Open Source Models: An Exciting New Business Model For Enterprise AI

0:00Okay, AI fans, I want to talk about AI tech a little bit today and really talk about this big new topic of open source models. Open source models are LLMs that are delivered by smaller software companies who take open source code, in other words, source code, not the actual platform, and they build their own model from that. There are at least eight really good ones out there. And what these providers do is they are software companies and they build, they take an open source model and they tune it and they tweak it and they improve it and they distill information and behavior from other models.

0:39Distill just means they run code that tries and uses the existing models to learn how they should train their model. And because they're open source, they can affect the weights and change the weights of the model and tune it, which you cannot do in one of these big frontier models. The two big frontier vendors are OpenAI and Anthropic, and the rest of them in sense are open source. By the way, Microsoft is getting into this. NVIDIA is getting into this. Meta is getting into this. These Chinese companies are getting into it, and this is very, very emblematic of many other things that have happened just like this in the software industry before.

1:18The database industry went this way. The browser industry went this way. Most software markets, when they become big, attract software companies who build smaller, cheaper, open-source versions of the tools. The companies that have prevented this from happening don't do it by preventing people from creating tools. They do it by market leadership. You don't see open source versions of Excel because Microsoft has dominated the space through OEM relationships and low prices and embedding it into other products. But there could be, if they charged a fortune for Excel, there would be open source versions of Excel.

1:55Believe me, there would. There were open source versions of relational databases when I worked for Sybase. They've kind of quieted down because the database market has been subsumed into applications. So what does that mean for you? It actually means three or four really, really important things. First of all, there's a stack of technology that's created in AI. There's the chips. There's the hardware systems and memory. There's the operating system layer. There's the large language model itself. There's the context layer on top of that. And then there's all the training and data on top of that.

2:29And then there's the tool or the harness that you use on top of that. And so when you use Gemini or OpenAI or ChatGPT or Claude, you're seeing the harness and you're not seeing all this stuff underneath it. And each of those layers has a bunch of companies that are trying to optimize their profit in the stack. So NVIDIA is charging, I don't know,$50 ,000 for a chip set per CPU set to a data center provider on the hardware at the bottom of the stack. And then OpenAI and Anthropic are charging you for the middle of the stack. And then the software application vendors, including us, by the way, with Galileo, are trying to make money at the top of the stack.

3:07And so what's going to happen, very, very clearly this is going to happen, is as these open source models become more popular, and by the way, they're performing at pretty similar capabilities as the frontier models, is that the value of the stack is going to move up, which is where we want it, by the way. We don't want to pay$100 a month for a model with no application value at all, which, by the way, is kind of what these cogeneration systems are. We want to pay, if we're going to pay$100 a month, we want it to be a whole application. We want it to solve an entire problem for us. So what's clearly happening is the cost and price, rather the pricing power of the frontier models is being threatened very significantly by these guys.

3:50And there's all sorts of shenanigans going on to try to stop them. Now, you know, OpenAI and Anthropic are lobbying with the government to try to prevent this in the United States, but they can't really prevent it because the software will make it into the United States and around the world. By the way, Mistral in Europe is an open source model. There's a lot of them. And then, so there's the protectionism of the big guys, but as I said, they really can't affect it. And since both of those two big companies haven't gone public yet, they're probably very, very nervous about their valuations given the massive amount of money they're spending on data centers.

4:22Because in a sense, the frontier vendors are investing at the bottom of the stack, building actual data centers at the very, very bottom, which is the most commodity-like part of it all, just to get into the market and stay ahead and generate the revenue they can. Now, the benefit to us in corporate is two or three things. First of all, pricing is going to go down, which is great. So you're going to have an IT department, if you're a medium to large company, who's going to know what these things are. And they might pick one. By the way, Meta is in this market too. And they might just use it. And they might just say, you know what?

4:55We're not going to pay for ChatGPT or Gemini because we feel capable of using this other model. So you're going to still get your company chatbot and your tools and all that stuff. But we're going to build it on one of our own, quote, systems. And we'll save a lot of money. And you guys can do a lot more with it. Or we're going to build an application that has a model embedded in it. And you don't have to use OpenAI. I was at NVIDIA yesterday talking with NVIDIA about Galileo, and one of the comments they made to us was, does Galileo work on Nemetron, which is the open source model from NVIDIA?

5:31And we looked at each other and we said, I don't know why it wouldn't. So there will be a project to spin up a version of Galileo on Nemetron so that NVIDIA can use it. And I don't think it's going to be that hard for us to do it, to be honest. So there's that. The second is the value of the stack of applications is going to move up. And, you know, I've really believed this from the beginning, that the platform is a technology and you do have to pay for it. It's obviously got value, but it's the application that really gives you a return on investment. You don't get a return on investment by just rolling out a chatbot and giving people the opportunity to play with it.

6:07You need to build solutions, build applications, connect it to your systems, connect it to your current data infrastructure, and then really getting it all to work. Actually, even at NVIDIA, they were actually very proud of the fact that every NVIDIA employee, all 55 ,000 of them, are building their own agents. And then we asked them some basic questions about what's connected to what, and they said, well, we don't really have this connected to that, and this isn't connected to that. They haven't spent as much time on the integration of all of these different data sources inside the company, at least within HR, yet because they're so busy building stuff and they're so innovative as a company.

6:42But those of you that are in a banking industry or healthcare or retail or manufacturing or financial services or insurance, you don't want everybody hacking around building their own thing. You want a solution that's integrated. So if you look at Sana from Workday, if you look at the new AI tools coming from many other vendors, the reason they're worth spending money on is because they solve a problem, a recruiting problem, a training problem, an employee experience problem, or an analytics problem. And so more money and more value and more pricing power will go higher up in the stack. The third implication of this to me is very much where we are, which is I think the AI market and what you're going to see in the value proposition from different providers is going to be much more vertical.

7:26So if you want an AI that's really good at leadership development or an AI that's really good at recruiting or an AI that's really good at supply chain optimization or an AI that's really good at sales automation or customer service automation, and you try to hack it up yourself with ChatGPT or Anthropic, you're building it on a platform of data that is completely nonspecific to your application. And we have just been doing a bunch of benchmarks, which we'll produce in a white paper, with Galileo versus Claude versus OpenAI. And it's twice as effective. And there are hundreds of errors in the open source models for typical or what we call the 30 gold queries that we have for Galileo.

8:14And that's not because the models don't work. It's because they're trained on so much heterogeneous information. I was kind of joking about this to a group of people yesterday. If you asked a PLOD or OpenAI a question about management, leadership, performance management, diversity or something, it's going to read articles from BuzzFeed, articles from the New York Times, articles from the Atlantic, articles from Wall Street Journal, articles from Mad Magazine, articles from People Magazine, articles from Reddit. and then it's going to take all that stuff and all the words and tokens and all those articles, by the way, where the word performance might have different meanings in different sources, and then it's going to mash it all together and it's going to answer your question.

8:56So you're getting a big, big average of a bunch of unknown stuff when you ask it a question. And so what these frontier guys have to do, since they are not specialists in any particular domain, is they have to spend billions of dollars on external data labeling companies like Mercore or Handshake to try to train their content. And that is an impossible task. It will never be as good as a domain-specific AI because the domain-specific AI is highly trained on one domain. And the reason I feel confident in even saying this is we have seen this in Galileo. Galileo, because of its deep knowledge and data set in skills and jobs, in all of the domains and the use cases and the examples and the case studies of different applications of management and leadership and recruiting and HR is very, very insightful.

9:50And we hear this every day from different companies. I just was talking to IATA, the Airline Industry Transport Association, about the skills model they built for the airline industry on Galileo and how quickly they were able to put it together because Galileo understood all the skills information they got from Boeing and Airbus and other providers. And so regardless of which part of the business you're in, you know, maybe you're in recruiting, maybe you're in L &D, maybe you're in employee experience, maybe you're in comp, maybe you're in executive development, maybe you're in sales and marketing, you're going to probably within a few years not want to use a generic solution for your AI in those areas because these vertical providers are going to be very, very specialized and they're going to know what all these use cases are.

10:38And we've seen this in recruiting that generic off the shelf tools, and even NVIDIA was saying this yesterday, are really not trustable for scoring and ranking candidates. They can do it, but they don't have all the right criteria or all the right training. So that's the third implication. The fourth implication is the cost is going to go down. I mean, the only way I can see the AI market growing to the size that the Wall Street investors believe it's going to grow to is for it to be embedded in almost every device we use, in our phones, in our clothing, in our eyeglasses, in our computers, in the future of our computers, whatever they may look like.

11:18And there's no way you're going to pay OpenAI and Anthropic for all that stuff. You're just not. And these hardware providers are going to embed the LLM into their hardware. I don't think it'll be very long before your PC or any other computer you buy has AI in it, and you're not going to know where it came from. It's just going to be embedded because the hardware providers want to move up the stack too. So this open source approach, even though it's kind of flaky and a little bit weird at the moment, is really the right direction. Now, the last sort topic that I'll just discuss on this is the government.

11:49And we have a very maybe schizophrenic government in the United States right now. And it's been schizophrenic for a while because during the Biden administration, there was sort of a political war against the Biden administration by venture capitalists that they were trying to regulate or control the industry. But now that the industry is out there and it's actually become democratized, the investor community wants the government to control it again, because the people that have invested in open AI or Anthropic do not want a small company in China to produce a model that's 80 or 90 percent as good as the one they produced.

12:28Well, you know, in the world of capitalism, luckily in the United States, usually the best solution provider wins out. And as these AI markets evolve, and certainly the way I see it going, the winning vendors are going to be the ones that move up the stack. If you're selling a component and it isn't the best component, you got to keep making your component better and better and better. That's what chip companies and database companies and other companies have learned. But if you take it upon yourself to build a stack of solutions like Oracle did in database and Workday did and so forth, you can compete at a different value proposition, sell to a different customer and create premium pricing power in different ways.

13:09I also think, so the government, you know, may have lobbying going on by the big vendors claiming to be national resources or national assets, and therefore the government needs to own part of them or whatever. I hope that doesn't happen because this open source trend is a huge source of innovation that we need. We need more innovation. We can't have, you know, 10 guys deciding every decision that needs to be made in AI, each of which making$100 million a year if we want this to really become as big as it's going to become. So I think this is a very natural direction. I guess the final point I'll make is that although you may not feel this way, I can see this technology becoming so ubiquitous that we don't focus on the technology so much anymore.

13:56And the analogy might be your phone or your PC or your ear headphones or whatever it is you use as a consumer item. Consumer products or mature technology reaches a point where we're not really that focused on what's under the covers. And so in the case of our world in HR, when we buy a chatbot, we're in some sense buying a tool that can be used to build solutions. And if the harness on the chatbot, like Microsoft Copilot, for example, has automatic access to your email and your calendar and so forth, it's more of a solution. But until you turn it into a solution, It's not that interesting. And so more and more of the solutions are going to be full stack.

14:39Like if I think about Workday or HiBob or SAP or Oracle or whatever system you use, those software companies, many of whom have had pushback on their financial prospects because of AI, are most likely going to embed AI into their systems. And you're not going to be paying for it separately. And they're not going to just charge you for stuff because they can. they're going to try to reduce the cost of their AI solution by embedding models in their systems that don't cost them a lot of money. Now, I don't know exactly what Oracle or SAP or Workday are doing in specifics, but one of the benefits of the open source models is that the applications we buy, whether it be those big ones or recruiting tools or whatever, are likely to be better integrated, better priced, more flexible, and more feature-rich because they don't have to pay massive amounts of money to these big frontier vendors.

15:35So any of you who are looking at this stuff, let me just conclude with a little bit of a discussion of what we're doing with HR 2030. I'm doing a massive webinar this morning for 4 ,600 people. HR 2030 is not just a bunch of technical white papers. It is that, but there There is a whole architecture under the covers of 161 agents with data interoperability, object model, and workflows for you to use to simulate and design your future HR agentic strategy. And we're not a software company and we're not planning on building all sorts of applications at this point, although we may build a lot more stuff on top of Galileo.

16:16The reason we did this is so that you can take your perhaps confused or evolving strategy for where you want to apply AI and HR, and you can literally use the tool set of HR 2030 to evaluate vendors and capabilities of the solutions you're buying and building against this model and see what the implications are for the overall HR function. And we're now rolling that out to clients. You can get it through us directly as part of a workshop we do, and we will be talking much more about it in the months ahead when we go out to the trade shows and do all our stuff out there. So I hope this is a good education.

16:54I look forward to seeing a whole bunch of you guys on the webinar today. And look up HR 2030 online and take a look at what's available. And please call us, and we'd be happy to walk through it and help you with your strategy. That's it for now.