Lenny's Podcast: Product | Career | Growth
00:15:34 · Lenny Rachitsky
[00:15:34] S1: I want to take a tangent and make sure people understand what Cisco even does these days. I think as a lay person, you'll think about Cisco and you're like, "Okay, they WebEx, yes. They make maybe some routers." You guys are key to this massive AI infrastructure build-out that's happening right now. You're a major player in this. I don't think people realize this, people listening to this podcast. Give us just a quick glimpse into how Cisco fits into this massive build-out and just what does Cisco these days?
00:16:03 · G2 Patel
[00:16:03] S0: Cisco is a critical infrastructure company for the AI era. What does that mean? But if you think about where the constraints are right now, if you think that AI is going to be one of the biggest movements, and then you ask yourself the question, "What could hold AI back?" There's three things where we feel like we can have a direct impact that can hold AI back. Number one is there's an infrastructure constraint. There's just not enough power compute and network bandwidth in the world to go out and satiate the needs of AI. Number two is, there's a trust deficit. If people don't trust these systems, they're not going to use them. And right now there's a lot of mistrust in these systems. Hallucination is a feature when you're writing poetry, but when you're trying to go out and run predictable systems, hallucination can be a bad thing. And these models are unpredictable, they're non-deterministic, and so they have to make sure that they have safety and security factored into them. And then the third area is a data gap. So far we've trained these models with human-generated data publicly available on the internet, but we are running out of human-generated data publicly available on the internet to train the models. And every company is going to differentiate based on their own proprietary enterprise data being used to train the models, synthetic data and machine data, which is where the most amount of growth is. And the third category of machine data we can play a massive role in at Cisco. So, what does Cisco do then? If you think about a GPU, which is what everyone now is very clear because of the great job that Jensen has done that here's what a GPU's core contribution is to AI. If these GPUs aren't networked together, you don't have AI, because it used to be that you could train a model on a single GPU, but then what happened was the model got too big to be put on a single GPU. So then you had a server with eight GPUs that got connected together. So, you could train a model with eight GPUs. But then that wasn't good enough. So, then what happened was you said, "I'm going to have a rack of servers that I'm going to network together." That at some point wasn't big enough. And so then they said, "I'm going to have a cluster of racks that are connected together." And that connected together is the operative word.
00:18:18 · G2 Patel, Lenny Rachitsky
That's what we end up doing is NVIDIA makes the GPUs and we connect those GPUs together. AMD makes the GPUs, we connect them together. And now what's happened, Lenny, is you have these data centers that might be hundreds of kilometers apart that need to operate like one coherent cluster, which means that they're completely in sync. Every GPU is in sync with each other when you're doing a training run. And that requires a very sophisticated set of technologies that we build to make sure that you could have two data centers, 800 kilometers apart, but boy, they run completely in sync with each other. And that's what Cisco does. We provide the networking, we provide the optics technology, we provide the safety and security technology, we provide the observability, we provide the data platform, all of those things together for making sure that we provide critical infrastructure for the AI era. [00:19:10] S1: So, being on the inside of this massive investment that is happening across the world, what do you think isn't being priced in into where things are heading into how much life will change, or just the scale of this build out?
00:19:25 · G2 Patel
[00:19:25] S0: Years ago I'd had a chance to meet with Ray Kurzweil. He's the chief scientist at Google for a while, and I think he still is. And he had talked about, he was writing this book called Live Long Enough to Live Forever. And so I was talking to him, I'm like, "What is the impact to human population if all of a sudden you can have 15 generations living simultaneously, because we have an indefinite span of life? Because now all of a sudden everything has to change. How does housing work, housing work? How does agriculture work? How does transportation work? How does ... Everything changes?" And he looked at me and he had the most profound answer and he said, "Most people can't think exponentially, because they always think exponentially maybe on a single dimension." But what ends up happening in these things is you can sometimes you have to keep in mind that exponentiality happens across multiple dimensions all at once. So, if you do have an indefinite span of life, you have to assume that humans are creative enough that they're going to find a way to have a three-day crop cycle. And they probably will have 5,000 story skyscrapers, and there will be a bunch of things in society that we have assumed are not solvable that'll now be solvable. So, when you go back to your question and say, what changes in this entire equation that has not been factored in well? I think today AI is looked at largely as a productivity tool and an aggregation mechanism. I have data all over and I'm going to be able to make sure that language can be used to compose the data in a way that I can give you, Lenny, the answer to the question that you're looking for that I think is the 0.0001% of the tip of the iceberg. The reality is, is we will have original insights generated that don't exist in the human corpus of knowledge, and we will have the physical world get augmented to language where capacity is augmented to humans. And what we have to be careful of is that that capacity is working on behalf of humans, but if that capacity is augmented to humans, you can now do things that you really care to do, and not do things that you don't care to do. And so our biggest realization that we had when we were using Codex, for example, when we were writing a code with OpenAI's kind of model and development tool, was the first three months we were screwing around with this.
00:22:05 · G2 Patel
And then there was this light bulb that runoffs. In fact, there was a former deployed engineer from OpenAI that told us about this, which is, "Hey, stop trying to think of this as a tool. Think of this as a teammate that got added to your team, and your framing will change, and the way in which you actually use the technology will change." And that essentially, if you compound that to how society operates, that's going to be pretty profound as an implication, while keeping in mind that these safety and security risks are non-trivial and they're real. And you can't be completely flippant about them, because how an AI identifies its own success and its own ambition will really matter. And we have to make sure that we actually keep guardrails around that, because it is in service of humans, it is not to go out and build a society by itself. And I do think that those are important kind of checks and balances you have to keep in mind. But the thing that people sometimes miss out in this very polarized narrative, which is we are either going to have nothing to do in society, or this is going to be completely useless as a piece of technology. I think that's not a helpful narrative. In fact, what is helpful is saying, "As we reconstruct society for the next phase, how can we make sure that life gets infinitely better? How can we make sure that diseases get solved? How can we make sure that poverty gets eradicated? How can we make sure that how people learn and find excitement and joy out of life gets compounded meaningfully?" If that happens, I think there's goodness that comes out of this.
00:23:55 · Lenny Rachitsky, G2 Patel
[00:23:55] S1: A line that I often think about is, Huang has had this thought that the best case scenario with AI, because he was a very AI doomer for a long time, and I think the reason he got leaned into AI is like, "I need to help steer this in a direction that isn't going to harm the world." The way he described it is, "The best case scenario for humanity is we're the house cat, where AI is just like, 'Okay, nice. Just keep sitting here with me and I'll take care of you.'" [00:24:20] S0: But by the way, the things that he is doing right now are nothing short of extraordinary. And for all the critique that one can have, the level of deep thinking that's going on with his company, it's just crazy. [00:24:36] S1: So, as you've been thinking about where things are heading, I've been liking to ask this question with people with kids, is there anything you're kind of shifting in how you raise your daughter, keeping in mind where things are heading? Are there skills you're trying to instill in her values you're trying to instill in her that will help her thrive in this future?