Less about Models; More about Architecture

Chaitan Gupta discusses the evolution of AI, highlighting the shift from small models targeting specific issues to addressing a broader range of challenges. He emphasizes the importance of making technology accessible, actionable, and safe for a wider audience. Gupta notes that Hitachi's focus on industrial AI, which began as a risky venture in 2016-2017, has become essential as the industry has matured, fostering confidence in internal talent and leadership.

The definitions of industrial AI and physical AI have evolved, initially concentrating on predictive maintenance and repair recommendations across various sectors. Gupta points out that, despite the differences among industrial verticals, there are commonalities in the problems faced and the techniques employed. He predicts that as industries accumulate more data and AI technology progresses, the significance of physical AI will increase.

Cultural differences in robotics acceptance are also discussed, with Japan being more receptive to robotics compared to North America, which lags in adopting robotics for elderly support and industrial applications. Currently, the center of gravity for industrial robotics is in China, while large language models are predominantly developed in North America.

Generative AI has democratized access to AI, enabling enterprises to engage with it despite challenges in model building and deployment. However, concerns about data sovereignty and potential intellectual property loss arise as AI becomes more operationalized. The conversation stresses the necessity for governance and assurance in AI model deployment to ensure safe and regulated behavior.

The architecture of AI systems must evolve to meet diverse enterprise needs and adapt to the rapidly changing landscape of model technology. Enterprises are encouraged to prioritize architectural thinking over specific model commitments to effectively navigate implementation challenges. Protecting intellectual property when using AI tools that rely on shared data is crucial, with Rackspace providing a private AI environment for secure operations.

While the complexity of AI architecture is acknowledged, it is framed within established design principles. Customers are advised to create an orchestration layer to manage multiple harnesses for different AI applications, recognizing that various harnesses may be necessary within an enterprise to effectively handle diverse AI applications.

Benchmarks serve as general guidelines but do not accurately reflect specific workloads, underscoring the importance of tailored evaluations for consistent customer experiences. An evaluation layer is essential within the orchestration framework to ensure that AI solutions align effectively with customer needs.

The architecture of AI solutions should be customized based on varying customer requirements, as there is no one-size-fits-all answer in AI architecture. Gupta emphasizes that building an internal AI solution enhances confidence when selling it externally, while also noting that governance, assurance, and orchestration are currently underserved areas in the AI landscape.

For AI solutions to be viable for enterprises, they must be safe, reliable, and cost-effective, according to Gupta's predictions. Rackspace is committed to developing AI that is sovereign and tailored to meet specific customer outcomes, advocating for a systematic approach to addressing AI challenges for successful implementation.