Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Would you like to insight in your inbox? Register for our weekly newsletters to only receive the company manager of Enterprise AI, data and security managers. Subscribe now
Decide AI models is also a technical decision and it is a strategic. However, the selection of open, closed or hybrid models are compromises.
During the speech at this year’s VB transformation, model architecture experts from General Motors, Zoom and IBM discussed how their companies and customers look at the AI model selection.
Barak Turovsky, who became the first Chief Ai officer of GM in March, said that there was a lot of noise with every new model publication and every time the ranking changes. Long before existing boards were a mainstream debate, Turovsky helped to start the first one Great -speaking model (Llm) and remembered the way open sourcing AI model weights and training data led to important breakthroughs.
“To be honest, that was probably one of the greatest breakthroughs that Openai and others contributed to starting starting,” said Turovsky. “So it is a funny anecdote: Open source has actually contributed to creating something that has been closed and may now be open again.”
The decision -making factors vary and include costs, performance, trust and security. According to Turovsky, companies sometimes prefer a mixed strategy – an open model for internal use and a closed model for production and customer fighting or vice versa.
IBMS KI strategy
Armand Ruiz, IBMS VP of the AI platform, said Own LLMSBut then it did not recognize that – especially as more powerful models came on the market. The company was then expanded to offer integrations with platforms such as Hugging Face so that customers can select an open source model. (The company recently made a new model Goal This gives companies an API to switch between LLMS.)
Other companies choose to buy more models from several providers. When Andreessen Horowitz examined 100 cios, 37% of the respondents stated that 5 or more models were used. Last year, only 29% used the same amount.
The choice is the key, but sometimes too much choice creates confusion, said Ruiz. In order to help customers with their approach, IBM does not make too much a question of which LLM they use during the Proof of Concept or the pilot phase. The main goal is feasibility. Only later do they begin to check whether a model is distilled or adapted to a customer’s needs.
“First we try to simplify all of this analysis parameter with all these options and to concentrate on the application,” said Ruiz. “Then we find out what is the best way for production.”
How Zoom approaches Ki
Zoom customers can choose between two configurations for his AI companion, said Zoom CTO XUEDong Huang. If the company’s own LLM’s own LLM with other larger foundation models. Another configuration enables customers who are concerned to use too many models to use Just Zoom’s model. (The company recently also worked with Google Cloud to take over an agent-to-agent protocol for AI Companion for corporate workflows)
The company has created its own small voice model (SLM) without customer data, said Huang. With 2 billion parameters, the LLM is actually very small, but can outperform other industry -specific models. The SLM works best for complex tasks if you work together next to a larger model.
“This is really the power of a hybrid approach,” said Huang. “Our philosophy is very simple. Our company is very similar to Mickey Mouse and The Elephant Tancing. The small model will do a very specific task. We do not say that a small model will be good enough … The Mickey Mouse and the elephant will work together as a team.”
