In our latest episode of Angel Insights, we spoke to founder, machine learning researcher and angel investor Zehan Wang.
After completing his PhD, Zehan Wang began his career as Co-Founder and CTO at Magic Pony Technology, which pioneered the use of neural networks for video enhancement, prediction and compression. After just two years of operation, it was acquired by Twitter for an estimated $150m in 2016. Zehan continued his career at Twitter as Head of Cortex Applied Research, as well as developing his angel investing portfolio.
You can listen to the full conversation here:
Or for a few key insights, read on:
1: Founders, take advantage of external, experienced perspectives to ensure you’re focused on the right things.
Zehan: “Almost every company I have worked with is somewhat unique in what they're doing. One thing that I have to remind people, particularly first time founders, is that there is no magic formula that you can follow to achieve success. You can't just redo what we did at Magic Pony and somehow expect that to be successful. Every case is unique and circumstantial. So what I try to do is help people talk through their particular circumstances and raise questions to provide perspective they wouldn’t otherwise get. I think a lot of it is just bringing a different, more experienced perspective to help founders understand what are the right questions to ask.
It's the unknown unknowns that get people: people always want to do their best, but how do they know if this is the best they can do? Or how do they know that they’re targeting the right things? So having an outside perspective is always important for that.
2. Investors, don’t put all your faith in past success.
Zehan: “One of the things I learned from that experience is that you should not judge people based on whether you've come across the name before or not. If I look back on myself, at that time, we were very much unknown. Many investors would pass on us because they thought, “Who are these guys? They're young, they've never done this before.” One thing I've learned is not to judge purely on a history of successes, because that doesn't always mean that there will be successes in the future, and in the same way, just because someone doesn't have a history doesn't mean they're not going to be successful in the future either.
I see this bias play out in the startup investment world, entrepreneurs that have made a name for themselves, or have some level of success tend to raise much bigger rounds, and so on, but it doesn’t necessarily mean they improve the odds, not compared to the differences in the funding between the unknown and the knowns, or at least I don't think that's always justified.
There's an interesting trend recently as well. If you're a famous machine learning researcher and you decide to start your own machine learning startup, you can almost instantly raise a fund from one of these VCs that follow these trends just because you're in the know in that field. But I would say if they applied that approach to us, that wouldn't have worked. Because we were not known in this field. My PhD background wasn't really pushing the frontiers of machine learning. It was very focused on something else. So I think there's a lot of people who have the right talent and capabilities who haven't shown the potential yes who are often overlooked by others.
One irony with somebody who has become a bit more famous is that they just have too many options for where they could invest their time. So they don't necessarily always have that focus. Then, because they have the options, it's hard to judge which of these options are the ones they're really passionate about, and really want to be pursuing and investing their time in. Whereas I think for some of the people who are new to this space, it's almost like they have something they want to prove about themselves. So they've chosen this field for a reason.
3. Since 2016, the machine learning and AI industry has come a long way.
Zehan: “There's several trends in how the industry has evolved. I think one is the commoditisation of machine learning as a whole. When we were starting, it was really only a very select few that had the knowledge to do anything substantial with machine learning. Within the industry, you tended to have more data scientists who weren't necessarily researchers.
Of course, you could argue that machine learning is really just applied statistics at the end of the day. But I think part of the question is how do you scale up these algorithms, and that's what deep learning has really enabled. You can scale up how you apply machine learning methods to a vast quantity of data. So in some ways, I think we've just really entered the Big Data age. We were there before, but the fruits of those big data outcomes hadn't really been seen to the same state as we've seen more recently. So I think there's been quite big leaps in terms of how we do it, particularly in natural language processing, which has given some very impressive results. People have seen things like Dali where they do text to image generation, as well, it's very impressive. But also for some of these areas, it's quite subjective in terms of how you assess the quality of the outcomes. There isn't an objective metric to say this is a good generation of an image. This is a bad generation, particularly when it's something that's very arty. So in some of these areas it’s interesting how we quantify progress in the future, as well.
So commoditisation of machine learning through open sourcing models, variable sourcing code, and then advances within the hardware that we build on as well. So we have more powerful GPUs that basically enable you to run much bigger models and train much faster, and so on. So really building on our ability to handle vast amounts of data and process that and derive learnings and statistics from it.
Bonus: Where did the Magic Pony name from?
Zehan: Matt Clifford, the CEO of Entrepreneur First, used to give this talk to other startups about how you should not build a magic pony, which is this thing that people want, but doesn't exist or can't be real. And then at the same time, some of the ideas we were initially pitching to potential customers, people just didn't believe that we could do them. Obviously, we hadn't proved it out. But we had a hunch of how they could work just because we've done the research and understood how that technology could work.
Either way, because people didn't believe what we're doing, success or failure we would be a magic pony. So yeah, that's kind of how it stuck.
Even now, most AI and machine learning companies have very similar types of names. So magic pony actually tended to stand out amongst all these other machine learning startups at the time: something mind or deep something, and so on. So we're different, not the same as the others.
Register to learn
more about our data,
fund and venture capital