AI researcher switches from Microsoft to OpenAI

AI researcher Sebastien Bubeck leaves Microsoft and joins OpenAI. In recent months, reports have generally read the other way around, with numerous employees leaving OpenAI to join competitors or set up their own businesses. Microsoft is repeatedly suspected of having a great deal of influence over OpenAI. Several authorities in various countries have even investigated whether Microsoft actually controls OpenAI. However, they have not yet come to this conclusion.

Due to the close relationship between OpenAI and Microsoft, Bubeck had early and extensive access to OpenAI's Large Language Models (LLMs). He was involved in the implementation of GPT in Microsoft's Bing search engine. He should also be very familiar with all subsequent OpenAI LLMs. Bubeck sees his fundamental focus on the development of an Artificial General Intelligence (AGI) – as he himself writes on a public profile.

With this focus, he is completely in line with OpenAI's plans, which have set the development of an AGI as their primary goal. "We call our approach 'physics of AGI' because we are trying to uncover at different levels (parameters, neurons, neuron groups, layers, data curriculum, ...) how the parts of the system come together to produce the amazing and unexpected behavior of these models," writes Bubeck himself.

Working on small, local models

The switch to OpenAI was first reported by The Information. Bubeck has most recently worked primarily on Microsoft's Phi models. This is a small language model family that is optimized for mobile devices, among other things. OpenAI released GPT-4o-mini in the summer, its first and so far only model optimized for speed and cost efficiency. SLMs are an area in which OpenAI has some catching up to do.

A recent article from nature magazine shared by Bubeck is entitled: "Forget ChatGPT: why researchers now run small AIs on their laptops" In addition to the size of a model, it is also important that they can run locally, i.e. do not require a permanent internet connection. Bubeck is quoted in the article as explaining how Microsoft trained Phi-3. Training tasks were created specifically for this purpose, thanks to which the response quality could be significantly increased. "If you are able to create a data set that is very rich in these reasoning features, then the signal will be much richer."

So it seems logical that Bubeck will continue to work on small models at OpenAI. In recent weeks, numerous senior OpenAI employees have left the company. Many have moved to competitor Anthropic, and some have announced their intention to set up their own business. Co-founder Greg Brockman is taking a sabbatical in the middle of such a successful period. Mira Murati, CTO, was the last to leave.

(emw)