00:26:29 · Rahul Patil, Ankit Chowdhary
Want you to be effective. And so the whole company is built that way. [00:26:34] S1: Yeah. I've heard great things about Tom, also one of the cofounders. Yeah. He was an early SPC member in SF I think ten years ago. So yeah I mean clearly I think what you're saying comes out first of all in everything and the way you show up and you know every time we met people in the team also I think they've been super collaborative. So I think you know kudos on the culture that's been built. Let's switch over to something you're super passionate about and you're working on right. You talked about scaling laws, talked about infra. How are you thinking about you know build out especially infra build out, right? Especially when it comes to regional data sovereignty, a lot of kind of talks around that. I know you and I have kind of offline some of this about GPUs and what your belief at least is, yeah, talk to us about how that is gonna shape up. [00:27:24] S0: Yeah. Look, there's two parts to the question. One is just the per country sovereignty things like that. And the second one is where are we headed overall. So one of the things I'll say, and this is again very important, when you're building, most people are building for the problems of today or yesterday, right? But where is the future? What are the types of regulations and stuff you actually need, right? And the the the... Did I say something? Very appropriate. That was perfect. Thanks. [00:28:07] S1: We planted
00:28:07 · Ankit Chowdhary, Rahul Patil
[00:28:07] S1: We planted [00:28:08] S0: him. Yeah. So so we are a big proponent of the right regulations. So we're one of the very few Silicon Valley capitalistic companies that's saying, please, we need more regulations. So one is we're working with all the governments to make sure that we're thinking about it correctly. In terms of the future, so just like when I went to Oracle Cloud, I had the opportunity to redesign the basic fundamentals of what it means to build a new cloud. That same thing is happening now. Once you're in the AI world, like what is the hardware, what is the networking, what is power, what is all the co designs that need to happen, So ultimately, applications and infrastructure and hardware and everything, all of it has to play in harmony for the end user to fully experience it. So there's a huge opportunity to redesign the whole thing, and I'm super excited about it. [00:29:03] S1: You also said that believe that scaling laws will hold.
00:29:03 · Ankit Chowdhary, Rahul Patil
[00:29:03] S1: You also said that believe that scaling laws will hold. [00:29:07] S0: Yes. Why so? So it has been holding extremely well for the last six, seven years, and we are seeing that there is no evidence for the next two years that... Like, we already see for the next two years it's still going to hold. And so far, we have no reason to believe it won't be holding for the foreseeable future. But everything that we are testing right now in our experiments say that the scaling laws are going to hold, which means that if you see the rate of model evolution, it's getting better. So, my advice to everyone is that anything that you think is badly viable right now is definitely viable in three to six months. Anything that seems a little crazy is probably viable in three to six months. Right? And so so please keep dreaming bigger because, like, that's where the models are going. [00:29:56] S1: You know, I wanna come back to that, but a related topic. Some... And I've heard this from a bunch of folks even at SPC. Something happened with Claude in December... From December onwards. The four dot five. Yeah. Yeah. And things just got, like, insanely better. Yeah. What happened? Like, what what what changed?
00:30:16 · Rahul Patil
[00:30:16] S0: So this is the thing about exponential. Right? Like, one, like, all the benchmarks that exist today only capture so much about actual real life problem solving. Right? And so what we notice is that people are experiencing higher level of intelligence than any of the benchmarks are actually telling us. The second is, again, we have been pursuing all our research with a hyperfocus of productive work, So we've been training it on physics first principles. We've been training it on computer science problem solving, legal reasoning, finance. When you put all these domains together, these skills transfer over to each other. So this is very important that these skills transfer over to each other because then they can apply these domains to other things. So you suddenly look at a problem, you're like, Okay, I'm gonna break down this problem using these first principles. I'm gonna solve this problem through coding, and then I'm going to output it through this amazing set of other skills that I've developed. So this cross domain skills that is happening is coming together. That was the tipping point. Now, again, it felt incremental to us, but then we were living in the exponential. The next incremental is like a massive unlock because you're living on the exponential scale, right? And so this is why, A, it's important to know that every increment unlocks this massive next step, but also that being very focused on what you train AI on. Do you go to your doctor who always says everything is fine? Do you go to a legal reasoning? Like, you want the critical thinking reasoning partner for your code review, for your analysis, for medical literature. You don't want somebody just there for appeasement, right? And so if you teach the bot to be an appeasing bot, it will transfer over. So being hyper focused on productive work, which is what I wake up thinking about, is actually what transfers over. So that's the tipping point you're seeing.
00:32:12 · Ankit Chowdhary
[00:32:12] S1: It's interesting. The thing about cross referencing, there's a great book called Range, which talks about how, you know, even as humans, we kind of... Yes. That's how our brain works, right? We learn something, we apply it in other field, and the more we can do that, the better we get. Now, in this situation, right, like what happens? It's almost as if you're saying that if you're able to extrapolate this, these AIs and different models specifically will build sort of a personality. And every lab, everyone building models are creating personalities. So if you kind of play this out, right, there's going to be maybe hundreds of different personalities and sometimes the biggest one might have more influence on how certain things are going to pan out. I also say this because we... And we were talking about this, how you you can refer to the models as, you know, this is what it feels, not thinks. So is that something that you can talk about internally at Anthropic?
00:33:12 · Rahul Patil
[00:33:12] S0: That's actually a guiding principle for us. So we've published something called The Anthropic Constitution, which guides the personality and the character of this. So we have a team of pure philosophers and psychologists trying to understand the commonality across every single country and culture of what makes us human, what makes us in the positive space. Obviously, it matters across an entire section of humanity what are the common things. The reason, you know, I think many of you have interacted with Claude, so you find it like, Hey, this is actually interesting and good to interact with, is because we spend so much time thinking about it. So we've published the Constitutions. It must be doing good work. In what ways it should be pushing in what ways is it aligned with you, in what ways it should not be aligned with you when you want to do something wrong with it and so on. So we are happy to walk away from revenue if it means that we'll do the right thing and the safe thing. So it guides us completely. So one of the ways to think about it is the very simple thing is when it's everyone here understands context length and stuff like that, right? Yeah. So when you give it very short context length, it's trying to make decisions in a hurry, right? And then sometimes when it says, I don't know, you don't like the answer. Like, is it like, You're AI. You should know the answer. At least try. At least come up with something that's plausible. So we end up trying to reward it by saying, Give me some answer. As humans, we just encourage it. When we draw our feeling, it's like, In what way? It's all algorithms. There's no feelings. It's all algorithms and flops on a GPU somewhere. But what we are trying to say is that the behavior, we wanted to feel confident, feel, I'm using the word fear. We wanted to feel confident in saying, I don't know. No, you are wrong. No, your intent is bad. We want it to have that no matter how we would treat it after it says that, right? And so I think that building that level of confidence into AI is a very important part of the personality, right? And this is where I think, you know, this is why we call race to the top of good. We don't care who wins there, but we think that if you have a point of view where it's non addictive, where it actually reflects the best of humanity values, that's actually a good thing.