• 5 months ago
Is Apple Intelligence just a reskinned version of Gemini? Dive deep into the architecture of Apple's new AI system and its groundbreaking partnership with Google and NVIDIA.

Category

🤖
Tech
Transcript
00:00This whole system is not Gemini with a skin on it by any stretch.
00:05The beginning of that section was very much talking about Apple's new Apple intelligence
00:11architecture, especially now they have verbally confirmed, even if they only mentioned them
00:15once in the whole show, about this partnership with Google's Gemini.
00:18We actually got a really cool, deeper dive into that architecture right after the keynote.
00:24It really answered the question of what is Google's involvement here?
00:29And it sort of added the thing of, oh, yes, and NVIDIA is involved, too.
00:34So it takes a village to build the AI that you actually need, right?
00:39And for the models, it is a real collaboration between Apple and Google.
00:46There's this image that they put up, this schematic that they put up on the screen of all these
00:50new models that they have, both a large model that's local and then a series of models that
00:56are in the private compute cloud.
00:58They color-coded them, and the color-coding indicated that all the models were built together
01:03with Google.
01:03This whole system is not Gemini with a skin on it by any stretch.
01:09It is, Apple has a system orchestrator that it built sort of this whole system on top
01:15of the models to manage the queries, to parse them out, to figure out where they go with the
01:22apps and where they go with the data and where they go locally or if they go into the private
01:27compute cloud.
01:29What Google's had to do here is that it has its private compute cloud, but the models live
01:35now on Google's cloud, on NVIDIA's GPUs.
01:39And they had to work with both of them to build bespoke cloud systems that would replicate the
01:48private cloud compute that Apple already has.
01:51And it is very important to understand that what that means, private cloud compute, is a
01:57replication of the privacy that you have on your phone.
01:59That's the idea of it.
02:01And the other little piece of their architecture, which is important, because they went from
02:063 billion parameters to like 20 billion parameters on the on-device model, is that they found
02:14a way, when they have a query, to only access a small set, what's needed for answering that.
02:23So instead of loading all of the parameters into memory, which would be overloading and kill
02:27the battery, they only load a very small subset, and that's what handles the queries.
02:32It doesn't seem to be different versions of its AI, whereas with Gemini, you've got Flash
02:37and you've got Pro and you've got all these different things you can choose.
02:40On Apple, it's just Apple intelligence and there's the one model.

Recommended