The future of AI is open

At this year’s Think, IBM’s flagship conference, the company unveiled a slew of new technologies aimed at making life easier for businesses. In his keynote to close out the conference, IBM Research Director Darío Gil explored how AI has moved so rapidly from a field of study to a set of technologies affecting the lives of billions around the world each day.

According to Gil, just about every single piece of publicly available data has already been incorporated into training data for foundation models in use today. But he estimated that less than 1% of enterprise data has made it into AI models. The problem is that enterprises are missing what is truly amazing about AI by focusing too much on the “name brand” models themselves, rather than the incredible architecture they use to represent data. Additionally, much of their enterprise data is not currently in formats or places that make feeding it into models easy. To address this gap, IBM is making foundation models open and collaborative, helping businesses to capitalize on their data and trust the results. 

Gil issued a challenge to the audience: to start a new mission to move your enterprise data into a foundation model that you own and control. The goal is to create a transformative path of sustained business value while reimagining what’s possible with their most valuable asset, their data. “Don’t be an AI user — be an AI value creator,” Gil said.

To truly see the benefits that AI can bring to their work, there are three steps that businesses can take. And they could be the difference between simply implementing a piece of software, and deploying game-changing tools that could unlock the value of an enterprise’s most valuable asset, its data.

For Gil, that process starts with trust. Businesses must find a base model that they feel comfortable using — one with transparency around where its training data comes from, how its model is weighted, and the components of the model. Only with a solid foundation can you hope to safely and securely build something meaningful. “We’re going to add the data to it, so we’ve got to know what is in it and how it works,” said Gil. That’s a large part of the reason IBM has open-sourced 18 of its Granite models, under an Apache 2.0 license, including its code models, time-series, language, and geospatial models.

From this base, businesses can begin to build a representation of their data that can be fed into models to solve their most pressing issues. In his keynote at last year’s conference, Gil said the world had seen an explosion of foundation models. Since then, developments in representation have enabled new AI applications, for things like payroll, social media, and internet search. Once a business’s data is in a format that can be added to a model, they can deploy, scale, and create value with their own AI applications. “It’s not really magic — it’s just math, human ingenuity, and a lot of computing power, "Gil said.

To this end, IBM and Red Hat launched InstructLab, a new method that enables foundation models to learn incrementally, more like humans do. InstructLab gives anyone the tools to feed a model new data and create new abilities for the model, without the need to retrain it from scratch. IBM has put its Granite-7B language model into InstructLab, allowing anyone to add new skills and knowledge, making them more useful for the specific needs of a business — without losing anything the model has previously been trained on. After a quick quality evaluation, the updated model is ready to use.

With InstructLab, this whole process can happen quickly and transparently, with lower computing costs. To show how much quicker models can be fine-tuned, Gil said that when IBM’s Granite code models were being trained on translating COBOL to Java, they had 14 rounds of fine-tuning that took nine months. Using InstructLab, the team added newly fine-tuned COBOL skills in a week, requiring only one round of tuning to achieve better performance.

It’s an exciting time for the future of computing. But no one company will be able to shape the future on their own. Innovation happens best when everyone has the same playing field to start from. That’s partly why IBM helped create the AI Alliance. Gil said onstage that there are now nearly 100 institutions involved in the Alliance. As AI continues to proliferate into new areas, it’s key that new tools and code are developed with rigor, trust, safety, security, and diversity included from the outset.

“The future of AI is open — no matter what some say,” Gil said.