Why is Microsoft creating a new model from scratch when it’s already a big investor in OpenAI and in the French startup Mistral? The tech corporation might be hedging its bets, given the regulatory scrutiny its current AI deals are undergoing.
MAI-1’s 500 billion parameters marks a larger amount than many models: Microsoft’s Phi-3 Mini was launched in March and just has 3.8 billion parameters, while Meta’s Llama 2 models has up to 70 billion parameters at last count.
Still, the number puts MAI-1 well below OpenAI’s GPT-4 and its reported one trillion parameters.
The model is under the auspices of Mustafa Suleyman, who previously served as CEO at AI startup Inflection. That startup sold its IP rights to Microsoft for $650 million a few months ago, in a deal that also saw most of its staff hired on at the tech giant.
However, according to The Information‘s scoop, MAI-1 is reportedly entirely built by Microsoft internally, and not a rebranded Inflection model — even if some training data and technologies may be carried over.
Before Inflection, Suleyman was at Google AI, yet another top AI big-business competitor. The shuffling around of AI workers may be reminicent of “Hunger Games but for GenAI companies,” as one LinkedIn commenter referred to this news.
