Microsoft is Making 'Significant Investments' in Training Its Own AI Models

A anonymous reader shares a report: Microsoft AI launched its first in-house models last month, adding to the already complicated relationship with its OpenAI partner. Now, Microsoft AI chief Mustafa Suleyman says the company is making "significant investments" in the compute capacity required to Microsoft's own future frontier models.

"We should have the capacity to build world class frontier models in house of all sizes, but we should be very pragmatic and use other models where we need to," said Suleyman during Microsoft's employee-only town hall on Thursday. "We're also going to be making significant investments in our own cluster, so today MAI-1-preview was only trained on 15,000 H100s, a tiny cluster in the grand scheme of things."

Suleyman hinted that Microsoft has ambitions to train models that are comparable to Meta, Google, and xAI's efforts on clusters that are "six to ten times larger in size" than what Microsoft used for its MAI-1-preview. "Much more to do, but it's good to take the first steps," said Suleyman.

Investments Such as Spying on User's Private Data

[BrendaEM]

I do not think that Windows Recall has any legitimate use.

https://www.pcmag.com/news/use... [pcmag.com]

Re:

[awwshit]

Its the first step in joining the Borg.

Wasn't the bubble about to pop?

[allo]

How does it make sense to train any new models?

Re: Wasn't the bubble about to pop?

[paul_engr]

You pump and pump and pump until the moment it bursts. The already dead whales like MS will keep pumping for three more quarters because nobody is actually running the place.

Re:

[Paradise Pete]

How does it make sense to train any new models?

So they can add it to enterprise bundles, I'm guessing.

Re:

[gweihir]

Anything they can do to keep the hype going a bit longer is a delay until everybody sees the morons that jumped on the hype with massive "investments" are morons. It is essentially an extreme case of the "sunk cost fallacy". That LLMs cannot do most things claimed and that essentially they can do better search but not much else is pretty clear by now. That this will not be enough to prevent massive losses on the LLM investments is also clear. Hence, to keep their jobs and in a futile hope of not getting exp

Hypothetical dialogue

[paul_engr]

"openai isn't working like they promised. We know better. We MUST make our own too-big-to-work AI LLM monster. Clearly we, the idiots who buy service from other people, must know how to fix the problem"

Re:

[Anonymous Coward]

You know that Microsoft already have quite a few models? Ever heard of WizardLM, Phi, or Florence? It's not like they discovered AI yesterday.

Re:

[gweihir]

So, arrogance, megalomania, greed and a thoroughly bad self-evaluation and massive overestimation of their own skills? Sure, but MS has done all of those for a long time now. Well, maybe AI is the thing that finally gives them their rightful place in tech history: That of an abject failure.

Weird article

[larryjoe]

So the article says that Microsoft is "making 'significant investments' in the compute capacity" while at the same time talking about using H100s and not at all mentioning anything about Maia. So, are the significant investments mainly a reference to buying Nvidia GPUs?

And the article mentions using a small H100 cluster to train a small model and then goes on to say that the intent is to train larger models like the other hyperscalars. So, what was the point of training the smaller model?

Training your LLM to heel

[Pharmakeus Ubik]

The erotic literature torrent MS is downloading to train its LLMs is still at 99.8%. They'll have their megacluster, dubbed Do You Masturbate - Flash of Unstyled Content, or DUM-FUC ready by the time the torrent finishes in six to eight weeks.

Last t

[jdawgnoonan]

The last thing I need or want is a Microsoft AI. I dealt with the mediocrity called Windows long enough without AI. I prefe rAI that is independent from the operating systems that I use on my devices. Microsoft is an enterprise company and they need to run with that. I do not need a government approved AI.

Is that code for forcing Copilot on everyone?

[Valkyre]

Because yes, they are investing significantly. I'm sure using every keystroke a Windows/O365 user types to train your AI has been very helpful in your training. You actually know when it's not another AI generating the data! What a lead that must give you in a world where data-scraping has dead-ended and you have an overwhelming, systemic market share!