The News
On Thursday, Microsoft announced two powerful AI models it built that it says perform at the level of the world’s top offerings, after years of backing others made by OpenAI.
The move officially pits Microsoft’s AI division, led by Mustafa Suleyman, against OpenAI and the rest of the industry, and helps it control its destiny in the high stakes AI race between tech giants and heavily funded AI startups.
“We are one of the largest companies in the world,” Suleyman said in an interview with Semafor. “We have to be able to have the in-house expertise to create the strongest models in the world.”
Microsoft unveiled MAI-Voice-1, a speech model that the company says is one of the most efficient in the industry, running on a single GPU and capable of producing a minute of audio in under a second.
It also announced a text-based model, called MAI-1-preview, that it says could help power future versions of Copilot, the company’s AI assistant that is woven into Windows and Office products.
Both models are geared toward cost-effectiveness. MAI-1-preview was trained on roughly 15,000 Nvidia H-100 GPUs, compared to models like xAI’s Grok, which was trained on more than 100,000 such chips.
Suleyman told Semafor the company used techniques, some gleaned from the open-source community, to stretch the capabilities of the models with minimal resources.
“This is a model that is punching way above its weight,” he said. “Increasingly, the art and craft of training models is selecting the perfect data and not wasting any of your flops on unnecessary tokens that didn’t actually teach your model very much.”
In this article:
Know More
Cost is one of the biggest issues facing AI companies. Hyperscalers like Microsoft, Amazon and Google are spending tens of billions of dollars each quarter to build increasingly massive data centers. Even as models become more efficient, consumer demand has only increased and newer techniques rely on ever more processing power to improve capabilities.
Suleyman told Semafor Microsoft AI is already working on the next versions of its models on some of the largest data centers in the world, equipped with Nvidia’s next generation of chips, known as GB-200s.
“We have an enormous five year roadmap that we’re investing in quarter after quarter. So I think that’s gonna continue.”
Suleyman helped co-found DeepMind, the AI firm acquired by Google in 2014. After leaving DeepMind, he founded Inflection AI. He was hired by Microsoft last year, along with much of the Inflection team, to launch the AI division there.
The move by Microsoft created tension in its relationship with OpenAI, which had agreed to exclusively run its AI models on Microsoft’s data centers. Its launch of its own high-end AI models could create even more competition between the two partners.
Suleyman downplayed the possible rift. “Our goal is to deepen the partnership and make sure that we have a great collaboration with OpenAI for many, many years to come,” he said. “I hope that continues. I’m pretty sure it will.”
Suleyman has long been an AI safety advocate and recently penned an essay warning against the dangers of “seemingly conscious AI,” that mimics human emotions.
He said Microsoft aims to remove qualities that make models seem like they have feelings and goals in the “post-training” process, which he likened to sculpting.
“We have to really answer the question as sculptors. What are the things that we are intending to do?” he said. “The purpose of my essay is to call attention to the fact that blind imitation of all human capabilities presents material risks, and now is the time for us to proactively think about those risks.”
Reed’s view
Microsoft is still early in its journey to build powerful AI models, so the world is unlikely to be blown away by the capabilities of the newly released offering.
But the move is probably crucial for Microsoft to stay competitive in the long run. Its Azure cloud business is one of the leading providers of compute that powers AI products like ChatGPT. But it is still dependent on Nvidia for chips and other companies for the models themselves.
It’s working to build its own AI chips, but those products will likely run more efficiently on models designed for them specifically.
In the end, the hope for Microsoft is that it can hold onto its Windows and Office distribution channels and use its own AI models and infrastructure to power them.
Room for Disagreement
Stratechery’s Ben Thompson argues that, while he’s still bullish on Microsoft, it has lost some of its AI advantage because of the fraying relationship with OpenAI, the stalled progress of Bing’s AI offerings and has lost the early lead its GitHub Copilot had over coding products like cursor. Its Office 365 franchise could also face potential disruption.
Notable
- Microsoft also unveiled in June an AI model it said was better than doctors at diagnosing complex health issues, though it was still far from ready to be deployed commercially.