AI models are moving beyond centralized providers

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 6:58

   All technology shifts go through a phase of diffusion, meaning they have to start somewhere. it's not like we turn on a switch and suddenly everyone can build an etsy store or a shopify store like that's not how the internet started with everyone suddenly benefiting from being able to be an entrepreneur online it took a couple of generations of technology to fusion before the idea of shopify and the high-speed internet all all over the place.

2. 7:30

   The country got everywhere, and people could actually stand up a Shopify store and run it. The first people on the internet, the first businesses that stood up, did very well. And so the technology started centrally. But then, you know, initially, people were like, oh, my God, Cisco is going to dominate the world, right? Like, Cisco's got the switches that make the internet switch. That's the technology that's going to, those guys that own Cisco, those guys are going to run the world. This is not fair. Like, they're going to control everything.

3. 7:56

   this is crazy nowadays it's NVIDIA to some degree it's google but like eventually every technology commoditizes that's what's so amazing about technology is it like it's always diffusing like this new innovation finds its way out like we've already seen in just the last couple of weeks this insane shift in ai where people don't have to run models in the cloud anymore they can run models on their desktop at home so there's no longer um like a dependency on google or a dependency on you know pick your favorite hosted model provider i can download an open source model and there's plenty of great models i can run it on a mac computer in my house and if you saw recently there's this auto research thing that went viral on twitter this weekend where Andrei Karpapi turned on auto research and he ran all of these agents on his computer

4. 8:49

   and they were just asked a bunch of questions to solve, make better LLMs. And they just ran 30 of them talking to each other and they just kept scoring the improvements they were each making to the LLM, to the underlying AI model. And they made a better LLM model than what ChatGPT had not too very long ago. In like a weekend on a computer at home, that's how quickly it's diffused. And so I'm like, I think that there's a, and so this whole thing of like, oh, data centers need to be stopped. I actually don't think that data centers are going to have much to do with the benefits we're going to realize. Like so much of AI is going to sit at the edge. It's going to sit in embedded devices. It's going to sit on your desktop computer. It's going to sit on your iPhone.

5. 9:27

   it's going to be ubiquitous it's going to be everywhere and i think everyone just doesn't see the benefit yet and so it's very hard to envision why i should do this and we can talk a lot about some of those benefits that could arise but over time all technologies have like a central feature where someone's making a bunch of money early on and then eventually everyone's like oh my god everyone's life has gotten better because of this thing like when the first car t therapies came out these are t cell i i totally jumped ship there but like car t therapy is this amazing technology that was developed where we could take T-cells, immune cells, out of the human body, program them to find a specific protein, put them back in your body, and then they go and find that protein and kill it, and kill that cell.

6. 10:12

   And it was used for cancer. So we could take a T-cell out of the body, program it to go attack a cancer cell, put it back in the body, it attacks the cancer cell and destroys it. When that first came out, it was like, oh my god, this is incredible. And a couple of companies made a couple billion dollars selling the first generation of those technologies. but to get that therapy is like millions of dollars initially you have to take all these cells out isolate them make sure they're clean engineer them put them back in make sure the person doesn't die so it became this very expensive initial process and now car t therapy is making its way into more and more cancer treatments and it is like almost a hundred percent like success rate in blood cancers when it works and so it's becoming this thing that goes from millions of

7. 10:54

   dollars to a million to 500 and pretty soon to 50k 20k and eventually 5k and that ends that disease that whole class of diseases goes away so all of these technologies start up with this aggregation of value small number of people get it small number of people get value but eventually all technologies diffuse and so i'm less like concerned about there being a monopoly we're already seeing every single model company getting disrupted by something else the next week um we're already seeing this idea of data centers being the requisite breaking apart there's a bunch of startups right now that are making technology that reduces the token cost by a thousand x so for every token produced which is a measure of output from ai it used to be 10 bucks or whatever you know pick your number a dollar 50 cents 10 cents it's coming down by a thousand fold because they're figuring out

8. 11:43

   better ways to architecture the underlying models to make distributed smaller models to use new chip architecture to use new systems of balancing energy across the different chips that you're using etc etc etc and these all compound so suddenly it's like the people that are going to build all these data centers really going to have a monopoly i don't know like this will allow anyone to stand up a small data center and run a bunch of ai stuff so you don't really need big 100 billion dollar data center so i think there's a lot that's changing very quickly and every step of the way that's it's happening so fast people just have all these reasons to have concern but But I'm pretty optimistic, as history being a predictor here, that the diffusion of these technologies will unlock value for every human on Earth.

9. 12:22

   And it's really just a function of at what point and at what point of value.