With its new MAI models, Microsoft is drawing a line under its relationship with OpenAI - and signalling a different approach to training data

I have been snarking the breakup of Microsoft and OpenAI for a while; Microsoft having to parachute an adult onto the board two and a half years ago was probably the first step and the point where Microsoft was considering suing them over breaking the contract by hosting models on AWS last month was probably the last step.

Mary Branscombe's avatar
3w

wait is OpenAI on the verge of being in lawsuits with both Microsoft (sued by) AND Apple (suing) as well as being in court with Musk right now? huh

They had got to the conscious uncoupling stage by April:

Mary Branscombe's avatar
1mo

Microsoft is no longer paying OpenAI to run all its models there; OpenAI can run on the massively polluting data centres Oracle and others are building if they can stay in business that long without Microsoft’s money. Amicable divorce splits childcare responsibilities, endorses dating around

And then came Build, where the keynote was full of everyone except OpenAI, who usually had a regular slot to chat to Tony Scott - and it was in their back yard!

I had a few things to say about the keynote at the time, including some snark about OpenAI being MIA, even before they announced the new and updated frontier models:

Mary Branscombe's avatar
1w

Satya on stage at #msbuild looking really excited to be on a pretty tiny stage and much closer to a room full of devs than a lot of big tech CEOs get these days as he talks up the local agent capability of PCs even without NPUs

Satya wearing grey and black lifting his hands to chest height in front of a slide reading: security & observability agent runtime models context tools infra[structure]

Putting it all together for The Stack:

Mary Branscombe's avatar
6d

Microsoft creating its own 'thinking' frontier MAI model (and polishing up other frontier models it already has) promises hills and moats for customers on Azure and off, with less data, safety and governance risk - and maybe leaves OpenAI in the ditch too (come for the analysis stay for the pun)

Microsoft's new models give it a better moat

Microsoft's new models give it a better moat

With its new MAI models, Microsoft is drawing a line under its relationship with OpenAI - and signalling a different approach to training data

https://www.thestack.technology/microsofts-new-models-give-it-a-better-moat/
Mary Branscombe's avatar
6d

yes, Microsoft still uses OpenAI models - including for the frontier tuning it sells customers, if only to compare with its own frontier models, but it's ripping them out of its own products left right and centre and reusing the custom MAIA accelerators for its MAI models; very conscious uncoupling

Microsoft's new models give it a better moat

Microsoft's new models give it a better moat

With its new MAI models, Microsoft is drawing a line under its relationship with OpenAI - and signalling a different approach to training data

https://www.thestack.technology/microsofts-new-models-give-it-a-better-moat/

  • Microsoft

  • Satya Nadella

  • AI

  • frontier models

  • OpenAI (or not)

  • Build

  • keynote