We are excited to announce that Mattermark has re-launched as an independent company! Learn More →
Back to All Articles  |  Editorial

Blood, Sweat and Years: Raising Money As A Deep Learning Startup

tl;dr: Deep learning startups are in vogue, but that doesn’t mean fundraising is easy. Here are the pitfalls startups should avoid to get funded and build a business.

screen-shot-2016-10-25-at-10-11-22-am

This how-to is for founders and technologists who know something about AI and are currently seeking out their very first investment.

I’m going to generalize a little bit about VCs and the deep learning landscape to keep it readable.

Beware the Superior Algorithm

Advances in deep learning have made AI hot again, and that heat has launched a lot of startups, investments, and headlines in the last couple years.

(For most of that time, I’ve been working on a deep learning startup called Skymind, and raised more than $3 million dollars from angels, seed funds, and larger investors on the idea that deep learning will transform most industries. Skymind built Deeplearning4j, an open-source framework for Java and Scala.)

But a quick warning to machine-learning grad students: You’re not going to raise on an algorithm. The marginal improvement you made in the basement lab, and which you’re planning to hide from the world, is simply not enough.

Vicarious is probably the only AI startup that’s really raised a lot of money on their secret sauce. But the real advances in deep learning are happening now in the public domain, accelerated by a slew of papers shared among major universities and corporate research labs.

It’s going to be hard for you to compete with a community of thousands of grad students and researchers backed by massive corporate GPU clusters, testing their ideas and iterating faster than you can imagine.

Solutions Hunting Problems

The Algorithm Fallacy is a special case of a more widespread mistake made in Silicon Valley. Technologists admire machines that run faster, do more, and break some record to become superlative. A billion row join! In under 3 nanoseconds! They build solutions in search of a problem.

But a lot of the time, those superlative solutions leave potential customers cold. So forget about the bright new algorithm and remember you just need to be good at two things:

  1. Machine learning.
  2. Applying machine learning to another field in a way that is helpful.

In other words, you have to make the algorithm useful in a new way, and only users can tell you if it is.

The world hasn’t caught up to last year’s AI yet, let alone the research we don’t even know about yet. The harder nut to crack is not better AI, but how you will sell the AI that works now. How will you change the world’s behavior so that it drops whatever it’s doing and adopts what you’ve made? This is not a technological problem; it’s a sales problem. And when you talk about sales, you talk about human relationships: listening, learning, and persuading.

Good investors know this, and they are going to ask you: Why should the world care? How will you sell this thing you’ve made? What’s your go-to-market? And many more pesky questions like that.

Pitching Ain’t Easy

The first 10,000 pitches are the hardest. Convincing other people to shower you with their money gets easier after that. That is, you’re going to pitch a lot of investors, and almost all of them are going to say no. Please don’t feel bitter. Don’t conclude that life is unfair. Saying no is their job. VCs, like hiring managers, mostly reject. They need to learn about a lot of companies to find a couple they like.

They also have much more time than you. VC time is measured in years and tied to the lifecycle of their funds; Early-Stage Startup Time is months or weeks. That means they won’t feel your urgency, and they may say silly things like “let’s circle back early next year” when you’re staring down the barrel of next month’s rent.

But you only need one good investor to say yes. After that, others will follow because they trust the signal from the first. So your job is to get through the no to arrive at the yes. Don’t tell yourself that you’re wasting precious time. When actors rehearse before their debut, they’re not wasting time. They’re getting better. Pitches are just rehearsals. Their purpose is to make you better for when your big day comes.

screen-shot-2016-06-23-at-7-59-55-am

Your final pitch will be to the pesky public markets.

One of the reasons we managed to raise a large seed at the end of Y Combinator is because we had pitching down. Not every startup is as lucky. Our secret weapon was a desperate year knocking on doors.

Investors come in all stripes and with all sorts of experience. Some investors will understand AI and the industry you’re tackling, and others won’t. The two things that all investors understand, regardless of expertise, are traction and social proof. Traction is how many users or customers you have, and social proof is how many important players are backing you.

So how do you get traction and social proof? To get users, you need a product, to get a product, you need funding, and to get funding, you need users. I’m sure you can appreciate the catch-22 implied in that reasoning. The typical workaround is a bootstrapping marathon of many nights and weekends alongside co-founders who are smart enough to build the tech and dumb enough to accept the risk for very little pay.

Their retirement plan is their equity.

Due Diligence, DeepMind And Big Data

Just because an investor isn’t an expert on AI doesn’t mean you can get hand-wavy about it. Non-experts will have you vetted by people who are smarter than you, me, or them, and the person vetting you will feel a righteous anger if you can’t back up your claims, and your potential investor will hear about it.

So be honest and know your shit. If you have a vision to add AI to your product down the road, be very clear about who you will need to hire, what they will do, and when that moment of adding AI will actually happen.

If you are ready to implement AI now, be ready to name your algorithm and compare it to others to explain your choice. There are usually tradeoffs. For instance, you might accept lower accuracy for more interpretability.

If you are still considering what kind of business you can build on deep learning, consider the following:

  1. Platforms. They’re powerful but rare. If you can build the platform, you control the API, and you have the potential to become the foundation of an enormous ecosystem. That’s what we did, and it worked. But the moment for new platforms in deep learning is passing quickly because network effects are in force. The leading platforms are established.
  2. Products. Now is the time to build products that are infused invisibly with AI. To your customers, the invisible AI won’t matter, because all they want is the experience your technology you can give them. Either you classify photos correctly or you don’t. But investors care a lot about these AI-infused products because they have an edge over the dumb objects we have wrestled with for millennia. This is the edge where new winners are established and old victors are felled.  
  3. Team. DeepMind raised and finally was acquired on the strength of its team. Starting with the founders, DeepMind gathered some of the best minds in deep learning, cleverly working their connections in academia to find the top graduate students until they almost cornered a very small market. Smart people gravitate to smart people. For Skymind, open-source gave us a way to send out a beacon and gather talent together in one team. A lot of the world’s smartest engineers want to work on open-source AI, and they consider it vastly preferable to polishing another mobile app.
  4. Data. Investors will want to know about your data. How will you get the data? What kind will it be? Is it proprietary? Is it sufficient? Whether your data gives you an edge or not, you should have a plan for how to gather, move, and store it, just like every business building an AI solution. If you can’t answer these questions, you’ll look unprepared.

Some people raise on pedigree. They get funded half way through their Stanford Ph.D., or immediately after leaving Google’s machine learning team, and more power to them. If you don’t have pedigree, you need a product with traction. If you don’t have a product, you need a personality that makes your idea compelling. If you can’t even raise on personality, all you’ve got left is persistence. Persistence is what we raised on, and it has come in useful in other ways since the round.

The first 10,000 pitches are the hardest.

Chris Nicholson is a co-founder at SkyMind, an enterprise-facing deep learning company.

Disclosure: When Y Combinator comes up, it’s worth noting that Referly, which later became Mattermark, was a Y Combinator company. As always, the editorial team at Mattermark is independent. 

Join thousands of business professionals reading the Mattermark Daily newsletter. A daily digest of timely, must-read posts by investors and operators.

© Mattermark 2024. Sources: Mattermark Research, Crunchbase, AngelList.
Shares