The Six Five Team discusses IBM Granite GAI watsonx goes GA with Indemnity Protection.
If you are interested in watching the full episode you can check it out 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.
Patrick Moorhead: IBM came out with it. It’s Granite foundational models, goes GA with that and they added a level of protection for their customers. What’s going on here, Dan?
Daniel Newman: Well, you kind of said it already, but thanks. No. But you and I have talked quite a bit about IBM Watsonx. We’ve talked quite a bit about the, it was Watsonx AI data governance, and now we’ve seen the company further establish its approach, which is enterprise centric to act as market leaders in terms of what it’s doing to enable… And Pat. The word democratize? Further democratize?
Patrick Moorhead: You can throw that out, dude, as much as you want.
Daniel Newman: I’m just having a little fun here. But further enable the enterprises to not only approach AI and data, but also drive governance. And the other thing that I liked about this announcement around Granite was really about the fact that we know increasingly, and this is what we keep saying about the proprietary data and its value versus commoditized large language models is that IBM from the very onset under the leadership of Arvin Christian as I understand it, that AI is not going to be the same for every company. And so this whole thing with Granite models is all about being delivered in different sizes with different levels of customization, so that companies can basically get to the specific needs and insights that they are looking for and not looking at every model offering the same set of capabilities to every company.
IBM has further really added that it’s not only going to focus on its own foundational models, but it’s also going to open the door for third parties. So it’s got Llama, it’s got Hugging Face… Which Pat, we’ve talked a lot more… It’s funny because remember when Hugging Face was the name everybody always mentioned and lately it’s been more Cohere and Anthropic. But it’s really the three of those and then others. And the other thing is that the company has come out now really explaining, IBM has an immense and incredible… You like the word incredible. Incredible set of data that has been able to train models on. They talk about their five domains, Pat. Academic, internet, code, legal finance. But it is training and curating these models so that what people are starting with Granite is something that’s got a lot of what they need, but then like you said, they’re making it very open so you can add your proprietary data and you can do it in a way that includes the important governance, so that you get the most and maximize the value and the return on your AI investment running Watson.
This is what I’ve always liked about the Watsonx strategy from the beginning was that it’s enterprise centric, it’s customizable, it’s foundational model focused, and it’s open both internally to what IBM has created, but also externally to the larger open source models that all seem to offer a little bits of disparity in terms of their value. So you can bring together Daisy Chain, STACK to get the most value. Pat, probably the biggest piece of news here with something that we’re hearing more and more about, and this is what IBM referred to as it’s contractual protections for AI models. Given the fact that we know there’s a lot of questions and concerns around responsible AI governance, privacy and data, is that how does intellectual property get protected when you’re building this out, trying to do it very quickly and you’re picking different horses, proverbially speaking, to which technologies you want to use.
Well, if you’re using IBM, what they basically come out with in this announcement is they’re providing IP indemnity or AKA contractual protection for the use of its foundational models. This allows customers that are now committing to working with IBM and Watsonx to know for sure that what’s being created, that they are not going to get sued or have to deal with legal ramifications for leveraging IBM’s foundational models. This is very important. Now something that I think is always going to be the caveat as these open source things get brought in external and internal models, so the protections do they sit if you’re using your data plus their data or do you have to use exclusively their foundational models? These are some of the questions I’m still curious about. But I have to imagine it’s set up in such a way that it allows for all the data, your own data plus the model data, just as long as you’re not using any data that’s technically not your data.
But this, Pat, was really big news and I think this is going to be a fast following announcement from IBM where we’re going to see others come out in this space aggressively offering something similar. But companies can be very fast and very dynamic and very efficient and very productive through the utilization of generative AI. But when we’ve seen some of the risk factors of data ending up in the wrong hands, model is being trained on data that isn’t owned, isn’t protected, companies have to say, hey, if this information is going to get to our customers, if it’s going to enable our users, customers, ecosystems, partners, vendors to interact with this data and this data factually inaccurate or even hallucination or ends up leading to a decision that costs has a big financial downside, for instance, to a company, where does the accountability lie?
And so this is a big step for IBM to basically say we believe 100% in what we do. We are going to back the customers that make this investment. But it will be very interesting to see how this plays out and if things like this, if this contractual protection ever even becomes a thing or if they’re able to really bet a thousand in terms of making sure it’s only the right data and only the right outcomes.
Patrick Moorhead: I view this as a major milestone for IBM, but also a good time to sit back and reflect on what the company has done in generative AI. They were one of the first to come out strong with a new generative AI platform with Watsonx. Three parts of that, AI, data, and basically the content protections. And IBM was first going GA, they were GA before Google, GA before Azure, GA before AWS. And we talked about that on the show and I do think that counts for something, right? Where IBM may not… They were first with Watson and Health, I think overall stubbed their toe a lot. It was based on analytics, not machine learning so it’s probably the wrong technology, but here they are, right? And I feel very comfortable if you’re an enterprise regulated industry that IBM has your back on this. If you want to train it at IBM, they’ve got their own Vela supercomputer. And in fact, this is where this specific model was trained, Granite, right? And the amount of information that they shared was really unprecedented. Neither OpenAI or LlAMA will tell you where they got their data from, okay?
So I think that is a little bit suspect to be honest, but also understand that it could be also viewed as part of their intellectual property. But if you go to the white paper, you can actually see right there the 14 sources that IBM pulls this from. And then they go through, and other people do that, that they talk about how much data they started with, which was 6.4 terabytes of data, took out 4.9 of dedupe, 3.79 terabytes of hatred, abuse, and profanity, which left you with a two terabyte model that’s ready for tokens. So the amount of detail is extraordinary here and I think IBM should be applauded for this by the way. The other thing is when you get the data set down to a meaningful, a smaller data set, guess what? The cost to do inference against it is less expensive and that’s a good thing.
I think this is also a really good model. And by the way, this is not a tiny model, right? It’s 13 point billion parameters, but it’s not this gigantic monolith that I think we’ve seen before. The other thing that I like, pricing. There’s freaking pricing right on the website. This is $.005 per 1000 tokens. Where have you seen pricing before on a website? I haven’t seen it anywhere. I’m sure the salespeople have it and they’re selling it, but it’s right there on the website. A final thing, a little twist up on the indemnity, just to give you an idea how far this is from other things. When you use Llama, Daniel, you have to sign a document indemnifying Meta, okay? From being sued as part and partial to a lawsuit against you. But here IBM is actually not guaranteeing a win, but guaranteeing that they’re going to pay for your legal bills on this.
Daniel Newman: That’s the whole thing about a real enterprise solution versus a consumer toy, right?
Patrick Moorhead: Keep going. Just fixing my earpiece.
Daniel Newman: It’s fine. I was just saying the big difference between a toy, a consumer product, and an enterprise product is if you want an enterprise, it’s kind of an ERP or any other solution, you want them to go big with your enterprise solution, you better be willing to back that your stuff works. You know what I mean? And this unfortunately, or fortunately, because it works at such a pace, there is literally no way to validate and fact check every single one of these generated assets at scale. So if you can’t do that and you can’t guarantee it, you can’t use it. This is like setting the standard. It’s a bit of a requirement. I don’t see how it isn’t a requirement going forward that these companies are going to say, look, we commit that what we’re generating is right.
Patrick Moorhead: Yeah. And listen, I’m just going to throw this out there. I asked AWS what their list of sources through their models were, and they said they don’t share that information for Titan. So it’s going to be interesting folks, right? And by the way, public indemnification of an entire program doesn’t mean that special people aren’t being indemnified. I’ll just throw that out there as well.