Marco Arment

The other danger about this, by the way, is they're like, that's not how it works in my brain when I do the problem. Therefore, this thing is not thinking. And that's always dangerous because even though we think we know how our brains work, we are often very, very wrong. And we don't actually understand. We don't actually understand everything about how our brains work either.

The other danger about this, by the way, is they're like, that's not how it works in my brain when I do the problem. Therefore, this thing is not thinking. And that's always dangerous because even though we think we know how our brains work, we are often very, very wrong. And we don't actually understand. We don't actually understand everything about how our brains work either.

So it's the example I always give is like.

So it's the example I always give is like.

a it doesn't really matter how our brains work uh airplanes don't work like birds but they still fly really well right and b we don't always know how our brains work so you can't just make assumptions like that by saying well i know lms don't think because they work nothing like how i think my brain works when i think of this problem i don't do anything like that what the lm is doing as far as i know therefore lms don't think you gotta test it so a plus to apple people i

a it doesn't really matter how our brains work uh airplanes don't work like birds but they still fly really well right and b we don't always know how our brains work so you can't just make assumptions like that by saying well i know lms don't think because they work nothing like how i think my brain works when i think of this problem i don't do anything like that what the lm is doing as far as i know therefore lms don't think you gotta test it so a plus to apple people i

to doing what I think most people would consider a boring and pointless paper. But you need to do these things. You need someone to actually test the things that everyone just assumes are true to try to show that they are or aren't. So kudos to Apple for confirming something that I already believed. Wow. Or trying to. Again, more papers will follow, I'm sure.

to doing what I think most people would consider a boring and pointless paper. But you need to do these things. You need someone to actually test the things that everyone just assumes are true to try to show that they are or aren't. So kudos to Apple for confirming something that I already believed. Wow. Or trying to. Again, more papers will follow, I'm sure.

And related to this, there's a series I always link to whenever we talk about LLMs is the three blue, one brown. They're called courses. I keep looking for playlists, but they're called courses somehow on their YouTube channel. There's a course on neural networks and chapters like five, six and seven are about LLMs.

And related to this, there's a series I always link to whenever we talk about LLMs is the three blue, one brown. They're called courses. I keep looking for playlists, but they're called courses somehow on their YouTube channel. There's a course on neural networks and chapters like five, six and seven are about LLMs.

I'll put a link in the show notes to I think what is the most recent video in that series called How LLMs Might Store Facts. that tries to explore, like, given how we know how LLMs work, see previous videos with just a bunch of numbers and matrices, how does the information in them stored? I think the example they give is like, Michael Jordan plays blank.

I'll put a link in the show notes to I think what is the most recent video in that series called How LLMs Might Store Facts. that tries to explore, like, given how we know how LLMs work, see previous videos with just a bunch of numbers and matrices, how does the information in them stored? I think the example they give is like, Michael Jordan plays blank.

How does it come up with basketball, right? Where is that information stored in this giant matrices of number? How does that even work? Which is an interesting question because we're like, well, I know how it's stored in my mind. I just know that he plays basketball. Everybody knows that. It's so easy, but... Now I just look at these giant arrays of numbers.

How does it come up with basketball, right? Where is that information stored in this giant matrices of number? How does that even work? Which is an interesting question because we're like, well, I know how it's stored in my mind. I just know that he plays basketball. Everybody knows that. It's so easy, but... Now I just look at these giant arrays of numbers.

Where the hell is that information in these numbers? This video tries to explain it, and also it's, as you note from the title, how they might store facts. It's actually kind of difficult to tell because we're not particularly good at reasoning about giant piles of numbers. So take a look at that if you're interested.

Where the hell is that information in these numbers? This video tries to explain it, and also it's, as you note from the title, how they might store facts. It's actually kind of difficult to tell because we're not particularly good at reasoning about giant piles of numbers. So take a look at that if you're interested.

Is Google in favor of that? Do they want to be broken up?

Is Google in favor of that? Do they want to be broken up?

I know you're surprised. The response is like, here's how the world will end if you make us split off Chrome, Android, or Google Play. It'll be bad for business. It'll be bad for you. Everyone will break out in a rash. Dogs will howl. Like, it's just, you know. I mean, it's so hard for me to tell because these things take so long to wind through the system or whatever.

I know you're surprised. The response is like, here's how the world will end if you make us split off Chrome, Android, or Google Play. It'll be bad for business. It'll be bad for you. Everyone will break out in a rash. Dogs will howl. Like, it's just, you know. I mean, it's so hard for me to tell because these things take so long to wind through the system or whatever.