A lot of people are saying like, this isn’t Apple’s real AI models. They’re just releasing a teaser and they’re gonna release the good ones at their next big conference. And I’m saying, I don’t think you people understand how bad these mmlu scores when every competitor in the same range can also run on your phone is doing four times that. So here you have mmlu, the score of how good the ad model is. Here you have how big the AI model is. This is the trend over time. Okay, these stars are the new ones that just came out. They’re very good and they will run on your phone. Okay, apple is like down here, like worse than GPT two.
These models were also trained on 90% less training data than the current generation of Lama model. So a lot of people are saying like, maybe these are designed to be fine tuned. These are just base models. Llama is also a base model that was trained on 10 times more data. It’s, it’s significantly, it’s an order of magnitude smarter and it’s also free and open source. And they could have just used that. There is absolutely no utility or benefit to any of these models that apple just released. There are an, there are hundreds of alternatives that are better, that are also open sourced and free that they could have just used instead. It’s like shockingly comically garbage, the quality of what apple just released. And this is like their big project they’ve been working on. So this person’s like, why would they release this? I don’t get it. What is the theory? The thing is, that’s very weird right now is like AI is very, this whole like current AI renaissance is very new and like a lot of incumbent engineers who haven’t already worked on it like really don’t know what they’re doing. And it’s this bizarre moment where you have like a million teenagers who know how to do this really well, and then you have people working at stuffy incumbent institutions that really don’t know how to do it really well.
So I’m, I keep comparing this to llama, but there are actually much better things than llama now, and Facebook in particular has had a lot of trouble attracting and retaining talent, understands how to do AI. And obviously, I think my argument here, obviously apple is 2. Facebook’s new AI is nowhere near as good as databerks new AI. So Facebook’s old AI is over here. Okay, the new AI is here. So it’s really not, it’s not orders of magnitude better like these ones are, which came out at the same time.
So these are like small startups. This is from database, which is a small startup in the same building where I go to school. Actually know these people. It’s wild. I had no idea they were working on AI. Mixture is a similar startup that’s in Paris. Okay, so these ones are an order of magnitude better than the best that Facebook has been able to come up with as small startups. And then you have up here, this is the best Elon has been able to do, which is significantly worse than average. These kind of people like Zach and Elon are having a lot of trouble attracting and retaining competent talent that can execute. Cute on this level.
Because these kinds of people know that they are this hot commodity and they can work wherever they want, right? This is this very interesting frontier. And these people are not interested, right? Like these are boring companies doing boring things. They’re not pushing boundaries. They’re not even using the modern architect mixture.
So these are mixed mixture of experts models, right? Like this is this very exciting new concept of how to do AI. These are not, these are like the same technology as apple down here in the garbage can. These are old, stuffy incumbents that aren’t doing interesting work. So I think this is really the best that they can do. I think that’s my answer. I think that apple is having the same trouble attracting entertaining talent that Elon is, that Facebook is, and the smart people who wanna work on cutting edge stuff don’t wanna do it there. And this is why what they’re releasing is just shockingly, comically bad.