Deciphering Trade-offs of Open and Closed AI Models: An In-Depth Look at Strategies Adopted by GM, Zoom and IBM

The process of deciding on artificial intelligence models brings about a healthy blend of technical and strategic decision-making. Open, closed, or hybrid models all have their string of trade-offs, a matter which was vividly explored in this year’s VB Transform by model architecture experts from companies like General Motors, Zoom, and IBM. They offered their insider perspectives on how their respective companies weigh and decide on AI model selection taking into account factors like cost, performance, trust, and safety.

Barak Turovsky, the first-ever chief AI officer for General Motors, averred that new model releases and leaderboard changes often create a lot of ’noise’. Making reference to the time before leaderboards were mainstream, Turovsky reminisced about how open-sourcing model weights and training data led to significant breakthroughs. He emphasized the paradoxical trend of how open-source methods ironically helped create elements that went closed but might now be reopening.

Meanwhile, IBM’s AI strategy was expounded by the VP of its AI platform, Armand Ruiz. He recalled how IBM started its platform with its own large language models (LLMs), but later expanded to offer integrations with platforms like Hugging Face, enabling customers to select from a variety of open-source models. Addressing concerns of too much choice creating confusion, Ruiz noted that IBM aims to determine first the feasibility before looking at whether they should distil or customize a model according to specific customer requirements.

Over at Zoom, the company provides two configurations for its AI Companion, according to CTO Xuedong Huang. Customers have the option of federating the company’s own LLM with other larger foundational models, or, for those concerned about using too many models, they can solely use Zoom’s model. The choice, much like AI itself, remains constantly evolving and largely user-driven.