AI adoption has more than doubled since 2017 and its evolution leaped forward in 2023 with advances in large language models. These advancements hold the potential to unlock trillions of dollars in economic value over the next decade. But, implementing AI for business has unique requirements – it needs to be open, properly governed, designed for the enterprise and most importantly, create value.
During a keynote presentation at ITEXPO 2024, Kate Soule, program director for generative ai research, IBM, discusses how to put AI to work for the enterprise.
“I think of generative AI as the movie, ‘The Good, the Bad the Ugly,’” said Soule. “Sure, it is good, but we need to look at the bad and the ugly.”
First, Soule looked at the good.
LLM can do things people thought unimaginable a few years ago. It can create photos, write poetry and plan trips. It is endless. For enterprise, it introduces a new paradigm of data-efficient AI than what many have used in the past.
“LLM use cases for financial services include improved CX, better compliance, better risk management and modernized systems,” said Soule. “However, there are growing concerns. LLM costs are growing at an unsustainable rate. We are not even sure how big they are at times. The bigger the model is the bigger the cost is.”
On that note, Soule switched gears to look at the ugly.
“There are dark corners on the internet that should not be used for an internet task,” said Soule. “There are steps to make models safer. These techniques rely on humans to go through and sort through disturbing content. That labor is getting outsourced into digital sweatshops.”
So, what is IBM (News - Alert) doing about it? IBM Granite.
IBM’s AI is responsible and governs. It starts with data acquisition, goes to dataset preprocessing (model agnostic) and finishes with data preprocessing (model specific). IBM’s Vela is a cloud native supercomputer for the foundation model age.
IBM’s AI is designed for the enterprise and targeted at business domains. Small, specialized models can outperform large, generalist models.
IBM’s AI is also transparent and designed to support a broader ecosystem.
“We share data sources in our model, the pipeline, all of our evaluation results. We are as open and transparent as we can be,” said Soule.
Soule also announced that IBM launched AI Alliance.
“Can combinations of smaller LLM models outperform one big LLM?” Soule asked the audience. “Even though big models do best on average, smaller can do as well if not better for a given task. But how do we choose which model to use? We have a report that can predict which model to choose from.”
The MIT (News - Alert)-IBM router predicts best model for each task in real time.
"Benefiting from an open ecosystem of models and model creators is why the Watsonx Platform is being built,” Soule said in her closing remarks.
Edited by
Greg Tavarez