Tom Godden:
So let's pretend I'm a late adopter here, Jake. Give me some advice. How do I get ahead? How do I catch up? Swing for the fences, big home run idea? Lots of small incremental improvements? What's your suggestion?
Jake Burns:
It's interesting, from the outside looking in, when you look at companies that do spectacular things, it always looks like a big bang. It always looks like they had one fat bat and they swung and they hit that home run. What really great companies do is they fail a lot, but they make those failures really invisible. And the most important-
Tom Godden:
The power of the cloud.
Jake Burns:
The cloud helps you do that. Of course, yeah.
Tom Godden:
Come on. We had to. You were saying it before.
Jake Burns:
I thought that would be obvious. But really what they've done, really what they've mastered, and I think this is the critical skill, is they've mastered lowering the cost of failure. And that's really it. Because the problem is these big great ideas, the HAQM Primes and all of these great products and services that each of us use and each of us know about that look like they kind of just came out of nowhere and were just instantly successful, behind those are dozens, maybe hundreds, maybe thousands, in some cases even more, of very small failures. So you need to be experimenting all the time.
But the problem is if the cost of failure is what it is in a data center, for example, then it's very hard to afford those failures. So you can do very few of them, maybe none. And so that's why I say that reducing the cost of failure is the key to those big successes. Because it's not about having the smartest people. Of course that helps. It's not about coming up with the best ideas. Of course that helps. But what it's really about, it's about iterating as much as possible and being able to fail as quickly and inexpensively as possible, so that when you finally hit that home run, when you finally get that great idea that just works, that you can really double down and invest in that, and then show that to the world.
Tom Godden:
Yeah, I like to say I like to have my failures have decimals, not commas. So keep them small on this.
Helena, what's your suggestion on how to help customers view that catch-up play?
Helena Yin Koeppl:
Well, it's interesting. It's such a new breakthrough piece of technology that everybody's learning. So even what you view are the early movers, and actually they've learned from making a lot of experiments, and therefore, that actually a lot of mistakes too. At the beginning of this journey, for example, let's say 18 months ago, everybody thought having the newest, most powerful model is the most important thing. And what people have learned from that is actually that for different use cases, and you need actually different types of models, and some might be much, much smaller, but you might need to... Sorry. You might need to customize more with your own data. It depends on what are you trying to do and what are you trying to productionize and what impact do you want to have. So that learning can benefit, for example, what we could call the late movers to start already knowing that, hey, model choices are actually more important.
So there are many, many more. Actually, what we would like to share with our customers, as an example, and what we are sharing with them and what we are developing are really that full, comprehensive, full stack of three layers of solutions. And you need the computing power and you need the model choices and easy and secure and with easy guardrails to enable you to mitigate risks. And you need basically, if you want to just get going, and there are some application level to let you actually experiment with your data as of tomorrow. So there are many, many ways that you can catch up very quickly.
Tom Godden:
Yeah. And don't underestimate the value of small, sustained incremental improvements. And it also has the benefit of this is a technology, and this is true probably with many, but it feels more the truth with this one, you learn by doing. So, yeah, iterate, experiment, do it on the cloud so that you can have those low-impact mistakes so that you can learn from them and you can add value.
We've all had this experience I'm sure in our careers, it's easy to build version one. You slap it out there, you put it out there, it's all good. Now I need to sustain it. I need to operate it. I need to do version 1.1 or version 1.2. How do I build that infrastructure around it? How do I look at generative AI and find a way that I can build it in a scalable, flexible fashion so that I can actually live with this and not just have that 1.0 type of release?
Jake Burns:
Well, we've been saying even long before cloud, that version one you always throw away. Version one is just practice, right?
Tom Godden:
It normally deserves it. Yeah.
Jake Burns:
But I would go much farther nowadays. When you've gotten to become a high-velocity enterprise, when you start thinking in the economies of speed, I think you throw away the first 100, throw away the first 200. You're no longer counting each individual experiment because you're doing so many of them. And so I think it's really about getting to that point, getting to where the cost of failure is that low when you're experimenting at scale to where you can't even count all the experiments.
And then, to your point as well, learn by doing. Because I think the longer you're in the analysis phase, the longer you're in the planning phase, you're not really learning much. And as we all know, you're going to throw that plan away anyway because it's going to be wrong. So there's some merit in doing some planning, but try to get through that process as quickly as possible, and get to the doing as quickly as possible. And then by doing, you'll figure out the right way. And by the way, the first way you try, it's not going to work very well. Second way, third way, fourth way. But eventually you're going to gain real expertise, and that's where real expertise comes from. Combine the theory with the practice, but don't forget the practice.
Tom Godden:
Yeah. Helena, thoughts?
Helena Yin Koeppl:
Yes. Additionally, we are talking about AI machine learning. It's always learning. So the first version, get out there and continuously with new data and with basically MLOps and managing that drifts coming from new data. But at the same time, you know that with reinforcement learning, with reinforcement... With reinforcement learning with human feedback, and all of this can help you to make the result better. So the most important thing is get started. And like I said, machine learning is always learning.