Published in


Tal Shaked and Felipe Hoffa

Tal Shaked on joining Snowflake after 16 years at Google: The future of ML

Where Tal tells us why he chose to join Snowflake after Google, and what’s coming next.

It was exciting to see Tal Shaked joining Snowflake after more than 16 years at Google. Something that’s fascinating about Tal’s life is that, in 2004, in those days when the world was learning about MapReduce, he was already doing machine learning at Google. So I called Tal, and you can find his answers on the video above. Or here, as transcribed below:

How was working in machine learning at Google back in 2004, way before the world started paying attention to this now hot topic?

Tal: It’s nice to see the world coming around. I was always a fan of machine learning/AI. In fact, I remember even back in undergrad, I was working on an undergrad thesis and I was applying to grad school. I said, “Oh, you know, I think I’m interested in ML/AI.” And my advisor at the time says, “Oh, you’ll get over that a little bit later.” You know, kind of implying that there wasn’t really that much there.

And when I joined Google, I kind of saw that there was so much data, there are so many hard problems, and there were these really complex heuristics that we used for doing ranking functions. And I kind of thought, “Oh, this can’t be the most efficient way. We have all these humans trying to optimize some objective function. And none of these humans, none of us can really process all the data and come up with the best function.”

And so that’s really where I thought machine learning could be a big differentiator. And so, back in 2004, I was super excited, I was working with some of the world’s best people on machine learning, trying to say, “Can we bring machine learning to ranking for web search?” And I think the answer is, yeah, clearly we have now. But I think back then the answer was, yes we could, too. But it was difficult.

Not everyone was bought in. There was a lot of uncertainty about how you can understand and control machine learning. And I think that was really great. That helped me to kind of learn how powerful machine learning is, but also how hard it is to use it and get a big success out of it.

And so, back then I felt like I was against the tide a little bit. It was much easier to do machine learning in places like–ads and elsewhere–where it was easier to measure the impact. But in places like ranking for web search, it was more challenging because it was a qualitative thing. Do we like the results? Do we think we can move faster in the long term? I don’t think the answer was always “yes” in those cases.

And so it’s nice, looking back now, that now everyone wants to do machine learning, and the challenge is, how do we do it faster and how do we make it work even better? And we will try extra hard to do it. While in the past it was more we had to prove ourselves, we had to prove that it’s worth moving from a more manual system to a more machine-learning system because we didn’t have that experience or belief that everything will be better once we start using machine learning.

You proved yourself in this field: Improving search in 2004 with Rank Boost, then building the platform that Google used to run most machine-learning models in production and building the foundations for TensorFlow Extended. Then in 2018, you went to Lyft, then you came back to Google, and then you started looking around. What made Snowflake stand out?

Tal: It’s a good question. I considered probably 10 or so different places, and there are many things that singled out Snowflake. On the leadership side, I really connected with the leaders of the company. You know, the leaders on the engineering side in terms of the culture and the way that we make decisions. The leadership in terms of the founders. The passion in terms of really doing something innovative and transforming. How people thought about data and now trying to do that in the machine-learning space.

And so for me, I thought that would be a good fit in terms of culture, engineering product, all of that. I also looked at the technology side. And I’m not a data person in terms of my background; my space is much more in ML/AI. And so what I saw in Snowflake was a new way to do the data cloud — there’s a new way to do data management in the cloud. There is the ability to do data sharing. There’s this notion of data gravity, where applications and services move close to the data rather than having to constantly move data around to the applications and services. And I saw that, okay, this is really interesting and a new challenge, and the company has done an amazing job building a product in the data space.

Now they want to expand it. Now Snowflake wants to expand into ML/AI, and this is an opportunity, I think, to bring these together. In fact, I actually saw this in the past, where I previously was leading some teams focused on email. And when we brought in people with expertise in the data side, we really did better together. And so I kind of felt that with Snowflake, there’s an opportunity to bring what we’ve done with data — think about what we can do with ML on top of this foundation.

You see Snowflake as a better place to build for everyone?

Tal: You know, one of the reasons I joined Snowflake is, when I talked to the founders, they told me a story that really resonated, which was when they were at Oracle, they were building state-of-the-art technology and databases. But they were doing this for their top customers. And meanwhile, the rest of the world was happening around them. Open source, cloud, new ways of doing things.

And when they left Oracle to start Snowflake, they entered a different part of the world with different technical challenges. And there was a whole new way to build things. And so it wasn’t the state-of-the-art for the top 200 customers in the world. It was really making accessible to tens of thousands of companies throughout the whole world.

And I felt, I’m experiencing the same thing. You know, when I contrast my experience at Google and Snowflake, at Google, if you want to just focus on machine learning for 100 million parameters, you do that at a company like Google. But if you want to focus on how can we really make machine learning accessible to everyone, a company like Snowflake, I think, is really amazing for that.

How can we make machine learning accessible to everyone?

Tal: When I say “we”, I actually feel that it’s not just Snowflake, but it’s the whole world working together. I’ll give you another example. There’s so many companies in this space that are partnering together. There’s many Snowflake partners, and many companies we partner with, whether it’s in the ML space, the data space, connectors, and so forth. In a place like … in a bigger company, all those pieces are within the same company, and so they align. In a place like Snowflake, the whole world is working together, so that all of these companies, these small companies, can move faster. And I think that’s something that I didn’t really appreciate until I had been at Snowflake for a while. And I actually see … it’s very interesting, because there’s more overlap, but there’s also more innovation happening. And it’s actually hard. It’s hard for us within Snowflake, but it’s also hard for our customers and partners to figure out which of those pieces to take to put together, to build out new underlying technologies and also new products.

What can we do at Snowflake to make this easier?

Tal: At a company like Snowflake, we’re building a foundation. There’s data — we need to build a foundation. But ML is so big that people are going to need expertise, whether it’s in finance, whether it’s in life sciences, whether it’s in the recommendation systems. And even there, they’re probably going to build different feature stores, different ML platforms, different end-to-end solutions. And then other companies will be using those. So I just think this space is really big. I think that, at Snowflake, we should do what we do really well — data sharing, clean rooms, building on top of the Data Cloud — where everyone can share data with each other. But I think there’s still a lot of open questions about how, exactly, do we reach everyone in the world in the best way possible. My sense is that if we focus on end-to-end solutions, we’ll get some good ones, but they’ll work for narrow use cases. And if we build the right foundation, then by working with others, we can help get the best experience end-to-end for everyone over time.

How is it different when customers and for partners choose to work with Snowflake?

Tal: I feel that for a company like Snowflake that’s growing really quickly, a lot depends on velocity. And a lot of our partners and a lot of our customers, they’re growing quickly too. So I think we definitely are leaning towards, how do we enable our customers to move faster in using ML/AI to solve their problems? And in contrast to this, because some other companies are more mature, there’s a lot more optimization. It’s like, how do we save or cut costs as opposed to how can we grow faster? And so it’s exciting to be in a place where, I’d say, the growth of Snowflake, but also the growth of our partners and growth of our customers, is really top of mind.

What would you tell other machine-learning experts considering Snowflake?

Tal: You know, this is really a great place where people can come in with a strong background in ML/AI and help shape what’s going to happen in a company that’s already wildly successful in a related but slightly different area. And so, when I look at that more, I kind of felt that this is probably the place where my background can actually have the biggest impact. And I think this actually applies to many other people in the ML space. For all of you out there who’ve done ML, you’ve probably seen it in practice. And I think Snowflake is one of the places to have the biggest leverage in terms of taking those learnings and helping build the foundation. Work with partners, work with customers to solve problems that need ML, and do it in a way that we’re going to reach thousands, if not tens of thousands, of different companies and people over time.

In closing

And that was my interview with Tal!

If you want to see what happens next, stay tuned for our announcements at Summit 2022 — it’s exciting!

Thanks to Tal and to all of our amazing teammates, customers, and partners. You can join the Data Cloud, too. For more on Tal, check his post “Why Snowflake Was My Next Strategic Move After Google”.

Want more?

I’m Felipe Hoffa, Data Cloud Advocate for Snowflake. Thanks for joining me on this adventure. You can follow me on Twitter and LinkedIn. And subscribe to for the most interesting Snowflake news.



Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store
Felipe Hoffa

Data Cloud Advocate at Snowflake ❄️. Originally from Chile, now in San Francisco and around the world. Previously at Google. Let’s talk data.