Why AI Companies Can’t Be Lean Startups

September 11, 2017 Sam DeBrule

A Conversation with Matt Turck of FirstMark Capital

Today it’s impossible to read business news without seeing some mention of artificial intelligence.

Recent breakthroughs in machine learning and AI have unleashed a host of new experiences that can be delivered to customers, and companies both large and small look to capitalize.

Large companies want to seize the opportunity to strengthen their existing positions, and small companies hope to fuel their rise to market leadership with these new technologies.

As an investor at FirstMark, Matt Turck has seen the machine learning-first startups that are poised to change the way we work from up close.

Here’s a discussion of the way machine learning-first startups are built, which ones are pushing the ecosystem forward, and why they look so different than the SaaS startups that came before them:


Sam: Matt, could you introduce yourself and talk a little about the work you do for the people who aren’t familiar?

Matt: Sure! I’m a partner at FirstMark in New York. In just a few years, we’ve become the largest early stage NYC-based venture firm, with about $1.6 billion under management, including $500M of new money raised last year.

From an investment perspective, while I’m interested in a bunch of things, for the most part I’m particularly focused on two major areas.

The first is “the world of data”, broadly defined. This includes Big Data, machine learning and artificial intelligence companies, as well as startups where data is the “secret sauce” or the core competitive moat I very actively invest in the space through companies like ActionIQ, Dataiku, x.ai, Sense360 and HyperScience. I also blog extensively on those topics and run Data Driven NYC, a big community of 14,000 Big Data and AI enthusiasts.

Frontier Tech is my second major area of focus. Frontier Tech includes emerging computing platforms, augmented/virtual reality, Internet of Things, and all the other buzzwords you can think of (laughs). Same deal, I actively invest in the space, I blog about it, and I run another big community, Hardwired NYC, that brings together 5,000+ people to explore frontier technologies.

The hype around AI is very real

Sam: Very cool — when did machine learning-focused startups first came on your radar as investment opportunities?

Matt: Well, pretty much all my career in tech, I have been immersed in data and analytics. A while back now, I was the co-founder of an enterprise search software startup, and we were all about applying Bayesians methods, a form of machine learning, to search and retrieval problems.

So machine learning has been very much on my radar for many years, but it’s certainly become even more interesting recently. The whole Big Data thing came along and provided the infrastructure to capture and process massive amounts of data at a reasonable cost and speed. That in turn accelerated machine learning dramatically, in particular areas that require large amounts of data to work, such as neural networks.

It certainly is a very unique time to be an entrepreneur or an investor in the field, as it feels like decades of work have “suddenly” come together, unleashing so many possibilities. The hype around AI is very real, in my opinion.

Marketing claims outpacing technical reality

Sam: Okay, let’s discuss a controversial question. Of the large companies working on AI today, which is worst positioned to deliver on its promises?

Matt: Based on what I hear from the market, probably IBM.

Largely because they have a very impressive marketing machine and big ambitions, and got themselves in a situation where they promised a lot and have gone after too many verticals at once.

And even though I hear you can do very interesting things with IBM Watson, as long as you’re willing to spend a few months and a significant amount of money training it, it’s clearly a situation where their marketing claims have gotten a bit too far ahead of what they can deliver.

Part of it is the curse of any gigantic company like IBM: to move the needle even a little bit at that scale, any new business needs to become very large very quickly, which creates a lot of pressure on everyone involved.

They find themselves in a situation where they’re competing on every deal in every vertical — and I’m hearing from the market that they’re losing a fair number of deals, often to much smaller startups that are more focused and nimble.

But hey, they’re IBM. Let’s not count them out just yet (laughs).

Access to machine learning talent is the limiting factor

Sam: Is Salesforce going down the same path?

Matt: I don’t think so. Sure, I’m told people internally at Salesforce freaked out a little bit when Marc Benioff made all those big claims about Einstein [Salesforce’s new AI] last year, but part of it comes with the territory of big personalities.

Connectivity to the startup ecosystem and a willingness to buy companies make a big difference, in part because they give you much stronger access to machine learning talent, which is really the gating factor at this stage.

Salesforce has much better connectivity to the startup ecosystem, which matters a lot.

They’re very active on the investment front through Salesforce Ventures, and Marc Benioff has personally invested in all these different startups that leverage machine learning. And of course, they’re also very acquisitive.

“…the sharpest minds in the industry (not just in startups) are working on making neural networks function with smaller amounts of data these days”

Challenge of securing access to proprietary datasets

Sam: It’s common knowledge that a huge amount of data is needed to execute on AI, which obviously puts startups at a disadvantage against large companies. Are you seeing data engineers finding ways to power their models with smaller datasets?

Matt: Startups can certainly be at a real disadvantage on that front, but as always, they end up being quite resourceful.

From a technical standpoint, I think a lot of the sharpest minds in the industry (not just in startups) are working on making neural networks function with smaller amounts of data these days. For the foreseeable future, that’s sort of the holy grail. I’m familiar with at least a couple of startups making some real inroads and working on interesting things around transfer learning. Having said that, it seems to be a particularly tough problem, so it may take a little while.

It once would have seemed like a crazy barrier to get large amounts of data, but startups are finding all sorts of creative ways to do it.

In the meantime, startups are figuring out all sorts of ways to get access to the large datasets they need. Take the example of some AI companies in the medical imaging space. I’ve seen several of them really hustle and somehow secure access to proprietary databases of radiology images after building a special relationship with a given hospital. I have seen others do the same in a bunch of different areas like collision insurance, industrial machinery, or agriculture.

There’s a German startup, TwentyBn, that’s created a crowdsourced library of hundreds of thousands of videos where people literally “act” a certain gesture or action in front of the camera, so that the computer vision system can learn what that action is. So they essentially built their own dataset.

It once would have seemed like a crazy barrier to get large amounts of data, but startups are finding all sorts of creative ways to do it.

By the way, getting the data is only part of the challenge, you also need to label it, for deep learning to work. And here as well, startups have been really resourceful in figuring it out.

I’ve seen some startups building little armies of people around the world to label their data in a “Mechanical Turk” kind of way. I’ve seen others recruit deep industry experts to label very specific types of data, such as an elite group of surgeons to label the most complex dataset of medical images, for example.

Kickstarting data network effects

Sam: Through products like Netflix, Spotify, and Facebook, many people have experienced the benefits of data network effects. What startups are building the next generation of powerful data network effects?

Matt: I wrote a blog post about data network effects a little while back, it’s a topic I find fascinating.

In theory, any machine learning company that can pool enough data from multiple users, run algorithms on pooled data set, and send back that learning to each individual customer, can benefit from data network effects.

In a B2B context, it’s often a bit harder to get data network effects going because corporations are particularly protective of their data…

To use an example from the FirstMark portfolio, x.ai comes to mind. The more meetings their AI assistant schedules, the more data they get, and the smarter their algorithm becomes. The smarter the algorithm, the better the experience. The better the experience, the more likely people are to want to schedule meetings using x.ai, the more data the company captures, and so on. This flywheel is the data network effect.

“In theory, any machine learning company… can benefit from data network effects.”

The beauty is that this applies to a number of circumstances, from companies like x.ai that help people to schedule meetings faster to computational genomics companies like Phosphorous that work with hospitals to help them run genetic testing labs.

In a B2B context, it’s often a bit harder to get data network effects going because corporations are particularly (and rightfully) protective of their data and are pretty uncomfortable with the idea that you could be commingling their data with that of other companies in their industry.

But creative solutions are appearing to address this issue. Google Research had this really interesting publication on Federated Learning a few months ago, where the idea was to enable collaborative machine learning without actually pooling the data. That would address all those concerns around data privacy and open the door to all sorts of data network effects.

Regardless, it’s worth bearing in mind that it can take years for data network effects to kick in because startups need to build a customer base to collect enough data from which their models can learn. But it feels like a very interesting competitive moat once you get there.

Complexity of machine learning slows sales cycles

Sam: Do these startups lend themselves to one type of go to market strategy over another?

Matt: I think ultimately, most AI startups will end up looking very much like any other startup. For example, enterprise AI companies will mostly look like any other software or SaaS company, with a variety of different go to market strategies available to them, depending on industry, customer size, price point, etc.

For any AI product to work well enough, startups need to capture significant amounts of usage data, and then they need to use that data to train the algorithms and customize the product.

But I don’t think we’re quite there yet. For now, the complexity around building a machine learning product itself requires a lot of R&D, and training the algorithms requires a lot of time, effort, technical resources, as well as a lot of data, as we just discussed.

To take the example of x.ai once again, it’s taken several years, dozens of data scientists and machine learning engineers and millions of venture capital money to build an AI backend offering high levels of automation and reliability.

As a result, at least for now, it’s much harder for a machine learning company to be a “lean” startup.

For all the hope (and hype) around TensorFlow and other machine intelligence open source libraries, I think it’s trickier for these machine learning companies to build a truly AI-driven minimum viable product, and iterate from there. For any AI product to work well enough, startups need to capture significant amounts of usage data, and then they need to use that data to train the algorithms and customize the product. None of this is quick and easy, and we’re still very much in the “deep tech” world.

This has important implications on go to market strategy. In the enterprise world, for example, I haven’t seen many machine learning startups follow a “bottoms-up” sales strategy, at least not successfully. I’ve seen the opposite much more often: AI startups going after larger customers with larger budgets, sell with a top down approach, and essentially follow a partnership strategy where they build a lot of the early product in close collaboration with a handful of customers.

Often those companies, initially, solve their customer’s problems with a lot of services, and not a lot of software. The hope is that they can build the software on the job, find repeatable use cases, and over time turn their service offering into products. That’s a long process with long sales cycles, it’s certainly tricky, but I’ve seen it work quite well.

But that’s probably just a temporary phase of the market. As machine learning becomes democratized, and you get more open source datasets and algorithms, as well as more trained engineers floating around, you’ll see ML-first companies looking increasingly like every other company, with the opportunity to be more nimble.

Then we’ll all think it was quite quaint that we ever called a company a “machine learning startup”. As any successful technology, machine learning will go from novelty to ubiquity to eventually disappearing in the background.

AI must make product performance 10x better than alternative

Sam: Are investors focused on artificial intelligence startups as quick acquisition targets, or do they actually believe they can succeed as massive standalone businesses?

Matt: Given the economics of venture capital, you have to be very much a believer in the latter.

Obviously we’ve seen a feeding frenzy of large companies snapping up all sorts of small AI companies. That’s what happens when everyone agrees more or less at the same time that that AI is the next big thing, and there’s not that much machine learning talent around.

So you’ve seen many companies that were closer to research labs than actual startups get acquired quickly, sometimes for pretty meaningful amounts. Those were great outcomes for founders, occasionally truly life changing money. But from an investor’s perspective, those outcomes were singles rather than homeruns — not how you achieve great venture returns. I think we’re slowly reaching the end of that phase, though.

This is exactly why investors like me are so interested in vertical AI startups. With a vertical positioning, AI startups can focus very heavily, position away from all the giant companies, and may have time to build a significant business before those come sniffing.

With the right positioning, I do think that AI companies have a window of time ahead of them, during which AI can be a true differentiator and accelerant against any company that is not a heavy user of machine learning. Of course, you want to pick a use case where AI truly makes a huge difference to product performance, and is not just an add on. For the right use case, AI startups can offer a product that is 10x better than existing alternative.

There are lot of use cases where this is not true, but there is a whole range of companies where machine learning does have the potential to make product that’s 10x better. It can create the opportunity to build a company that is a true market leader.

Whether you’re a founder or an investor, it’s all about building companies that leverage the next big market inflection point. A few years ago it was SaaS. Now machine learning is the next evolution. Eventually, that window will close, but for now a lot of those AI-first companies have a real shot at leading their categories, or creating new ones.

Sam: Matt, this was fantastic. Thank you so much for your time.

Matt: My pleasure, Sam!

Join 30,000+ people who read the weekly 🤖Machine Learnings🤖newsletter to understand how AI will impact the way they work and live.

Why AI Companies Can’t Be Lean Startups was originally published in Machine Learnings on Medium, where people are continuing the conversation by highlighting and responding to this story.

Previous Article
Facebook enabling anti-semitic ads
Facebook enabling anti-semitic ads

sourceWhat we’re reading.1/ Facebook’s algorithms create groups like ‘Jew Haters’ for advertisers to target...

Next Article
Companies benefitting from advances in machine learning
Companies benefitting from advances in machine learning

SourceAwesome, not awesome.#Awesome“When comparing the outputs of the machine learning models against the t...