IBM EMEA's Ana Paula Assis explains why enterprise AI needs multiple specialized models, not one giant one. A European bank deployment required six risk assessments and seven regulatory certifications. IBM open-sourced its Granite foundation model and launched InstructLab, an open-source tool that adds new capabilities to models without full retraining.
Chapters
really transitioning from the experimentation phase to realizing value from AI. Just to give you an example, I was with a big bank this week here in Europe to talk about an agreement that we just celebrated. They are adopting our Watsonx platform. And it was interesting to see the steps that we needed to walk through in order to get to this point. We had to pass 6 risk assessment tests. We had to provide 7 certifications that are specific to address compliance and regulatory requirements that they have. So the path for AI adoption in the corporate sense, in the business world, is a very different game. It's very different from the general user that is using AI for their own consumption.
IBM is not buying the NVIDIA GPUs like others, but you do have a partnership with NVIDIA. So just explain what that is.
Yes, we actually buy. So we have our Avela supercomputer that we use to train our own models. So we have that partnership on the infrastructure side as well. But we are developing software stack on top of the NVIDIA GPUs to accelerate the development and make sure that developers have a user interface that is more easier for them to do, that is easier for them to develop AI applications.
And I think it's fascinating what you say about that use case with the bank and how difficult it was to get there. So on the one hand, as Karen was saying, we're seeing huge demands for NVIDIA chips from the big cloud players. But is that— is the demand there for the end customer? Does the end customer want all these cloud products that the hyperscalers are selling?
Yeah. So it's It's a very good question because what we're going to see in this world of AI is that one size is not going to rule all of the models. We are going to have very targeted models. We are going to have models that are going to be focusing on specific tasks. And those are probably going to be the most efficient models for you to run inferences on the AI and make sure that you are not using computing capacity that is not necessary to address a specific task. Yes, there will be lots of applications that will probably require very large large models. And one of the things that we are introducing, that we introduced actually this week, is an open-source project in partnership with Red Hat to accelerate training models based on the existing capabilities so that you can have additional skills and knowledge added to a model without having to retrain it all over again. So to give an example, this will allow us to take from probably months to train a model to do code development on a certain programming language week. So that's kind of the acceleration that we're looking at. And that capability is called InstructLab and we are announcing it on GitHub.
You mentioned open source there. It's a big debate right now, the open source versus the closed systems we see from the likes of Microsoft. Why are you convinced that open source is a part of this AI future?
IBM has been a collaborator in the open source community since we started participating in the Linux project 30 years ago. Right. So we believe in open innovation. We believe that this is the way that technologies can scale faster. and in a more secure way. And in that sense, this week we announced that we are bringing one of our most powerful foundation models, Granite, to an open source adoption using a license of Apache 2.0, both on Hugging Face and GitHub. And we are bringing those models to the open source community to develop on top of it.
Can I bring up the use of data? So we were talking about how these models get trained up, whether they are using information that is out there, and we're talking about night about OpenAI having a collaboration with News Corp, the owner of The Wall Street Journal. But after the use of original data is there, the next phase is synthetic data. This is data that the computers themselves generate. What does that next phase signal?
Yeah, so as I was mentioning, right, so InstructLab, for example, one of the things that it does is it extrapolates using synthetic data from a few examples that you can provide from your own corpus of data. And then it extrapolates hundreds of thousands of examples as part of that with the concept of synthetic, because it's not going to be able to generate all the amount of information that is out there, or usually consider just the amount of information that is out there to train the models. We'll have to rely a lot on synthetic data. And as you do that, you have to make sure that you have the proper algorithms, the proper methodologies to, to make sure that as you are using synthetic data, that data is aligned with the models that you want to train.
