UPDATED 09:25 EDT / MAY 23 2023

AI

IBM doubles down on generative AI and hybrid cloud

Returning to live in-person events, our overriding impression from IBM Corp.’s annual Think event a couple weeks back was that the company showed unusual discipline by confining the focus to a couple of core themes: generative AI and hybrid cloud.

Given the hype around generative AI with the public preview of ChatGPT and the central role of Red Hat’s OpenShift as IBM’s platform modernization strategy, the choice of those themes was not surprising. What was surprising was that IBM stuck quite close to the script, as a stroll through the expo area reinforced.

From a product announcement standpoint, the spotlight was on the new watsonx family of products targeting, AI builders and data professionals. To the uninitiated, watsonx is not a typo; the branding is purposely all lower-case, which plays havoc with spellcheckers. The brand, which is distinguished from existing Watson, encompasses AI-based applications and tooling, focusing on AI model lifecycle management, AI model governance and a new data lakehouse.

Lower-case watsonx is supposed to represent a new generation of enterprise AI. The first generation was largely centered around machine learning, developing linear models using algorithms around regression, clustering, classification and so on. Not that machine learning, deep learning or neural networks are old hat or that IBM is moving away from them. Quite the contrary (and in fact, watsonx includes a number of “classical” capital-W Watson tools).

But here’s a reality check: Anecdotal conversations in passing with IBM Think attendees found most of their organizations still dipping their feet into machine learning, but likely taking for granted it that has already been embedded in the applications that they use day to day. And we’re also not omitting mention of deep learning or neural network models, but development of such complex models in the enterprise has to date only represented a tiny tip of the iceberg.

The foundation model train is leaving the station

IBM Research has actually been developing foundational models for the past three to four years but hasn’t exactly shouted about it until now. Neither had much of the rest of the industry, but all that changed with the hype around ChatGPT, which has dominated the news cycle this year. But consider this: Six months ago, how many people heard of GPT?

The new generative AI generation is premised on supply of foundation models that engorge sublime volumes of data with highly complex algorithms; such models would be practically impossible for mainstream enterprises to develop from scratch. The guiding notion is for organizations to start with a prebuilt foundation model and customize it for their needs. The good news is that IBM is showing that generative AI is not limited to Large Language Models, or LLMs, of the kind associated with GPT or Bard.

IBM is not alone here. Amazon Web Services Inc. recently announced private preview of its new Bedrock service that will include foundation models for LLM, text and conversation, and text-to-image that are run on specialized training and inference chips. For its part, Google LLC just unveiled a choice of foundation models for coding, image generation and conversation in addition to the rough-cut Bard LLM.

Initially, IBM is readying a series of models that are individually targeted for generative or traditional machine learning use cases. Like AWS and Google, IBM will address LLM, but also offer foundation models for geospatial, molecular chemistry (often used for drug discovery), information technology events (for addressing IT operations), code generation and documents (which could provide a form of knowledge management). IBM has identified “digital labor” (e.g., contact centers), IT automation, cybersecurity, sustainability and application modernization as the highest-demand use cases. For LLM models, we view code generation as the first likely killer app.

A key challenge for customers is navigating and choosing the right foundation model, or models, for the task. IBM will be prescriptive in some cases, designing a specific model for a specific use case. For instance, IBM is looking to enrich processes such as human capital management, procurement, and cybersecurity with specific models.

Because generative AI is still quite new, getting enterprises up to speed will initially require high-touch engagements, ranging from fixed-term jumpstarts to traditional consulting. In the long run, we’d like to see generative AI itself applied in helping organizations navigate through selecting the right foundational model, and providing guided experiences for customizing them. Yes, this could get quite meta.

Model lifecycle management

Watsonx will cover model build and lifecycle management environment. As such, it will draw upon classical upper-case Watson tools such as Watson Orchestrate, Watson Assistant and Watson Discovery, and introduce new ones (e.g., Watson Code Assistant), while replacing Watson Studio with a new environment for training and validation in the build stage, and tuning and model serving for production.

A highlight is the addition of a tuning workbench designed specifically for foundation models. Model governance, which is arguably part of the model lifecycle, will be handled concurrently through watsonx.governance.

The watsonx portfolio won’t be limited to IBM-supplied models but will also support the use of models harvested from the open-source wild such as Hugging Face. It will leverage open source enabling technologies and frameworks such as Ray, for scaling distributed compute, and PyTorch, for optimizing Python models for production.

The governance side will also be ecumenical in its reach across models. IBM adapted several capital-W Watson tools for AI governance along with capabilities from OpenPages, but with this proviso: These governance tools were designed for classical machine learning models. For generative AI foundation models, identifying practical approaches for governance is still a work in progress.

Models need data

The other piece of the puzzle is watsonx.data, which is IBM’s new data lakehouse based on Apache Iceberg. We’re not surprised as to IBM’s choice of Iceberg, as it is the open-source lakehouse table format that has garnered the most cross-industry support, and because IBM views Databricks Inc., which is behind Delta Lake, as a rival.

Although IBM is just the latest provider to support Iceberg, its implementation is differentiated with remote distributed caching, which allows organizations with data distributed across multiple physical instances to cache it where they want. And it supports hybrid cloud deployment. By contrast, most other lakehouse implementations restrict caching to the local cluster adjacent to where the data is physically stored.

IBM’s implementation also supports interoperability with Db2 Blu and Netezza, providing existing customers a lift-and-shift upgrade that allows them to take advantage of lakehouse capabilities, with the most important being the ability to extend ACID to data sitting in cloud object storage. This accomplishes two goals: By supporting bidirectional integration with Db2 Blu (Warehouse) and Netezza, IBM lives up to the requirement for hybrid cloud support. By integrating with the rudimentary Iceberg data catalog, IBM customers get access to popular open-source formats in the wild.

And, in IBM’s implementation, they can also use Spark and Presto open source query engines. We expect that IBM will subsequently update Watson Query (a.k.a., IBM Data Virtualization) to support these open source engines and, of course, connect to Iceberg.

But there’s another piece on our wish list. We would also like to see IBM make Python a first class citizen in the lakehouse, just as Snowflake Inc. has already done through its implementation for Iceberg.

Sorting it all out

With a new lower-case watsonx brand joining existing upper-case Watson, there’s bound to be confusion as to whether watsonx is the new, replacement version of Watson. The same goes for watsonx.data and Cloud Pak for Data; is one the replacement for the other?

In actuality, watsonx is the environment for building, running and governing AI models. But then again, there is IBM Cloud Pak for AIOps. Clearly, the existing Cloud Pak offering was geared around managing the lifecycle of machine learning, rather than more ambitious foundation models. Then there’s Cloud Pak for Data, which has been IBM’s primary data, analytics and AI environment for hybrid cloud and watsonx.data about the lakehouse.

Let’s zero in on governance. AI models feed on data and the algorithms, features and hyperparameters that comprise the model. The relevance of both model and data are closely intertwined. You could have technically correct data – that is, data that passes the right quality, currency, sovereignty/localization and security/access control requirements – but if the model is based on false assumptions, the house tumbles down.

And the reverse is true if the model is built with the right attributes: If the data is biased, or conditions change requiring different features, proverbially the surgery could still be successful, but the patient dies.

So, IBM not only needs to implement full data governance in watsonx.data, but it also needs to integrate the data and model governance functions so that neither functions are siloed or implemented as afterthoughts. Under watsonx, data and model governance are supposed to be concurrent activities. We believe that, at minimum, the activities should be coordinated and managed through a single pane of glass and, in the long run, have remediation capabilities such as sliding bar controls that could juggle model features, hyperparameters and data set selection.

By the way, IBM is hardly alone here. The data, AI and analytics industry still needs to figure this out.

The same goes with rationalizing Cloud Pak for Data with watsonx and IBM’s emerging intelligent data fabric architecture. As the lakehouse, with Apache Iceberg support, watsonx.data would be a logical extension of Cloud Pak for Data. It makes no sense for IBM to offer two separate “big data” technology stacks or product portfolios. And the data discovery, orchestration and governance capabilities that are delivered through the data fabric also play in.

With watsonx being a good start for delivering a coherent hybrid and multicloud build environment for AI models, we’d like to see IBM finish the job by integrating it with the Cloud Pak for Data portfolio. IBM states that these pieces fit together; our response is that they should not be separate pieces.

Tony Baer is principal at dbInsight LLC, which provides an independent view on the database and analytics technology ecosystem. Baer is an industry expert in extending data management practices, governance and advanced analytics to address the desire of enterprises to generate meaningful value from data-driven transformation. He wrote this article for SiliconANGLE.

Photo: IBM

A message from John Furrier, co-founder of SiliconANGLE:

Your vote of support is important to us and it helps us keep the content FREE.

One click below supports our mission to provide free, deep, and relevant content.  

Join our community on YouTube

Join the community that includes more than 15,000 #CubeAlumni experts, including Amazon.com CEO Andy Jassy, Dell Technologies founder and CEO Michael Dell, Intel CEO Pat Gelsinger, and many more luminaries and experts.

“TheCUBE is an important partner to the industry. You guys really are a part of our events and we really appreciate you coming and I know people appreciate the content you create as well” – Andy Jassy

THANK YOU