the data, because it comes back to garbage in garbage out. And if you don't have the right data in there to either actually at the foundation or in there to tune the model, you're really going to end up with just garbage. And that's when you start to get those hallucinations and you start to worry about what is leaking back into the models as well as you further tune those models on your own data. So I think, again, accuracy and cost are two things that definitely go up at as the parameters go up as the number of tokens go up. And I think what is going on is that from a sustainability and a power perspective, that also becomes more power hungry as you start to scale up as well. So the power law of Gen AI, something to develop by the Cube Research Team, it's on the horizontal access is or the vertical access is the size of the model. So you're going to have those big foundation models like open AI, like Anthropic, certainly Google that are really going to sort of dominate the market and get a lot of attention. But then you're going to have a long tail and that long tail is going to be all about domain specificity and model specificity, which is on that sort of horizontal axis. And then you have the torso, which traditionally in a lot of these power models is a hard right angle. That torso is pulling up because you have all this open source models. You've got meta, you've got, you know, Falcon, you've got stability AI, inflection, gleam, light tricks, Jasper, on and on pulling that torso up and to the right. So that's kind of an interesting dynamic here. And I think IBM is really playing off that with its emphasis on governance. I do. I believe that that pull and that openness pulling up on the torso is really about, Hey, how do you do larger models at a lower cost, lower power on premise with data, you know? And maybe you start out at 20 billion, maybe you move to a Lama two with 70 billion, maybe you go somewhere else, but do you really need a trillion for your use case? And I think that becomes the tradeoff. Do you need to go take over 100,000 GPUs to actually train your model or to tune your model on that foundational model? I look at it for most organizations. The answer is probably no, if they're just trying to automate their HR and their people trying to figure out how to do direct deposit, I do not see people needing a trillion parameter model to go and do HR. Now, it may be nice if that's also going and rewriting all your HR policies. And I think it goes to that breadth of use case that you're looking at really ties very tightly to those parameters and amount of tokens and follows the Chinchilla law, but it's use case specific. Yeah. And that very that domain specifically, you're right, we didn't need a trillion parameters to train the cube AI, the cube AI dot com. Check that out. I think high marks on the initiative. I love the focus on the governance, the model specificity and the collaboration across the industry. It's unclear what the governance structure is of this. When we had our little private briefing, you know, that was kind of TBD, but I have no doubt they'll figure that out. Maybe modeling after similar to the CNCF. But I think overall high marks is something the industry needs here. Give you the final word. Yeah, I think it's desperately needed. In fact, the CNCF has actually partnered up with some of these other open, I guess could say AI consortiums that are looking at governance and things of that nature. And I think that this is a key to really bringing, I guess you say transparency and some amount of sanity to the governance over AI, so that we have kind of a almost a united voice, a little bit more united voice going back to the governments, because I think that's going to be the key to when regulation comes down the pipe. Great analysis, Rob. Thank you so much. And thank you for watching this Analyst Angle. We'll see you next time.