From Parking Sensors in Shanghai to ML Platforms at Snowflake - Making AI Accessible to Everyone
Guest: Paul Zhao, Senior Manager of ML Platforms, Snowflake; Founder, Black Swan Tech
Hosts: Seth Earley, CEO at Earley Information Science
Chris Featherstone, Sr. Director of AI/Data Product/Program Management at Salesforce
Published on: February 8, 2022
In this episode, Seth Earley and Chris Featherstone speak with Paul Zhao, most recently Senior Manager for ML Platforms at Snowflake and previously on the speech recognition team at AWS, whose career spans IoT entrepreneurship, cloud insurance, and enterprise AI. Paul traces a path that began with his parents giving up careers as a surgeon and embassy translator in China to immigrate to the US, runs through a bootstrapped startup that deployed 7 million IoT sensors across parking garages, airports, and train stations in China before being acquired, and arrives at a deceptively simple thesis: the most powerful AI tools in the world are still not in the hands of the people best positioned to use them. He explains the three distinct ways that gap destroys business value, makes the case for why the build-versus-buy decision comes down to whether you have differentiating data, and offers a calculator analogy that cuts through years of AI hype to explain where each role in an organization should actually be playing.
Key Takeaways:
- Being early to market with genuinely transformative technology is not the same as being wrong - the IoT parking sensor startup that got 50-plus rejections in the US found a receptive market in China because Chinese businesses in 2008 had a much higher appetite for experimentation to gain a competitive edge.
- Early adopters perform an inadvertent service for the world by running the experiment that proves whether an innovation works, footing the bill for that discovery, and creating the reference point that makes adoption easier for everyone who follows - pioneers get the best land but also get arrows in the back.
- The goal of AI is always the same regardless of sophistication: identify value, then capture that value for whoever the business stakeholder is - whether the algorithm is a simple rule or a self-learning model, that question never changes.
- AI supplements human judgment rather than substituting for it - the practical value of ML is accelerating certain tasks and finding patterns that humans cannot process at scale, not replacing the contextual judgment and domain expertise that humans uniquely bring.
- Democratizing AI means eliminating three distinct failure modes for non-technical users: not knowing what value exists in their data, knowing value exists but lacking access to the right tools, and having access to tools but misapplying them because the tools were not designed for domain experts.
- The build-versus-buy decision in AI simplifies to one question - do you have differentiating data? Companies like Google and Amazon build because their data gives them a competitive edge; companies without uniquely differentiated data in their domain are almost always better off buying and then configuring at the layers where their specific knowledge and experience actually lives.
- The future division of AI labor looks like calculator design versus calculator use - ML scientists will focus on building better and better models, while business analysts and domain experts will apply those models to extract value without needing to understand the underlying mechanics, just as accountants use calculators without understanding transistor design.
Insightful Quotes:
"We're still heavily reliant on folks who have relatively specialized expertise to apply AI. And that's natural, but it's also in some senses a tragedy - because it means that some of the most powerful tools at our disposal are really not being put into the hands of people who could use them. That deprivation is severely limiting because there's all this unknown value that could be excavated." - Paul Zhao
"If you're a data moat like the Googles and Amazons of the world, you're going to want to build - because whatever you build will be differentiated because of data. If you're a company that really doesn't have access to very differentiated data in your domain or your competitive space, you might be better off buying and then tweaking." - Paul Zhao
"You can't outsource your competitive advantage. You have the available capability at the commodity level, but the question is what level of granularity you need to own - where does your competitive advantage actually come from? At some level you have to build, even if you're buying all the components below that." - Seth Earley
Tune in to hear Paul Zhao explain why getting fifty-plus rejections from US parking garage operators before pivoting to China felt less like grit and more like ignorance mixed with arrogance, how a revenue share model on incremental improvement only - not on the base business - finally unlocked his first customer, why his father (a Business School professor) told him that an MBA is not where you go to learn things but where you pay for the privilege of knowing great people, and why he believes confusion about what "AI experience" even means in a job description is exactly what is creating anxiety about whether AI will replace human workers.
Contact Paul:
Thanks to our sponsors:
Podcast Transcript: Democratizing AI - From IoT Entrepreneurship to Making Machine Learning Work for Everyone
Transcript introduction
This transcript captures a conversation between Seth Earley, Chris Featherstone, and Paul Zhao about what it actually takes to bring transformative technology to people who are not already experts in it. Drawing on his experience building an IoT startup from scratch in China, working on speech recognition and ML platforms at AWS and Snowflake, and founding a cloud insurance company, Paul offers a practitioner's view of why AI accessibility is still far from solved - and what it would take to genuinely level the playing field.
Transcript
Seth Earley: Good afternoon, good morning, good evening depending upon your timezone - welcome to today's podcast. I am Seth Earley.
Chris Featherstone: And I'm Chris Featherstone. Good to be with you.
Seth Earley: Before we get into it - our guest today is someone who sits comfortably at the intersection of technology and business. Most recently he's been at Snowflake as a Senior Manager for ML Platforms, working on making machine learning accessible to folks who don't always come from an ML background. But his entrepreneurial spirit has guided him on a journey of very interesting waypoints. From Seattle, please welcome Paul Zhao.
Paul Zhao: Hi Seth, hi Chris. Thanks for having me. Happy to be here.
Seth Earley: Chris, you and Paul connected at AWS - do you want to tell us how that happened?
Chris Featherstone: I'm super excited to have Paul on. He's not only a phenomenal mentor to me but a great friend. One of my main goals in choosing companies to work for has been to go find great people to work with. When I met Paul I was super impressed by his approach, and as we got to know each other I was equally blown away by his background and what he's been able to accomplish. Surrounding yourself with great people means some of that excellence just rubs off.
Paul Zhao: So most recently, like Chris introduced, I was at Snowflake as a Senior Manager responsible for the ML strategy - specifically working on ML platforms that give accessibility to folks who are not always coming from an ML background. Preceding that I was at AWS where Chris and I met - I worked on speech recognition under the ML group, and Chris and I had a chance to tackle some pretty tough enterprise customers together. Even farther back, I ran my own business - a company called Black Swan Tech, which sold cloud insurance. If you're a company running on public cloud and that public cloud has downtime or failure that causes business disruption, our company would come in through a contract and make you whole. And before that I went to Business School, and before that I ran my own business for about five years in IoT in China.
Seth Earley: I love hearing the longer version about the IoT initiatives. What year was this, because this was fairly new territory?
Paul Zhao: This was from roughly 2008 through 2013, right at the height of the economic recession. My co-founder and I had the - I'll package it as wisdom, but it was probably stupidity - to fly to China and say: we'd like to retrofit large-scale commercial facilities with sensors and hook them up to software so that operators can make real-time decisions based on real-time data. That's how it all began.
Seth Earley: And one of the areas you chose was parking facilities, correct?
Paul Zhao: That's right. You tried it in the US first but couldn't get much uptake. People didn't really understand it at the time - it wasn't even called IoT yet, it was just sensors. You had to explain the entire paradigm. That's a tough sell when you're early to the market.
Seth Earley: Being early to the market - a little too early. People can't conceptualize what the problem really is and how it can be solved. We face a very similar situation today with machine learning and AI. At the granular level, there's a lot of uncertainty about what it can actually do for specific processes. Go back to what you were facing and how you tried to combat it.
Paul Zhao: There were really two important challenges - one macro, one micro. The macro challenge was that 2008 was a time when we started in the US and didn't have a lot of traction. A lot of businesses were still pen and paper. It doesn't seem like a long time ago, but 14 years ago, that really was the world. And many of the garage operators we approached had an attitude of: why do we need to care about optimizing our pricing in real time? People are going to park if they want to park.
There's a lot of non-data-driven thinking surrounding operations like that. They didn't know what they didn't know - they put everything in the context of what they already had, and couldn't envision the shift.
The micro challenge was about what we were really asking them to do. We were presenting two theses: first, you have to digitize, and second, once you've digitized, you need intelligence - software to process all that data and actually make decisions. That combination was simply not palatable for a lot of customers in 2008. We were basically selling to parking operators in Chicago and New York and getting no after no. Over 50 formal meetings.
Seth Earley: Think about that. When people get discouraged after two or three or five nos. You went to fifty and were still believing in your vision.
Paul Zhao: You're generous. From the inside it's really a combination of ignorance - not knowing what we were up against - some degree of stupidity because we weren't learning fast enough up that curve, and some degree of arrogance: hey, we're introducing something novel, the old guard just can't see it. Those three things together is what people on the outside package as grit and resilience. I just didn't know better.
Seth Earley: Pioneers get the best land, but they also get arrows in the back.
Paul Zhao: Exactly. And any form of innovation has early adopters who essentially do the world a favor by running the experiment - consuming whatever that new innovation is and footing the bill. Sometimes it works, sometimes it doesn't, but either way there is that inadvertent experimentation. We wouldn't be anywhere without those early adopters.
Seth Earley: So how did you finally get traction? You went to China.
Paul Zhao: We had run out of runway - coming down to the wire on our savings, which we'd bootstrapped everything from. My co-founder and I had both formerly worked at Goldman, so that helped a little as young guys out of university. But that's never enough to keep you going for multiple years.
On the business side, we felt China was a highly competitive place where businesses were willing to try anything new to gain an edge on their competition. The appetite and culture of experimentation felt extremely open. And the other thing was: everything in China is big. Even a small percentage of success would eke out larger absolute value. That was the supposed wisdom at the time.
Seth Earley: And it wasn't 50 conversations this time.
Paul Zhao: More like 15 or 16, and by that point we had some good indicators - MOUs signed. Here's where we started: we said we'll do experimental setups, forgive you all the costs, do a proof of concept in the field entirely at our expense. And then we'd do a revenue share on the incremental improvement - not on the base business, because our whole point was that we would drive incremental value. That took all the economics off the table. Basically: you have nothing to lose. That bred a degree of confidence. These guys are that confident in what they're selling. And then customer success breeds more customer success.
Seth Earley: How did the business eventually grow? What did you cover?
Paul Zhao: The business grew to cover commercial parking facilities, airports, sporting arenas, even train stations - basically anything where you could measure physical or virtual environment, we deployed our sensors. We began to saturate most of Guangzhou, which is one of the tier-one cities in China, then spread to other cities like Shenzhen. At some point our largest customer, Grandview, decided it was just more profitable for them to buy us out and integrate our technology into all their new development going forward. We wrapped it up, and I came back home.
Seth Earley: What scale did you get to?
Paul Zhao: We had over 7 million sensors physically deployed across southeast China, and we were at around 100 employees at that point. That gave us significant regional scale and what I call data monopolization - because data piggybacks off itself. Having data about parking garages not just for one customer but across many customers meant we could create a data map of vacancies across an entire region. We could change the flexibility of when you adjust prices, not just what the prices are. Supply and demand timing matters a lot. And from there more incremental value cropped up - like notifying garages of overflow situations. That couldn't have been done before.
Seth Earley: At that point you probably needed machine learning to make sense of all that data.
Paul Zhao: We had started tinkering with it, but to be quite frank, our team hadn't been equipped to really engage with ML just yet. And then Grandview came in and made a nice offer, so we said not a bad exit.
Seth Earley: And after that?
Paul Zhao: I told myself I had no idea how to do a normal job. The title of CEO and founder of a startup is really confusing to a lot of companies that hire. So I went to business school - I saw the brochures, everyone's holding glasses of wine or wearing graduation gowns with perfect smiles, and I said I want to go to that.
Actually my father teaches business school, at St. Louis University today. I asked him: what's your take? He said something really simple - if you really want to go learn stuff, don't get a PhD. You learn in undergrad, you learn in a master's program. But an MBA is probably not where you go to learn deep knowledge about anything. What you go there for is to pay for the privilege of knowing really great people and building that network. So I did that for a couple of years.
Then Black Swan Tech - cloud insurance. And then AWS, where Chris and I met working on speech recognition for enterprise customers. And most recently Snowflake, working on the ML platform accessibility problem.
Chris Featherstone: Let's get to the core theme - how do you make AI easier for the common person who doesn't know ML to actually take their data and use these models to generate outcomes?
Paul Zhao: That's a really good question. I think we're still at the very beginning of this journey. We're still heavily reliant on folks who have relatively specialized expertise to apply AI. That's natural, but it's also in some senses a tragedy - because it means that some of the most powerful tools at our disposal are really not being put into the hands of people who could use them.
There are actually three distinct types of deprivation here. First: you don't know what value is sitting in your data. That's the not-knowing-what-you-don't-know problem. Second: you know value is there, but the tools aren't accessible to you. Third: you're a sharp person, you try to use those tools, but you misapply them - because they weren't designed for folks like you and me, so you derive the wrong value from the analytical process.
That's why it's so important to truly make these tools easy to use - to take the PhDs out of the process. I always go back to when I have a conversation with my parents about a certain subject and they go into PhD mode. That's wonderful - say, in ML - but as soon as you want to connect with someone more practically oriented around business value, you lose them. If you can imagine that same analogy happening at the application layer of AI, that's a tremendous blocker to value accessibility and transparency.
Most of my work has been around: if you're interested in speech recognition, can it just be plug and play? If you're a business analyst at a large insurance company trying to understand patterns in claims behavior, today you commonly rely on an ML specialist to go crunch the data and do all the modeling. But as an analyst you have a lot more domain knowledge about actual customer behavior and what it means. Your job is slowed down because of a dependency on someone else. What if we put those tools in your hands and eliminated that interdependency?
Then your ML scientists can be freed up to do really cutting-edge R&D - which they love, by the way, because they don't want to spend their time on mundane data crunching. And for you as an analyst, you get to immediate value realization because you're treating this as a tool. We're taking the mythos and the mystery out of ML.
Seth Earley: That's the idea of layers of abstraction. At the deepest level, you have computer science, mathematics, deep algorithmic expertise - fine-tuning hundreds of ML outcomes, working with hyperparameters, really refining those models. Then a layer above that, taking those algorithms and abstracting certain parameters so they can be better tuned to a specific application. Then building blocks that can be assembled into applications. And then the fine-tuning of business processes and integration across an ecosystem. The AWS suite of tools lets you come in at any of those levels. Where should organizations be playing?
Paul Zhao: It really does come down to the build versus buy conversation, and the short rationale is this: it comes down to whether you have differentiating data. If you're a data moat - the Googles and Amazons of the world - you're going to want to build, because whatever you build will be differentiated because of the data. If you're a company who really doesn't have access to very differentiated data in your domain or your competitive space, you might be better off buying and then tweaking.
There's a lot of confusion across that value chain, across the tech stack. If you're not actually a data-driven company and you pretend to be without really knowing your identity, you're likely to face a very, very tough roadmap.
Seth Earley: We were working with an organization trying to build capabilities to better serve their customers. They do have differentiated data - it's their first-party data about their knowledge of their customers, that's what they're competing on. The question is what level of the tech stack they try to leverage. The last mile in the cognitive AI space is the organization's own knowledge. You can have lots of off-the-shelf tools you rent from big providers, but your differentiator - your competitive advantage - is really that knowledge layer. You can't outsource your competitive advantage. You have the available commodity capability, but the question is at what level of granularity you need to own it. Where do they build, what skills do they hire internally versus what they rent?
Paul Zhao: Right - you buy the commodities. General compute, network, and storage are commodities now, infrastructural stuff. A little higher up the stack you might do a mix and match when you get to the analytics layer. And then when you get to the capabilities and knowledge and experience that are truly differentiated for your business - that's where you might need to build, or at least configure and adapt, because commodity sellers don't have that specific knowledge. They can't package your secret sauce.
The analogy I keep coming back to is the calculator. In the future, ML scientists are the people who design that calculator - making it better and better, from the TI-83 to the TI-89 and beyond. They're focused on and happy doing that because that's their expertise. But then there are the accountants and the mathematicians who don't want to be bogged down in the mechanics of how the calculator works. They know how to apply it really well - they know when to use a square root versus a quadratic, even if they don't think about the mechanics of how that's solved by the hardware. That's the world we should aspire to, because that's what specialization does: it allows each person to be really good at what they do for the maximum ROI based on their expertise, training, and education.
Seth Earley: And the anxiety about whether AI will replace human workers connects to this.
Paul Zhao: Exactly. Part of the problem is we haven't answered what "AI experience" even means in a job description. Is it application of models or design of algorithms? Until we answer that clearly, it's going to create a lot of nervousness. I think people in application roles - not in design roles - have very little to fear. It's the specialization that has to be understood.
Seth Earley: Paul, it's been great. I love hearing the stories. Hopefully we can have you back and dive into some other areas.
Chris Featherstone: As always, thanks so much. Much love. Part of this is I love having my friends on the podcast. Thanks for doing this with us.
Paul Zhao: Thanks so much, both of you. Really appreciate it. Have a good one.
Seth Earley: And thank you to our sponsors - the Marketing AI Institute, CMSWire, and Earley Information Science. And thank you to our producer Sharon for all the work she does keeping things on track. Look forward to talking to everyone next time.