Although Microsoft Corp. is said to own just under half of ChatGPT creator OpenAI LP, it’s also working on alternative artificial intelligence models that are less powerful, but also much more efficient.

That’s according to a report today in The Information, which cites unnamed OpenAI employees as saying Microsoft is concerned about the sometimes staggering costs of running powerful models such as GPT-4, which underpins ChatGPT.

In response to those concerns, Microsoft’s research chief Peter Lee has reportedly ordered some of the company’s 1,500 AI researchers to focus on developing smaller conversational systems that are less expensive to run. They will likely be less capable than GPT-4, however.

The report comes as Microsoft races to build AI features into most of its software products, including a Copilot for Windows that is based on GPT-4. But the company is worried that, with Windows having more than a billion users worldwide, the costs of running these AI features could mushroom. The company doesn’t want to stunt the economic benefits of its new AI offerings, so it’s now looking at lower-cost alternatives, The Information said.

Microsoft’s pivot to more efficient AI models is currently at an early stage, though the company is said to have begun testing internally developed models in services such as Bing Chat.

Microsoft’s head of search Mikhail Parakhin has previously said Bing Chat is 100% reliant on GPT-4 in its Creative and Precision modes. However, in its Balanced mode it’s using a new model called Prometheus, as well as its Turing language models. The latter are not as powerful as GPT-4: They can recognize and answer simple questions, but when they face more vexing questions, they’re designed to pass these on to GPT-4.

In coding, Microsoft recently unveiled its 1.3 billion parameter Phi-1 model, which was said to have been trained on textbook-quality data. It can generate code in a more efficient way, but it isn’t quite up to GPT-4’s standards, The Information said.

The company is also said to be looking at alternative AI models such as Orca, which is based on the open-source Llama-2 model developed by Meta Platforms Inc. Orca reportedly comes close to matching the performance of OpenAI’s models, even though it is smaller and less resource intensive.

Microsoft’s AI research division is said to have around 2,000 graphics cards from Nvidia Corp. at its disposal, and reportedly Lee has now ordered that the majority of them be used to train more efficient models that are focused on performing specific tasks, as opposed to GPT-4, which is more general-purpose.

“The quest for cost-effective AI is likely to become the second major trend in the industry, following on from the inital trend, which was all about getting the best generative AI capabilities out of the door,” said Holger Mueller of Constellation Research Inc. “Now, it’s all about trying to reduce the total cost of ownership of these generative AI models.”

Microsoft reportedly owns a 49% stake in OpenAI and has injected billions of dollars in funding into the startup. It’s also expected that it will continue its close collaboration with OpenAI for years to come.

However, it’s believed that Microsoft’s and OpenAI’s interests don’t always align. Despite being collaborators, the two companies are rivals as well, with their ChatGPT Enterprise and Bing Chat Enterprise products essentially competing against one another for the same target audience.

Photo: efes/Pixabay