At VB Transform 2024, IBM’s David Cox championed the importance of open innovation in enterprise generative AI, drawing on the company’s long-standing dedication to open-source technologies. As the VP of AI Models and Director at the MIT-IBM Watson AI Lab, Cox presented a vision that both challenges and inspires the tech sector.
“Open innovation is really the story of human progress,” Cox articulated, emphasizing that this concept is crucial for technological advancement. He underscored the importance of the current moment in AI development: “We must make decisions about where to invest and how to avoid lock-in.”
Cox challenged the simple binary perception of openness in AI, stating, “Open isn’t just one thing. It encompasses many meanings.” He pointed out the emerging ecosystem of open models from various sources, including tech giants, universities, and even governments.
However, he raised concerns regarding the quality of openness in many large language models (LLMs). “In some instances, you’re merely presented with a bag of numbers, leaving you unaware of how it’s produced,” he cautioned. This lack of transparency, he argued, complicates reproducibility and undermines essential open-source principles.
Cox drew parallels with traditional open-source software, highlighting characteristics that have contributed to its success: frequent updates, structured release cycles, security patches, and active community contributions. “Everything is well-defined, allowing for incremental contributions from both companies and the broader community,” he noted.
He criticized the current state of open LLMs, stating, “Despite being impressive, they often lack core open-source attributes.” Cox pointed out the inconsistent release schedules of some providers, with models sometimes not receiving updates after their initial launch. This inconsistency, according to Cox, undermines the true essence of open-source and limits community-driven innovation in AI.
Cox emphasized IBM’s commitment to transparency through their Granite series of open-source AI models. “We disclose everything about the model,” he stated, highlighting that they have open-sourced all processing code to ensure quality and objectionable content filtering.
Cox argued that this level of openness does not compromise performance. He presented benchmarks showing Granite’s models as state-of-the-art, asserting, “You don’t have to sacrifice transparency for superior performance.”
He offered a novel perspective on LLMs, framing them as data representations rather than just conversational tools. With estimates indicating that LLMs will soon encompass nearly all publicly available data, Cox noted a significant gap: enterprises’ proprietary knowledge is largely unrepresented in these models.
To address this, he proposed a mission to incorporate enterprise data into foundation models to unlock its full potential. While techniques like retrieval-augmented generation (RAG) exist, Cox argued they often fail to leverage unique enterprise knowledge and proprietary information.
Cox outlined a three-step approach for enterprises: identifying a trusted open base model, creating a new representation of business data, and then deploying and scaling for value creation. He emphasized the need for transparency in selecting a base model, especially in regulated industries, noting that many model providers do not disclose their data.
The challenge lies in effectively merging proprietary data with the base model. Cox argued that the chosen base model must be high-performing, transparent, and open-source to provide the necessary control and flexibility for enterprises.
To bring his vision to life, Cox introduced InstructLab, a joint initiative between IBM and Red Hat designed to integrate enterprise knowledge into AI models. “InstructLab enables genuine open-source contributions for LLMs,” he explained.
The project employs a structured taxonomy of world knowledge and skills, allowing users to enhance model performance precisely. This structured approach facilitates the integration of enterprise-specific insights, lowering barriers for domain experts to customize models.
InstructLab employs a “teacher” model to generate synthetic training data, seamlessly blending proprietary data with base models without sacrificing performance. Notably, it accelerates the model update cycle to just one day, contrasting with the traditional, lengthy release cycles.
Cox’s insights and IBM’s InstructLab signal a shift in enterprise AI adoption, moving from generic models to customized solutions that reflect each company’s unique expertise. As technology evolves, competitive advantage may hinge on effectively turning institutional knowledge into AI-driven insights. The next chapter of AI isn’t merely about smarter machines; it’s about machines that understand business as intimately as their users do.
Ranking
