Group of people working in office.
An arrow pointing leftHome

Weeding out the AI pretenders in the PNW's startup sphere

  • Adam Bluestein

A Q&A with Frank Chang, co-founder and managing director of Flying Fish Partners.

Flying Fish Partners' co-founders and managing partners (in order from left to right) Geoff Harris, Heather Redman and Frank Chang. (Jason Redmond | Microsoft Alumni Network)

Seattle-based Flying Fish is an early-stage venture capital firm focused on AI and machine-learning startups. Founded in 2016 by veterans of Microsoft and Amazon, Flying Fish is the one of the — if not the — only regional investor groups with a laser focus on AI. And it’s one of very few such specialists nationwide.

We recently spoke with co-founder and managing director Frank Chang, soon after Flying Fish announced the launch of a $70 million second fund, which will invest in roughly two dozen promising startups in the Pacific Northwest and beyond.

This interview has been edited and condensed for clarity.

Do you have a background in AI? And how important is that for investors in AI and ML startups?

Our three, co-founding managing partners, are all former operators. Heather Redman was most recently at a data company called Indix. I spent 17 years at Microsoft, and nearly three years at Amazon. Our other cofounder, Geoff Harris, and I worked together at Microsoft on and off for about 12 years. The last thing we did together there was running the speech and natural language team, which turned into Cortana. All of the voice recognition technology for Microsoft came out of our team. We did that for five years together. So we’re fairly steeped in a hard AI problem. When a founder comes to us, we can leverage a deep network of contacts to find people we know who can work with them. Furthermore, everybody claims that they have AI and machine learning in their pitch deck today. Lots of people don’t really have it. We’re able to weed out the pretenders.

What made you decide that AI startups in the Pacific Northwest represented a niche that was worth focusing on?

Geoff was doing angle investing and realized there was a gap when it came specifically to seed investors in the Pacific Northwest. There were lots of angel investors, and people from the Valley and VCs here, like Madrona, will do bigger deals. Not a lot of people were doing seed deals — $1 million to $2 million — so that’s where we saw the opportunity. And with our experience working in AI, we had a firm belief that AI was going to be sort of the computing paradigm for the next several decades. Really, that’s the basis of our thesis.

Are there any unique challenges that AI startups are facing now?

As more and more platform-level technologies become available, you don’t need to be an expert in all of the lower-level algorithms [to build AI applications]. However, you still need to understand data, and understand what kinds of data you need. You still need to tune your models. You need data science expertise. Most AI problems also suffer from a cold-start problem. On day zero, you usually have no real-world data to feed your models. You need time to get data, to build a feedback loop, to tune the models, to make sure things are really accurate. People may not even know how to interact with your system — this is a very new field, relatively speaking. You may have the best model and technology on the planet, but if someone’s not comfortable talking to an AI chatbot, if you don’t build the user interface in the right way, it may fail. There are a lot of nuances to building an AI system. You need the patience to understand that it takes time to get it right. This isn’t traditional software where you just find a bug and you fix it and then tomorrow it works. I think a lot of investors will get frustrated by that.

So, what is your typical timeframe for working with a company?

It’s a 10-year fund, as most are, with two optional one-year re-ups. So each fund could theoretically go for 12 years. We’re just three or four years in with many companies from our first fund. Typically, we would expect exits to start showing up around years five to eight, and most of those are going to be a large tech company buying them out. We’re not banking on huge IPOs to get the returns that you need.

Do you tend to be very hands-on with portfolio companies?

It totally varies by the founder. We’ve had companies where it’s a founder’s third or fourth startup, they’ve got everything really buttoned up, and they don’t need a ton of help from us. They may need the odd connection here or there, and that’s great. There are others who are a couple years out of college, they’ve never run a board meeting. They want the advice, and they want us to be more hands on. I’ll sync with them for an hour or two every week, walk through financials and board decks, product planning, milestone setting, all these sorts of basic things. That’s one of the benefits of our operating experience.

You’ve been involved in natural language processing for most of your career. What’s happening in that space now?

Natural language processing has come a very long way since I was deep into it, and I have a slight worry that there’s a lot of hype built up around it. There’s that Google engineer that came out and said that their system was sentient. Eh, I don’t think so. But the systems are getting better and better. Something like GPT-3 [the open-source language modeling platform created by OpenAI] is trained on just an enormous amount of data. It’s almost like [the idea] “give a million monkeys a million typewriters, and eventually they’ll make Shakespeare” — but not quite that random. If you throw enough data at something, it’s going to be able to answer everything in a seemingly intelligent way.

Are there other emerging opportunities for NLP startups?

There’s so much legacy data out there. I think one of the biggest opportunities for NLP is in trying to gain insights from all this data, in finance, insurance, health, law — just terabytes and exabytes of data that are out there just sitting in storage not being leveraged. So, I think there’s still plenty of green field.

How is AI impacting manufacturing?

There’s a lot to unpack here. There’s the baseline problem of shortages in labor. People don’t want to work in these repetitive, high-stress, potentially dangerous environments, and so anywhere you can put some sort of AI smarts into a robot that can handle the task, the better. We’ve got one company in our portfolio, Apera, based in Vancouver, that uses computer vision to be very accurate with picking parts out of a bin. That seems like a mundane task, but it turns out that lighting is a big deal and parts that are transparent or shiny really screw up a computer vision camera. To be able to have a system that is highly accurate in different lighting conditions is no small feat. There are lots of improvements like that happening across the industry. Line workers that are today just sitting there, literally picking parts out of a bin and putting them in a container, those jobs will go away, and they can up-level those people. We’ve also seen companies using computer vision to understand where factory workers on a production line are more or less efficient — and to identify patterns of behavior that might result in a repetitive stress injury, so you can catch that and nip it in the bud. There’s a great opportunity for AI to come in, observe and make some corrections.

Which current portfolio companies exemplify some of the larger trends that you’re betting on?

Phaidra is one. There’s this technique called reinforcement learning — AlphaGo from DeepMind is the most famous example. This game system played thousands of games, learning from its mistakes, and eventually went on to beat Go world champions. The company is founded by two former members of DeepMind, and they’re using reinforcement learning to make industrial processes more efficient. Things like glass manufacturing, aluminum, smelting or pharmaceuticals are super high-energy usage processes that are also highly sensorized. You’ve already got the data coming in, and you can put an AI on that to learn and understand what the parameters are and what you can and can’t tweak to make the system more efficient. They’re focused on data centers now — an internal project they did at Google’s data centers saved Google around 40% in energy costs.

Who else?

Picnic are the guys that are making the robots that make pizzas. COVID accelerated a lot of technology adoption and acceptance, and sort of highlighted two things: People don’t really want other humans touching their food, as much as possible; and nobody wants to work in restaurants. So here you have sort of a two birds, one stone solution, a machine that is super sanitary, way more efficient, never calls in sick and always makes the pizza super consistent. Restaurants love the fact that there’s no food waste. The machine makes it with the same amount of cheese every single time. Originally, we said we weren’t going to invest in hardware — it’s too difficult, too expensive — but it turns out that now, when you’re dealing with AI, you’re sort of interfacing with the real world.

How does sharing a geographic location with two of the biggest acquirers of AI technology — Amazon and Microsoft — impact the startup ecosystem there?

It’s a double benefit. Having them here draws the talent, and people don’t stick around at either company now that long. There’s been a lot more sort of Valley entrepreneurial focused people moving up into Seattle, and every Valley company also has an engineering outpost here. So you don’t have to work for Amazon. You can go work for whichever tech company is the flavor of the month right here. That drives a great talent pool for us, also lots of potential acquisitions. It’s not just Microsoft and Amazon, but also the Valley companies that will acquire companies here, leave the team intact and use this as a sort of base of operations. Apple’s acquisition of Turi, in 2016, is a great example of that.

Are any other firms focused on AI and machine learning the way you are?

The only one that really pops to mind, that’s very focused on AI like us, is Glasswing Ventures out in Boston. Other firms invest in AI and machine learning, of course, but it’s one of five or six focus areas that they have, versus an entire fund focused on that. The most common piece of feedback we get from founders when they talk to us is, “nobody’s ever asked us those questions before.” For them, I think, it’s refreshing that we put so much focus on the technology and the product. It’s not just the financials and the revenue and the pipeline — 99% of VCs just ask about that.