Harnessing the Power of AI With Foundation Models

There is no disputing today that the invention of the printing press changed the course of human history. But when it was first introduced, large swathes of the world viewed it with great suspicion—some even banning it outright.

Speaking at the DataScience&AITrends Asia Summit last month, Kitman Cheung, chief technology officer for data, AI, and automation at IBM Asia Pacific, made this remark as he drew parallels between the printing press and generative AI today.

As with almost every technology invented by man, there are two sides to it. Just like printed books could be used to educate or spread propaganda, generative AI today could be leveraged to further our knowledge or generate fake news that harms us.

Using AI for good

How can we use AI for good? Cheung put the spotlight on sports, a universal language that has brought people together through the ages. When the BBC first incorporated live commentary in a broadcast boxing match, it was revolutionary and changed the way people watched television, says Cheung. 

“Instead of passively watching the game as it happens, viewers can now experience the buzz of the stadium from the comfort of their homes. It's like sitting there with a superfan who knows everything about the sports and starts telling you about those things.”

At the 2023 Masters Tournament earlier this year, IBM and Masters partnered to create a custom solution to improve the user experience for millions of golf fans globally. By using AI, engagements with fans were enriched through the use of automated generated spoken commentary.

Cheung noted that the idea wasn’t to replace golf commentators but to enhance the experience by bringing audio commentary to video clips. With some 20,000 created over four days of the golf tournament, this task would have been impossible to do manually.

The right model for the job 

While the interest in generative AI is undeniable due to its incredible ability to generate coherent, congruent responses, Cheung observed it is not suitable for every task. 

For instance, despite its apparent confidence in answering all types of questions, generative AI and its use of statistical probability to produce a ranked list of words would make it “terrible” at predicting the score of a golf championship. “That's what traditional machine learning is for.”

“As a technology, generative AI is just as important as predictive AI. Use them side by side, and you're going to drive a much better outcome than just trying to use one versus the other,” he elaborated.

But considering the complexities of training AI models, both generative AI and traditional AI, how can organizations deploy AI to move the needle in a meaningful way? This is where foundation models that can be fine-tuned to perform different tasks come into play. 

“We had a foundation model that speaks sports. We fine-tune it to speak golf one week and tennis the next week. The work is already 90% done, making it much quicker to adopt AI in your organization.”

Foundation models for business

Cheung outlined foundational models businesses could adopt as building blocks, ticking off large language models (LLMs), IT automation models, digital labor models, and cybersecurity models as examples. 

“We can easily see some business use cases that foundation models can help today. LLMs can be used to answer questions for knowledge discovery and knowledge management for internal employees. We are already seeing those use cases on a regular basis.”

IT automation is a logical extension of programming, where such a model can be used to write scripts and automate IT tasks. And in a call center, AI can also serve to enhance productivity.
The idea is to create foundation models with domain-specific focus. 

“We are not trying to build a massive model that does everything. We want to build a smaller model that does specific things; there are a few things that matter when it comes to those models,” said Cheung. 

“When they're smaller and lighter, you can host them anywhere; you can transport them anywhere. And you can retrain them a little bit easier compared to larger models, especially when you want to do fine-tuning.”

Looking ahead

The momentum behind foundation models is growing. According to Cheung, IBM and NASA earlier this year announced a collaboration to build a trio of foundation models to address the challenge of climate change.

One of these is an LLM explicitly designed to ingest a corpus of some 300,000 climate change research papers, culminating in a dedicated knowledge repository on climate change. 

“What we think is important is the ability to try different AI models, to fine-tune those models, to evaluate those models as an end-to-end process. And then to deploy and monitor those models in an integrated manner,” said Cheung.

“We want to use technology to share knowledge, as well as to enhance and augment human intelligence. And that's actually what IBM believes in—augmenting human intelligence with AI,” he summed up.

Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].​

Image credit: iStockphoto/Andy