IN 2004, IN THE MAZE-LIKE aisles of Stanford’s computer
science department, I spoke to a man who resembled Santa
Claus. This bearded man was John McCarthy, who coined the
term Artificial Intelligence in the 1950s and was one of the
founding fathers of Artificial Intelligence, along with Marvin
Minsky. McCarthy spearheaded the effort for some time,
including creating the language Lisp for the purpose of AI,
among other innovations like time-sharing for computers,
garbage collection, and lambda calculus. I was a graduate
student studying natural language processing, and AI wasn’t
as cool as it is today. Neither was natural language processing.
It was far from the awe-inspiring concept it has become. But
the thawing of the so-called AI winter was starting.
On that day in 2004, I stared at an old thermostat in the
room, and my conversation with John McCarthy moved from
the inability of relational databases to be introspective to AI.
The thermostat was the boring kind found in every university
and hospital. John, however, believed that thermostats could
“think” and “have emotions” and “beliefs”, as described in
his essay found at http://stanford.io/2alWwVr. He was disappointed
at the state of the affairs of AI—or databases,
for that matter. I don’t know if the founders of Nest know
of or were motivated by his thinking on thermostats when
they invented their beautiful device, but every time I look at a
Nest, I remember John McCarthy and how quickly we moved
from a dumb thermostat in that office to the one from Nest.
While I am not sure how much my Nest “believes” in things, it
certainly does a good job at managing its narrow task. It has
models that predict the future and goals—set by me—that
drive its behavior. It may not be everything we think as AI,
but in only a few years, the thermostat moved a lot closer to
McCarthy’s vision.
Learn faster. Dig deeper. See farther.
Professor John McCarthy passed away in 2011. Quite rapidly,
AI moved from the labs of computer science departments and
failed research attempts into the real world. The question I
often ponder, along with everyone in the field, is whether AI
is here to stay or if we are susceptible to another AI winter.
The majority of people involved in AI are quite pragmatic and looking to solve practical problems, which gives me confidence.
I appreciate every Watson commercial that I see on TV
because IBM is investing valuable marketing dollars in popularizing
the AI vision, but I also
get nervous about the possibility
of over-promise and under-
delivery by this very nascent
technology.
Everyone is jumping into the
fray. The CEO of Google recently
announced that AI and
machine learning will the central
component in all of their products.
They are actually trailing
companies like Amazon, which
have already released really smart home products like Echo
and Alexa based on AI and natural language understanding.
This report aims to cover the current market of AI and its commercial
adoption beyond the academic labs into industry. We
are at the cusp of mass adoption of AI. Big market predictions
are being thrown around, and we must
ground where we are with data. The goal
is to create a landmark that can be used to
study its future growth, though I do try to
provide some color along with data on what
is happening in the current business world
around AI. The backing data provided is
meant to be stand-alone and my comments
are just one interpretation. The goal of the
report is to provide guidance to industry
on how their peers are adopting AI, and its
general direction and use cases. The report
makes no claims of predicting AI’s future, and the scope of
the project is restricted to companies operating in the U.S.
The Primordial Soup for Artificial Intelligence
AI REALLY JUMPED INTO MAINSTREAM industry in 2011
and 2012—ironically, right after the death of its founding
fathers, McCarthy and Minsky. Turns out, there were many
material reasons for AI to sprout around this time and many
foundational technologies came together to create this perfect
storm. The following are some technological innovations
and market conditions that made AI accessible to mainstream
developers and companies around the world:
- Big data Infrastructure
- The original MapReduce paper by Google spawned projects
like Hadoop, which provided the infrastructure required for
cheap, massive data processing required by AI. - Cloud computing
- This advancement provided the ability for a graduate student
to hire 100-node machines for a data processing job
for a mere $1000, something that would have previously
required $100 million in investments to build. - Massive amounts of data
- Open source crawlers like Nutch have made knowledge accessible
on the Internet. Also, copies of most pages found
on the Internet are easily available to everyone thanks to
open source repositories like commoncrawl. - Watson and Siri
- While not always impeccable, both Watson and Siri should
be credited for popularizing AI and making it approachable
to the masses. - Venture funding
- Since 2009, over $10B of venture funding has been invested
in the big data infrastructure required to build today’s AI
applications. - Qualified people
- The number of people who can perform the various tasks
for AI development, from data processing to data science,
has grown tenfold.
A Word of Caution
BEFORE I DIVE INTO THE current state of AI in the business
world, I would like to point out that most technologies
available today are still far from a generalized AI. I define
generalized AI as a system that can reason about the world,
understand general problems, and solve them at super-human
or even human-level intelligence. The main argument against
modern peddlers of AI is that most are trivial bag-of-word
models (aka counting words) being passed off as AI—they
cannot think or do anything labeled as cognition. Please refer to
these series of blogs by an AI researcher (http://bit.ly/2aCfYyN)
on why we should not peddle our current AI as the AI promised
to us in the 1960s. There are claims that Google DeepMind
is generalizable and in theory it looks like one, but to me,
it’s still just playing games, and we haven’t seen any other
application.
Having said that, there are pockets of technologies available
that are achieving human- or super-human-level intelligence
on a given human task (aka “narrow intelligence”). And it’s
not just a task done by a disinterested party, but a human
who is very good at completing it. DeepMind beat the Go
master—an achievement since, unlike Chess, Go requires
human intuition and not just brute force calculations of every
move. Image recognition has also reached and sometimes
beat human-level performance. For example, our own AIbased
engine for Spiderbook can best a good salesperson in
coming up with a list of customers to target.
However, it is not the purpose of this report to argue what
Artificial Intelligence is or is not’. Rather, I take a practical
approach to the definition of AI and present an analysis based
on self-identified businesses that claim to be using or building
AI. I do not attempt to verify what people are calling AI, or
discern between “good” AI or “bad” AI.
Research for this Report
To conduct research for this report, my team used
a graph-based machine learning model developed at Spiderbook
that learns industry vocabularies around AI, reads the
entire business Internet, and then classifies businesses into different
levels of maturity and investments in AI. We canvassed
almost 500,000 companies around the globe to develop a
data-driven, in-depth understanding of the AI landscape and
various related technologies, like cognitive computing, deep
learning, machine vision, natural language understanding, and
chatbots. The engine reads and understands billions of publicly
available documents, including all press releases, business
relationships, forums, job postings, blogs, tweets, patents,
and proprietary databases that we have licensed. We use this
data, which largely represents the business Internet, to create
a knowledge graph that represents how companies are interlinked
and who is using what products or has employees with
given skills. On top of this knowledge graph, we performed
network-based machine learning to create a near-real-time
snapshot of a company’s priorities, projects, and investments.
Let’s dive into the results.
Investment in AI by Industry
As one would expect, the largest share of AI is being used by
software and IT-related companies. Although the figure that follows provides a breakdown of the industries investing in AI, the actual counts are still very low. Only a few dozen companies in each industry, outside of software and IT, are actually
involved in AI.
Investment in AI by Company
There are only 1,500 companies in North America that are
doing anything related to AI today, even using its narrow,
task-based definition. That means less than one percent
of all medium-to-large companies across all industries are
adopting AI.
The table on the following page shows some of the companies
that are actively investing in AI, organized by industry.
Even though less than one percent of companies in any industry
are adopting AI, the companies that are adopting it seem
to be the leaders of their industry. They are household names
and the biggest, most successful companies in their fields. It’s
hard to discern the causal reason for this finding: is it because
they are paranoid of their leadership positions? Or do they
have extra resources to try out any new ideas, not just AI? Or
perhaps these are the early adopters, laying out the groundwork
for others in their respective industries to follow?
Top Companies Investing in AI
The following list shows the companies investing the most
in AI, and talking about it as a core strategic driver for their
business. There are the usual suspects, such as Google and
Facebook, but also companies like MITRE Corporation, a
nonprofit that operates federally funded research and development
centers, that aren’t household names:
- Rocket Fuel
- IBM
- Amazon
- Yahoo!
- Intel
- Microsoft
- Deloitte
- MITRE
- Baidu
- Apple
- Cylance
- Lockheed Martin
- NASA
- Sentient Corporation
- Electronic Arts
Use Cases of AI
I recently watched a panel of luminaries in AI, organized by the
Milken Institute, speaking about their vision on what is going
on in AI and what’s now possible using such technologies.
Ideas suggested by the panel were a lot more exciting—some
extreme, and many more humane—than the actual applications
of AI today. The ideas ranged from human disease diagnostics
to farming to elderly care. However, based on our machine-intelligence-
based research, the predominant applications of AI
seem to be more banal and routine automation of tasks done
by humans. The figure that follows quantifies how corporate
budgets are being spent on specific AI-based use cases.
There are some novel applications in this graphic that are beyond
task automation. For example, use cases like telematics,
IoT, and robotics have industry-wide implications, and represent
more than just human task automation.
Cyber-Intelligence and Security: A Major Driver for AI
It is also surprising to see such a wide application of AI in the
world of cyber-intelligence, an area that isn’t a big topic of
conversation in AI circles yet, although large amounts of budgets
are clearly being invested in this area.
There are more companies building, consulting, or using AI for
cyber-intelligence than any other use case. Perhaps there are
more threats in society than what’s reported, since companies
do not have natural incentives to publicize them. Or, perhaps
this is an epiphenomena of continuous funding from the U.S.
government focused on this vertical.
Adoption of Technologies in AI-Mature Companies
Over the last decade, there have been waves of AI-related algorithm
du jour for solving classical problems such as classification
or natural language processing. Some algorithms stick around for
larger adoption, based on their efficacy and applicability to the
problems, but most fade out. Latest innovations in algorithms
have been in the area of deep learning, a position previously held
by latent dirichlet allocation (LDA), semi-supervised learning,
Latent Semantic Indexing (LSI), Support Vector Machines, and so
on. Some of these technologies have become a class all their own,
even though there is a lot of overlap in the problems they solve.
For example, deep learning can be used for natural language
understanding (NLU), cognitive computing, or even autonomous
vehicles, although it’s mostly used for image processing.
A breakdown of AI adoption does not provide a fair picture of the
current level of AI maturity in the market. The following two figures
detail how many companies are using these AI technologies
beyond lab experiments (i.e., those developing applications based
on it or deploying it across the company).
The second figure that follows shows subcategories of AI technologies,
and the number of companies investing in those spaces.
Physical Locations of AI-Oriented Companies
The physical presence of companies adopting AI is very bi-modal, even more than high-tech in general. For example, my previous report on The Big Data Market showed companies more geographically distributed than the companies adopting AI. Perhaps we can all guess where the singularity will arise. See the figure that follows for a state-by-state breakdown.
Summary
The AI winter has thawed, and it’s moving into spring. While
the AI movement is still very much in its infancy, the promise
and recent gains around task-based AI has created a buzz.
There are some focused sets of use cases like cyber intelligence,
sales and marketing, and manufacturing automation
that already have AI-based products in the market. Generalized
problem solving and healthcare applications are topics of
intense conversation, but not many companies are investing in
them with their budgets.
The rise of AI was triggered roughly five years ago by government
and private investments in big data technologies, cloud
infrastructure, and, most importantly, the general availability
of talent. Just within the last few months, major companies
like Amazon, Tesla, and Google have highlighted AI as the
driver for the next decade of innovation within their companies.
But there is a small core set of companies adopting AI
across the board, and less than 1,500 companies operating
in the U.S. are investing anywhere near the space. While that
is a small percentage of industry, the companies leading the
movement are the biggest and brightest, and certainly have
the most to gain (or lose) from AI becoming a reality.
