00:12:43 · Aditya Agarwal, Sridhar Ramaswamy
[00:12:43] S1: Yeah. Yeah. [00:12:44] S0: Yeah. The opportunity was there for the taking. I'm sure most of you don't remember Sydney, but Sydney was way better than ChatGPT. ChatGPT would like spew out trash through much of 2023 anytime you asked a factual question. If you ask for a high query to do a great job, if you ask for a fact, it'll just like happily make it up. Sydney Bing chatbot had solved all of this. It had this fatal flaw that every interaction with it would take two minutes. They never fixed it. These are the people that are supplying capacity to open AI. They didn't see the opportunity. Bing itself makes 10 plus billion dollars. So they had a lot to lose by pushing on Sydney as the definitive search experience. But these are the people that actually created a much better search, not OpenAI, and they still blew it. So [00:13:34] S1: It's hard building new products. It's really hard. Before we kind of get into Snowflake a little bit, let's talk. I mean, I'm curious to get your take right now. We are... Is this kind of an interesting time to be an open source kind of like model right now? We obviously had GLM five two come out, which is getting great reviews. And in general, it feels as though that, yes, GPT 5.5 is exciting. Fable, whatever the... Whatever it is, you know, if I could use it, is sounds pretty cool. [00:14:02] S0: There's a line of thinking, by the way, that says that all model progress is just dependent on underlying hardware. And Fable and 5.6 are merely B200 models.
00:14:02 · Sridhar Ramaswamy, Aditya Agarwal
[00:14:02] S0: There's a line of thinking, by the way, that says that all model progress is just dependent on underlying hardware. And Fable and 5.6 are merely B200 models. [00:14:13] S1: Yeah, yeah, yeah. Know. So I guess my question coming to you is, what is your current take on the model ecosystem? Kind of like the leading lab providers like obviously Gemini, OpenAI sorry Gemini, ChatGPT, Claude, and then also obviously the six or seven kind of big open source models. Like how do you think about navigating in that world from your current position as the CEO of Snowflake? [00:14:38] S0: The Snowflake CEO, I mean, first of all, these have been recurring cycles of like frontier model labs, creating models that are significantly better along some variety of dimensions on the open source models- Catching up, yep. Catching up. I think, by the way, what is remarkable is that there's just a small handful of companies that can truly create frontier model. Still blows my mind. [00:15:09] S1: Still kind of crazy.
00:15:09 · Aditya Agarwal, Sridhar Ramaswamy
[00:15:09] S1: Still kind of crazy. [00:15:09] S0: Yeah. Yeah, I'm kind of like, what money can't buy stuff? No, it can't. I still find that just wild. I think definitely what is happening with the power of the Frontier Labs is that at least with lots of enterprise customers, there's a little bit of a worry of like, wait, all of us often you see like Satya had both a tweet and I think an article, Wall Street Journal, on this, where if you look at a world in which kind of all software disappears and these models are the interface between pretty much all of humanity and data underneath, let's say, it's a very different kind of world. [00:15:50] S1: Correct. [00:15:51] S0: So there's an amount of trepidation about being too reliant on a single model source. [00:15:57] S1: Right. Which is really kind of an intrinsically a trepidation about centralization. 100%. Yeah.
00:16:03 · Sridhar Ramaswamy
[00:16:03] S0: And so one of Snowflake's strengths has always been that we were the neutral data player when it came to the CSPs. There are tons of customers that will run their primary on AWS and run a disaster instance on Azure. And part of our selling prop has been your implementations are easily portable from AWS to Azure. In ways, by the way, all of you know this, you can't actually move software that way. This whole thing of cloud computing, blah, blah, in practice is just miserably hard to move any modest sized deployment from one CSP to another. There's just a whole lot of hooks that are super CSP dependent that most companies routinely fall trapped. And so part of our setup now when it comes to AI is that with both our kind of like our sort of Snowflake orchestrator, which is Coco, the coding agent for Snowflake, as well as what we think will be the work front end called co work, we are now able to say things like, well, you use it, we will make sure that they work just as well on both OpenAI and on Anthropic. And that's a selling proposition. And the open source models in many ways reinforce that because we have a team that's working on running our harnesses on top of the open source models. We have a very good inference team. We can run those models directly ourselves, Very different economics for us as Snowflake the company. So this is a pretty good setup overall. We can also do things like cost optimizations. One of our engineers came up with a very clever compilation technique for skills, which is almost deja vu all over again, which is you had code and then you had models and then you had English describing what to do. And then you go back to code for the most simple things that you're doing, but it's very cheap. It's like, wouldn't it
00:17:59 · Aditya Agarwal, Sridhar Ramaswamy
[00:17:59] S1: be nice if all of these skills were actually converted into simple code that you could then have essentially byte code that can run much more efficiently that a compiler could then essentially optimize? [00:18:09] S0: Percent, yeah. 100%. And so you see this cycle coming back. And so in that sense, it's actually pretty good news for companies like Snowflake. But it also, again, I think there are many fundamental lessons that will continue to be true. The harness is the new browser regardless of what anyone tells you. Being like buried under someone else's harness, some of us might have to do it. It carries a bunch of risk. And so we have consistently believed this is partly my consumer background. I have my co founder who also worked in a consumer background. So we place a lot of value on create products that directly touch the end users. And so that comes in very handy in moments like this.
00:18:50 · Aditya Agarwal, Sridhar Ramaswamy
[00:18:50] S1: If I can push you a little bit here to make a forward looking kind of prognostication about how do you think that the model ecosystem evolves over the next twelve to twenty four months? Do you think that... Or maybe even the next few years, do you think that there is enough of a concentration of capitalchipsmaybe research talent at the best three or four organizations, again, OpenAI, Anthropic, Google, maybe Meta, such that they will reliably have a six to twelve month lead? Or do you think that the gap between open source and closed actually starts to narrow? I don't actually have a point of view on this myself. I think it comes in cycles, you're right, that you kind of get a big closed model release and then open source kind of chips away and then you kind of get a big and we'll see how Fable does. Obviously, I think it's a little hard to say. So it's hard to know how big a jump Fable is right now. [00:19:45] S0: You either make a lot of predictions or you make none. So I think it's just hard to tell. I think what's been unique about like both chips and AI models has been the infinite demand for all kinds of different chips and for all kinds of different models. I think it's pretty unique to have a lot of jobs to be run on top of A10 chips. Like they're perfectly good. We used them six years ago at Neeva, five years ago at Neeva. They're still super useful today to the latest B200s. [00:20:22] S1: And I think their price per compute hour has actually gone up relative to same because there's enough kind of demand for it.