Building on IBM’s long-standing commitment to open-source technology, David Cox of IBM presented a strong argument for open innovation in business generative AI at VB Transform 2024. A vision that challenges and inspires the tech industry was presented by the director of the MIT-IBM Watson AI Lab and vice president of AI models.
Cox described open innovation as the story of human progress, defining the concept as crucial to technological advancement. Cox stressed the importance of the current period in AI development, noting, He thinks this moment is especially crucial because they all have to make decisions about where they want to invest. How do they intend to prevent lockdown?
All kinds of open
The IBM executive questioned the idea that openness in AI is a simple binary concept and instead presented a more nuanced perspective. Openness is more than one thing. Actually, Cox argued, it might mean a number of different things. He referred to the expanding ecosystem of open models that come from a range of sources, such as nation-states, universities, and even tech behemoths.
But Cox voiced worries about the level of transparency in many LLMs. He advised being cautious since sometimes you’re receiving something that’s more akin to a binary. You don’t know how it’s made; it’s just like receiving a bag of numbers. A fundamental tenet of open-source principles may be undermined, according to Cox, by this lack of openness, which can make it difficult or impossible to replicate these methods.
Cox compared these initiatives to more established open-source software and listed a number of traits that have contributed to their success. These include of regular security patches, planned release schedules, frequent updates, and engaged community involvement. He pointed out: Everything is properly defined since it doesn’t vary significantly from version to version; little additions can come from both inside an organization and the community at large.
LLMs: Open in name only?
Subsequently, Cox shifted his focus to the present situation of open LLMs, highlighting that numerous of them lack these crucial open-source characteristics. Even though open LLMs are wonderful, they don’t have many of these qualities these days, he noted. He expressed disapproval of certain corporations’ erratic release schedules, claiming that they should be free to discontinue “new generation models” at any time. Some model suppliers provide a model, but they never update it again.
This strategy, according to Cox, restricts the possibility of community-driven AI research and enhancement while falling short of real open-source principles. His observations call for more standardized, transparent, and cooperative approaches to AI development and push the AI industry to reexamine its policies regarding open-source models.
Cox used IBM’s Granite line of open-source AI models as an example of the company’s own initiatives in this area to support his points. As Cox noted, they fully share all of the model’s contents, highlighting IBM’s dedication to openness. They have actually made all of theie processing code open sourced so you can see precisely what they did to eliminate any objectionable information and filter it for quality. They will tell you exactly what’s there.
According to Cox, performance is not compromised by this degree of transparency. He said, “These are state of the art models,” as he displayed benchmarks contrasting Granite’s code model with other top models. Models that perform well don’t necessarily need to be opaque.
The gap in enterprise data
Additionally, Cox offered a fresh viewpoint on LLMs by describing them more as data representations than as conversational aids. This change in knowledge occurs at a critical juncture, since projections indicate that in the next five to ten years, LLMs will contain almost all publicly accessible data. However, Cox identified a big gap: the proprietary “secret sauce” of businesses is usually unrepresented in these models.
Cox proposed a mission to express enterprise data within foundation models in order to solve this issue and maximize its usefulness. Although retrieval-augmented generation (RAG) approaches are widely used, Cox contended that they are not as effective in utilizing the distinct knowledge, policies, and private information of a company. He asserts that it is crucial for LLMs to fully comprehend and take into account this enterprise-specific context.
Cox proposes a possible three-phase strategy for businesses, which includes selecting an open, reliable base model, developing a new business data representation, and finally implementing, growing, and adding value. He stresses how crucial it is to choose the base model wisely, especially for regulated businesses. Transparency is essential since it is required by many different businesses, including regulated industries, other industries where it is required, and many instances where model suppliers refuse to provide the data that goes into their models, according to Cox.
Integrating proprietary data with the base model is a challenge. Cox contends that the selected base model must fulfill a number of requirements in order to accomplish this. As a minimum, it ought to be extremely performant. Above all, it needs to be transparent so that businesses can completely comprehend what’s in it. It goes without saying that the model should be open-source in order to give businesses the flexibility and control they require.
Teaching AI your business secrets
Cox announced InstructLab, an IBM and Red Hat collaboration initiative that realizes his idea of merging enterprise data with open-source LLMs. InstructLab tackles the difficulty of integrating private company knowledge into AI models. It provides a truly open-source contribution mechanism for LLMs, as Cox characterized it.
The methodology employed by the project is based on a taxonomy of global knowledge and abilities, which allows users to pinpoint specific areas that require improvement in the model. The inclusion of enterprise “secret sauce,” which Cox pointed out is lacking from current LLMs, is made easier by this methodical approach. With InstructLab, domain experts can contribute to model customisation at a lower barrier by using simple examples or relevant documentation.
One way that InstructLab tackles the problem of combining proprietary data with base models is by creating artificial training data using a “teacher” model. This novel method adds enterprise-specific functionality while maintaining model performance.
Notably, the model update cycle is accelerated greatly using InstructLab. In contrast to conventional “monolithic, sort of one year release cycles,” Cox said, “We can even turn this around one day.” Because of their agility, businesses can quickly incorporate new data and modify their AI models to meet evolving business requirements.
It appears that enterprise AI usage is changing, based on Cox’s observations and IBM’s InstructLab. Personalized solutions that highlight each business’s distinct area of expertise are replacing generic, off-the-shelf approaches. Those who can most effectively transform their institutional knowledge into AI-powered insights may stand to gain a competitive advantage as this technology develops. A machine that understands your business as well as you do is what the next chapter of AI is all about, not just smarter machines.