Stay updated with our daily and weekly newsletters for the latest news and exclusive content on the forefront of AI. At the VB Transform 2024 event, IBM’s David Cox highlighted the importance of open innovation in generative AI for enterprises, reinforcing IBM’s dedication to open-source technology. As the VP of AI models and a director at the MIT-IBM Watson AI Lab, Cox presented a vision that both challenges and inspires the tech industry.
Cox framed open innovation as essential to technological progress and stressed the critical nature of the present moment in AI development, urging individuals and companies to consider where they want to invest and how to avoid restrictive lock-ins.
Cox offered a nuanced view of openness in AI, debunking the idea that it’s a simple binary concept. He described an ecosystem where open models come from various sources like tech giants, universities, and even nation-states. However, Cox expressed concerns about the real quality of openness in many large language models (LLMs). He noted that the lack of transparency in how some models are produced undermines the core principles of open-source development.
Drawing parallels with traditional open-source software, Cox outlined key characteristics that contribute to the success of these projects: frequent updates, structured release cycles, regular security fixes, and active community involvement. He pointed out that true openness in LLMs is often lacking, with companies releasing models sporadically and not providing updates, which hinders community-driven improvements and innovation.
Cox illustrated IBM’s commitment to transparency with their Granite series of open-source AI models, which disclose everything within the models and the exact processing steps taken. He argued that high performance doesn’t have to come at the cost of opacity, showcasing benchmarks where Granite’s code models match leading competitors.
Highlighting a gap in LLMs, Cox proposed seeing them as comprehensive data representations rather than just conversational tools. He stressed the importance of including proprietary enterprise data in models to unlock their full potential. While common techniques like retrieval-augmented generation are useful, they don’t fully leverage enterprise’s unique knowledge and policies. Cox emphasized the importance of choosing a transparent, trustable base model, especially for regulated industries, to integrate with proprietary data effectively.
To support this vision, Cox introduced InstructLab, a collaborative project between IBM and Red Hat. InstructLab exemplifies Cox’s three-step approach to enterprise AI: selecting an open, trusted base model, representing business data, and deploying and scaling to create value. The project offers a genuinely open-source contribution model for LLMs, enabling enterprises to integrate their “secret sauce” through a taxonomy of world knowledge and skills. This approach lowers the barrier for domain experts to customize models with specific examples or documents.
InstructLab uses a “teacher” model to generate synthetic training data, mixing proprietary data with base models while maintaining performance. This significantly accelerates the model update cycle, allowing rapid integration of new information and adaptation to changing business needs.
Cox’s insights and IBM’s InstructLab signal a shift in enterprise AI adoption toward tailored solutions reflecting each company’s unique expertise. As this technology matures, the competitive edge will likely belong to those who can most effectively leverage their institutional knowledge into AI-powered insights. The future of AI isn’t just about smarter machines; it’s about machines that understand your business as well as you do.