AI models sometimes have an almost mythical status. You can do a lot with them, but building and deploying them is no easy task. If it were up to SAS, that wouldn’t be necessary at all. How does it want to achieve that? Some announcements during SAS Innovate shed light on this question.
Organizations that want to start using AI do not always have the easiest of times. The benefits of AI are often quite clear, but how do you get started with it? Building it themselves can be a bridge too far for many organizations because of the complexity involved. They simply do not have the resources and/or expertise to do so. Yet they want to get started with it.
One solution to this problem would be to simply purchase AI models from somewhere that organizations can immediately deploy. Another is to significantly lower the threshold for developing AI models. For both approaches, SAS announced updates at SAS Innovate, taking place this week in Las Vegas. First is the general availability of SAS Viya Workbench for AI model development. A second piece of news is that the company will launch a series of lightweight AI models aimed at specific industries. In this way, it hopes to lower the barrier to getting started with AI models a little further. Finally, SAS had several things to share around generative AI (GenAI), and what developers can do with that when building applications.
SAS Viya Workbench
SAS Viya Workbench is not new. We already wrote about this last September, when SAS launched it. Today’s news is mainly that it is now generally available. It is part of the overarching SAS Viya data and AI platform. SAS made this to give developers and AI modelers a relatively easy way to build AI models. It is a self-service environment with scalable compute (CPU/GPU) based on the needs and requirements within a specific project. Within this environment it is possible to do data prep, analysis and the actual development of analytics and AI models.
Lightweight is again the theme here, as we have already indicated with the packaged models (which we will return to below). It should enable users to increase productivity and ultimately improve model performance as well. These users (the developers and modelers) can work in different languages within this environment. Initially, it supports the SAS language and Python. Support for R is slated for the end of this year. As an IDE, there is a choice of Jupyter Notebook/JupyterLab and Visual Studio Code.
Furthermore, Viya Workbench uses SAS’s PROCs. These are procedures for (statistical) analysis of all data in SAS datasets. These procedures can take care of rearranging data as well as providing all kinds of insights in the form of tables, graphs and reports, to name some examples. In addition, given the option to program in Python, native Python APIs are available as well as Python libraries developed specifically for Viya Workbench. These do not deviate hugely from what developers are used to, but do provide a significant jump in speed and performance, according to SAS.
SAS Viya Workbench will be available in the second quarter of this year through Amazon’s AWS Marketplace. There are plans to make it available on other clouds in the future, and there will also be a SaaS version of Viya Workbench.
Industry-specific packaged AI models
As already indicated, not every organization has sufficient resources and expertise to build their own AI models. This while there are clear performance improvements possible even for those organizations when they deploy AI. For those types of organizations, SAS has good news during Innovate. Today it introduces a new line of products/services that enable organizations to start using AI quickly and in a focused way. That is the intention at least. These are what SAS itself calls packaged AI models. These are models that can be more or less rolled out out-of-the box.
These types of packaged models are obviously not suitable for general-purpose environments. That is already difficult with if not impossible with LLMs, let alone with the kind of AI these models offer. SAS came up with packaged AI models for specific use cases in specific industries. SAS is also immediately announcing today the first pre-packaged AI model. This one goes by the name AI assistant for warehouse space optimization.
The name of this first AI model indicates two things about this new offering from SAS. First, it is really very specific. It is not a model for optimizing the use of space in general, but specifically focused on warehouses. Second, you can tell from it that it is an assistant for someone else. Since most warehouse workers probably don’t have a background in data science, this means it has a low threshold for adoption. In other words, it is a lightweight model suitable for people without technical baggage.
More generally, we see the announcement of SAS’s prepackaged AI models as an important trend for the AI industry as a whole. AI is (thankfully) beyond the great hype of 2023 and is getting back into some more realistic and practical territory. We saw this earlier this week at Appian World as well. That company added some very pragmatic and sensible AI skills and features to its platform. SAS is doing the same with its packaged AI models. That’s an excellent thing as far as we’re concerned. It allows organizations that can’t build AI models themselves to get started with AI and reap the benefits of it.
GenAI in SAS Viya
We can’t write an article in 2024 without talking about generative AI (GenAI). Needless to say, SAS also puts a lot of emphasis on this during SAS Innovate. Again, the main emphasis is on the pragmatic approach, which should provide developers with useful GenAI additions to the applications they build, for different industries.
There are several components that SAS is emphasizing in terms of GenAI. These include the ability to leverage external LLMs to add the capabilities they provide to an application. In addition, a private preview of Viya Copilot is available. This does what every Copilot does. It helps developers get their work done faster. Also interesting is that it is possible to build your own Copilot, for the specific industry in which an organization operates. All this with the guarantee that data never goes out of its own environment, no matter which LLMs you use.
A third component has to do with the type of data used for GenAI applications. Real-world data is not always reliable or sufficient. Hence SAS’s emphasis on synthetic data, to add to the real-world data. SAS Data Maker is a new product SAS has developed around synthetic data. This is a no-code/low-code platform that you can use to ensure that GenAI always has the right data at its disposal. It is in private preview as of today.
Finally: nutrition labels for AI
Being able to build and deploy AI models is one thing. However, doing so responsibly is also very important. The ethical side of AI has traditionally received quite a bit of attention at SAS. It certainly continues to do so today. Indeed, today it announced so-called model cards. You can think of these as nutrition labels like those you see on food packaging. Model cards give you at-a-glance insight into how good a specific model is for your organization.
Model cards will become part of SAS Viya as soon as they are available. The idea is that a model card will be automatically generated for models that use content derived from SAS products. These nutrition labels will also become available for open source models. This is due to the fact that within SAS Viya it is also possible to manage open source models. This also makes them immediately suitable for generating model cards. Initially, this new functionality will be available for Python models.
Model cards provide insight into all kinds of indicators that say something about how “healthy” a model is for you. Think accuracy, fairness and drift. You’ll also immediately see all sorts of information around governance. An example of this is information on when the model was last updated. You also see who is responsible for the model. This gives organizations insight into how to address any problems internally. Furthermore, the model card allows you to see what a model was developed for and whether it was used for use cases that are not within the scope. It also provides insight into the limitations of the model.
Model cards should lead to more transparency and thus better reliability of and trust in models. This is ultimately crucial when it comes to AI. You can come up with many great AI models and use cases for AI. If it isn’t possible to determine the trustworthiness of it, you might as well not bother at all.