Angel Insights Blog

By: Ronald F.E. Weissman, Ph.D., Band of Angels and Member, ACA Board of Directors

(Editor’s Note:  Ron Weissman will present a session on artificial intelligence at the 2019 ACA Summit in Chicago, April 24-26.  Get a sneak peek of the “intelligence on artificial intelligence” ahead of the event.)

Artificial Intelligence is on the verge of transforming almost everything, from your medical diagnosis to the way you select your next home, to detecting an audience’s emotional reaction to an ad or TV pilot, to predicting, in silico, if a drug is likely to kill you.  Given the success and spread of AI, how should we approach investing in this technically complex and rapidly evolving field?

AI is complex, but it is not new. It dates back to the 1940s and the two most popular AI paradigms used today both emerged in the 1950s: 1) rule-based AI systems using knowledge engineering to codify what experts “know” into rules, and 2) modeling human intelligence by simulating the neural structure of the brain. 

Whether based on rules or neural nets, AI has had boom and bust cycles throughout its history. The “bust” cycles are periodic enough that they’ve earned a name, “AI Winter.”  AI Spring turns into Winter when investors fear that AI hype has far outstripped the value that AI has, so far, actually produced.

So, where are we in the hope vs. hype cycle in 2019?

Artificial Intelligence‘s most recent winter thaw occurred in 2012, when a team, guided by the work of Geoffrey Hinton, won a prestigious AI competition by classifying images with such a dramatically improved accuracy that the industry was forced to take notice.  In March 2019, Hinton and two colleagues (Yann LeCun, Yoshua Bengio) were awarded Computer Science’s top honor, the Turing Prize, for their collaborative development of those breakthroughs in neural net-based Deep Learning, first published in 2006.  It is those advances that enabled the ImageNet victory six years later.

For now, at least, the neural computing model has won.  After 2012, Deep Learning, a form of Machine Learning, supercharged by Hinton and his collaborators, became the hot new thing.  AI fever hit hard, and the number of machine learning patents and published papers skyrocketed - as did the number of newly funded AI startups.  If AI’s development were a Broadway play, one might call the current era “Springtime for Hinton.”

In the half decade prior to 2012, AI was all about developing core technologies like Hinton’s back propagation, a method of feeding the errors produced by neural nets back into the net to tune and improve it.  But recently, attention has shifted from developing AI core technologies and platforms to harvesting their fruits: using AI in practical business applications.  We’re now deploying AI in autonomous vehicles, computer vision-enabled robotics, medical diagnostics, and customer targeting in sales and marketing, as well as lesser known but equally dramatic AI improvements in law, real estate, insurance, logistics, manufacturing and many other sectors.

As Chinese venture giant Sinovation’s Chairman and former head of Google China, Kai-Fu Lee has recently written, a deployment race is on between the US and China. And it is in China where 70% of global AI dollars are currently being invested.  While the US leads in the development of advanced theory and core technology, Lee believes that China, with its enormous, digitally connected population is poised to lead the practical application of AI and the creation of the massive data sets needed to power AI.

Over the past several years, AI has become such a dominant wave that MIT estimates that two thirds of recently formed startups should really be considered AI-related companies.  If, as Marc Andreesen has suggested, “software is eating the world,” then AI is clearly “eating software.”  As more and more companies seek to gain the valuation halo effect of AI, it will become even more important to distinguish real, transformative AI from cheap wine in attractive bottles.  If we don’t, the seeming ubiquity of AI will make the term less and less meaningful, particularly if it becomes a mere synonym for “technology.”  Beware the coming crapware invasion where everything is “AI”!

As investors, what do recent AI trends mean for us?

Today there are more than 6,500 AI companies, 44% of which have received venture funding, totaling $46B worldwide.  Since 2013, US investors have invested more than $26B. On an annual basis, US AI startup funding has grown 8x in the past five years. On a quarterly basis, it has grown 10x, from $160M to $1.6B per quarter. And according to CB Insights, 32 of the 160+ Unicorns across the globe are AI companies. 

While private capital flowing into AI has received much attention, only a minority of AI “investors” are angels or VCs.  Three of every four dollars invested in AI are invested internally by large corporations.  And beyond their internal investments, an AI arms race is on between technology giants like Google, Apple, Facebook, Amazon, Microsoft, Baidu, Intel and Twitter, fueling a surge in AI M&A. 6.5% of all active AI companies have been already been acquired, more than half in the past two years and at a median value of $197M, and some, for billions of dollars.

AI’s very success has created problems to be solved by a new generation of startups.  Now that AI is being taken very seriously, to avoid liability, AI’s neural networks need to explain themselves.  The European Union’s General Data Protection Regulation (GDPR) requires that AI-based decisions affecting humans be explainable.  A new technology movement, XAI (eXplainable AI) is now creating opportunities for companies to solve AI’s “black box” problem, because we really do need to understand how our models reach their conclusions and recommendations.  Patients need to understand why they were diagnosed with cancer, prisoners need to understand AI-authored parole denials, and accident victims need understand why the car headed for people in the crosswalk rather than the oak tree on the corner.

Data Scientists build AI models and Data Wranglers develop and curate the massive data needed to power those models.  But Data Scientists, Wranglers and data stores containing tens of millions of objects are all in short supply.  New opportunities exist for companies remedying these market gaps.  Some are building tools to create synthetic data or make real data curation easier.  Others have created data marts, Data as a Service, or complete outsourced solutions, Data Science as a Service, combining scientists and data for customers lacking both.

With nearly 7,000 active companies and tens of billions of dollars already invested, the current Machine Learning wave is no longer an uncharted wilderness, at the beginning of its adoption curve. It’s not 2012 anymore.  Today, angels need to adopt a mindset different from the “wide open frontier” of 2012 when investing in AI startups.

AI in 2019 differs from AI in 2012 in three ways: changing investor dynamics, the growth of the AI consumer and the maturing of the AI exit market.  First, AI is now a crowded market, with thousands of funded companies and dozens or more each in popular vertical niches.  Many venture investors who might provide follow-on capital to our seed investments have already made many of their AI bets.

Second, we’ve evolved from AI as a set of investable foundation technologies, to mainstream AI used by real businesses. Early adopters buy tools.  Mainstream buyers solve problems.  Real businesses care about AI only if it produces better results.  We’re about to become far tougher in how we evaluate the AI models that companies actually buy, since the real world is less forgiving than the AI lab.

Finally, hundreds of AI companies have already been acquired by their most logical buyers.  Clearly, the M&A race is far from over, particularly as we’re only at the beginning of the second wave of Deep Learning M&A.  The first wave was all about arming technology giants like Google and Facebook with foundation AI technologies.  The second wave is about acquiring companies for the business value of their vertical market solutions.  Whether we’re investing in core technologies or in next-generation vertical market AI, we need to be aware of who has been acquiring what.  For platforms as well as vertical solutions, will major buyers still have a strong M&A appetite by the time our startups have matured?

Given where we are now, what are some practical rules of the road in investing in AI?  What questions should we should ask during AI investment diligence?

• Product.  Is this real, cutting-edge AI or is it a relabeled expert system or repacked BI (business intelligence)?  What kinds of AI frameworks is the team using?  Can the AI evolve, adapt and learn, or is it a static, hard to maintain and manually constructed rule-base?
   
• Data. While new methods are emerging for reducing the need for training data, such as Unsupervised Learning and Lifelong Learning Machines (L2M), most AI solutions require millions of objects in order to build and validate a neural net model.  Does the company have sufficient data to produce valid results?  Is the data accurate?  Is it biased?  If it isn’t ready today, what is the company’s data development strategy?  When will they have sufficient data to test, validate and refine their models?
   
• Talent. AI is a hyper-aggressive market, where top companies compete for customers, for investor dollars, for mind-share, and, above all, for talent.  What’s the evidence that the investee team is world class?  One way to judge is if the team can attract talent that is as good as or better than themselves.  Since M&A exits are often predicated on the presence of world class AI scientists, top tier talent should be a top tier investor concern.
   
• Regulatory Challenges. Is the startup aware of the current regulatory and cultural challenges facing AI companies?  Is the team on the path to eXplainable AI?  Are there methods for detecting and correcting the biases that all-too-human developers introduce?  And what about the company’s datasets?  Has the team adopted methods, particularly in areas like medical diagnostics, to guard against cultural, ethnic and gender data biases?
   
• Capital. Does the company have a capital strategy?  My data suggest that AI companies typically require an average of $27M in funding to reach an M&A exit.  How much capital do you anticipate they’ll need over the long term?  Is the capital requirement so great that it threatens investor outcomes via a potentially crushing liquidation preference stack?  What about the risk of a future down round or later-stage cram-down?  And given a startup’s market and focus, who are their likely next round investors?  Who are their likely acquirers?  Do the logical investors and acquirers already have companies in their portfolios similar to this startup?
   
• Exits. “Acqui-hires” (buying the company to acquire their AI talent) have been a popular form of AI exit, particularly in today’s Big Tech arms race.  But acqui-hire exits for talent alone run the risk of realizing little value for their investors.  What has been the exit experience of other companies in this sector?  What does management really want?  Do we face the risk of a low-value acqui-hire exit?
   
• Business Dynamics. Deep Learning AI is fast transitioning from theory to deployment.  We’re now building solutions for mere mortal customers for whom outcomes are more important than tools or theory. Is this company a clever algorithm, or is it on the path to becoming a real business?  Does the company have any real mainstream customers beyond the usual suspects—the tech companies, universities and government labs who buy one of everything?  Does the team know the target vertical market “in their bones?” Was the solution designed in partnership with expert customers or was it a faith-based design?

And, finally, the ultimate question: does the company produce outcomes that are materially better than the competition and, of course, better the current, pre-AI state of the practice?  If it is too early to know, when will the team be able to validate their models, methods and results in a way that’s clearly reproducible?  Ultimately it is all about outcomes and real-world results, not about AI theory.  If the outcomes are not best-in-class, why should we invest?

We’re at the beginning of AI’s latest wave of deployments in high value vertical markets.  Collectively, the opportunities may well dwarf the initial Deep Learning investing wave focused on core technologies.  But as with any emerging technology, the prudent investor should pay particular attention, not only to the hype cycle, but also to industry pragmatics.  With nearly 7,000 AI companies already in the market, smart angels have plenty of experience and data to mine to guide a potentially very profitable AI investing journey.  Like any good AI, let’s put that data to work!