BETA
This is a BETA experience. You may opt-out by clicking here

More From Forbes

Edit Story

AI Decoded: Evaluating Artificial Intelligence Startups

This article is more than 4 years old.

Since the term was coined in 1956, “Artificial Intelligence” has endured a lifetime of misunderstanding.

The root of the problem lies in the interpretation of the word “intelligence.” In the words of legendary computer scientist Edsger Dijkstra: “The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.” If I were to ask you who is a better swimmer, would you choose Michael Phelps or the USS Illinois? Swimming is a human activity, certainly not something a submarine does. Yet, a submarine travels much faster and over much greater distances than any swimmer. 

Intelligence is like swimming; we associate it with living beings. Adding to the confusion, humanity has a long history of "moving the target" on what constitutes intelligence. In 1642 Pascal wrote that a device "will eventually perform all four arithmetic operations without relying on human intelligence.” Pascal, it seems, would have considered the ability to add, subtract, multiply, and divide as intelligence and the ability to do so on a calculator would have been a form of artificial (machine) intelligence. Fast forward to today where a calculator is considered rudimentary. Herein lies the problem: by promising intelligence, AI sets a high, ambiguously defined bar that undermines all that it actually can and has achieved.

Though we're certainly in an AI boom, the technology itself isn't new. Rather, AI is the continuation of the computation revolution in the same way the internet was the continuation of the communication revolution. The roots of each revolution run deep. Electronic communication was invented in 1827 with the telegraph. Like the telegraph, the internet is still fundamentally electrons relaying encoded information over a medium. The differences are simply speed of transfer, robustness of protocols, and number of nodes on the network. 

The calculator is the computation revolution’s equivalent to the telegraph, carrying out rudimentary calculations just like the telegraph transferred rudimentary communications. Today’s personal computers evolved from calculators. Artificial intelligence is the next step in this progression: it is still a series of computations, but now enabled by more processing power, advanced algorithms, and data availability. Although artificial general intelligence (which aims to create a general system that can learn anything a human can) gets a lot of attention in film, that field is a long way off. Today, AI is about solving specific problems, and we have come to the point of maturation where it is actually yielding commercially viable businesses. 

Over 350 companies applied for a spot on the inaugural Forbes AI 50 list. Some of the companies are on the path to commercial adoption, but most have launched and are already earning revenue. Many are even reeling in 8-figures.

The 50 featured companies fall into two categories of business:

  • Horizontal AI is basically AI infrastructure, and is the "picks and shovels" for solving problems with artificial intelligence. These are companies that sell tools to help their customers implement AI. DataRobot, Domino Data Labs, and Scale are examples of companies that provide"horizontal AI." Specifically, DataRobot provides a fully-automated AI product that a business analyst with modest training can use to build virtually any prediction engine, given the right data. Similarly, Domino Data provides an opinionated framework that enables advanced AI engineers to accelerated development with tools like version control that are custom made for AI applications. Further, a business like Scale outsources the data-labeling problem so that autonomous vehicle teams can focus on building software, not labeling data. 
  • Vertical AI companies solve a specific vertical problem using AI. For example, ClimaCell uses AI to better predict weather microclimates, Kodiak to operate semi-trucks, Viz.ai to accelerate the identification/treatment of strokes, and Verkada for security and people identification.

Regardless of whether a company provides horizontal or vertical AI, it’s only interesting if it’s solving a real-world problem. When evaluating a company I always think about the problems that the AI company solves, how it solves those problems, and why its solution is superior to alternatives. The customer is the ultimate judge of if an AI product is working. 

Disclosure: Meritech Capital is an investor in DataRobot and Verkada, which are used as examples. Konstantine is an individual investor in Viz.ai.

The question of "how" the AI works is important to a company's defensibility, scalability, and future performance. Broadly speaking, most of AI today is a form of optimization. Many search algorithms could even be considered AI. However, the most notable subfield of AI today is Machine Learning, which optimizes a “cost function” using training data to come up with a prediction, like whether an email is spam. Training data is historical data that helps inform a prediction. For example, when predicting whether an email is spam, a machine learning product  could analyze millions of past emails, some of which are identified as spam and others which are not, so as to distinguish the two email categories. This is called “supervised learning” because the algorithm is supervised via label of data. The main types of supervised learning are classifications (for example, "spam" or "not spam") and regression (for example, predicting the time of day someone is most likely to open an email). Unsupervised learning, on the other hand, involves no labeling. We usually use unsupervised algorithms if we want to find patterns but aren't quite sure what we're looking for a priori. For example, we could run a clustering algorithm on unlabeled emails and find that emails relating to scheduling meetings cluster in one group while emails related to explaining AI cluster in another very different part of the hyperspace. Deep learning is a set of algorithms that in recent years have proven to be particularly fruitful in solving many supervised and unsupervised problems. The below chart isn't exhaustive (for example it doesn't capture all the exciting generative algorithms out there), but it is a high-level taxonomy of AI.

Finally, when evaluating a company I ask why its solution is and will continue to be superior to alternatives. A solution isn't inherently more valuable just because it uses AI. There must be some reason why AI is advantageous. 

  • Speed can be an advantage when time is of the essence. In the case of Viz.ai’s stroke detection technology, every minute equates to a loss of 1.9 million brain cells, so an algorithm analyzing an image faster than a human makes a meaningful difference.
  • Cost can also be a driver. For example, Scale.ai uses AI to pre-label data, thus delivering high quality results at lower prices.
  • Accuracy matters, especially for tasks that are tedious or boring to humans. Verkada can accurately analyze hours of video footage in seconds, in a powerful way humans can’t.

In order to determine the continued viability we must also consider defensibility:

  • Technology defensibility isn’t guaranteed just because a company uses AI.  Companies like DataRobot and Algorithmia have hundreds of brilliant engineers working to simplify AI for their customers to consume. More so, there are a plethora of open-source frameworks that enable engineers to leverage AI quickly and easily. Although accessibility is wonderful for the industry, it is also important to have a technology moat. To understand defensibility it’s imperative to understand the methodology that the technical team is employing. The more insightful, process-driven, and perpetually-improving the approach, the better. Another proxy for defensibility is the caliber of the AI team. If it's packed with people who studied or researched the cutting edge of what's possible, that's a good sign.
  • Data defensibility relates not only to the unique access to data, but also to whether that data will improve the product over time. If a company is using nothing but open source data, others using similar algorithms will come to similar conclusions. Yet, if a company has proprietary data that improves the accuracy and speed of an algorithm, that is a long-term sustainable advantage.

 Artificial Intelligence has some big shoes to fill, especially since “intelligence” is a moving target. AI is both much less and much more than we think it is. Though it would be unfair to expect humanoid robots powered by AI in the near future, AI is ready to deliver solutions to important real-world problems with remarkable speed, cost savings, and accuracy.

Follow me on Twitter or LinkedIn