A (VERY) BRIEF HISTORY OF THE AI INDUSTRY

In the early 2000s, Marvin Minsky—a cognitive scientist and one of the fathers of artificial intelligence—liked to say that we can make a computer capable of beating the reigning genius of chess, but we can’t make a robot capable of walking across the street as well as any normal two-year-old child.

The real world is not a strictly regulated closed system like a chess game. Sensing moves on a virtual board and responding within agreed-upon rules is one thing. Sensing and physically responding in actual reality—where a huge number and type of unexpected events might occur—is quite another.

In fact, the entire AI industry has been through multiple attempts since the 1980s to grow into a mature market. Each of these efforts collapsed because the technology was unable to meet the unrealistic public and investor expectations generated by non-real-world computing triumphs like those of IBM’s Watson.

SMART SYSTEMS ARE REALLY ABOUT THE DATA MODELS

As networks continue to invade the “physical” world, traditionally unique components and interfaces between and among electronic as well as electro-mechanical elements are becoming more standardized. Product and service design is increasingly influenced by common components and sub-systems. Vertically defined, stand-alone products and application markets will increasingly become a part of a larger “horizontal” set of standards for hardware, software, communications and data.

As it becomes easier to design and develop smarter more adaptive systems, competitive differentiation will shift away from unique product features towards how the product is actually used, how the product fosters interactions between and among users in a networked context, and, most importantly, how the data from the product will inform these new insights.

Machine data from cyber-physical systems of the real physical world can offer extraordinary business advantages to the companies that understand how to organize that data and model the behavior of the physical world. The ability to develop models from sensor and machine datasets allows not only data patterns but a much higher order of intelligence to emerge.

Widespread adoption of AI and machine learning systems is inevitable. But that doesn’t mean that every participant will automatically be shaking a money tree. Value and returns from AI/ML are playing a new game of hide-and-seek. They’re still there, but not where they used to be. If you keep looking in the old places…well, you know what’s going to happen. We think that the economic impact of AI/ML developments will recapitulate the tendency we’ve seen for decades in digital technology generally—less and less physical value, and more and more metaphysical value.

IT’S STILL EARLY IN THE GAME

Of course, digital computing has radically transformed human affairs. But so far that transformation has taken place entirely on the computer’s terms. Note that even the most remarkable recent achievements of AI and machine learning—autonomous driving, natural language processing, text generation, facial recognition, algorithm design and vaccine discovery—have occurred in domains of our physical environment that are subject to rigid sets of rules and laws.

We’re in the third decade of the 21st century, and the question still remains, “How many engineers in white lab coats does it take to make AI valuable?”

Rapid advancements in silicon, computing and networks are clearly forming the foundations for AI and machine learning capabilities to advance. But these systems, sophisticated as they are, are still in their early stages, and many intended use cases for AI can still be accomplished with more cost-effective traditional tools like basic regressions. It seems clear that real business value from machine learning and AI will be realized unevenly across markets, applications and use cases.

NEW AI TOOLS MUST BE ALIGNED WITH INTENDED APPLICATIONS AND USE CASES

Possibly most important for the growth of AI is that multiple parallel technology developments are now increasingly reinforcing and accelerating one another. Cloud infrastructure resources are providing unprecedented computing scale. Mobile computing devices are extending the reach of computing. Embedded systems and IoT technology are connecting and integrating a broad array of physical and digital applications. And of course the signature achievement of the age of “big data,” the ability to capture and process massive amounts of raw intelligence from the physical world, has the potential to inform many new and diverse capabilities.

Each of those technologies is powerful on its own, but creative combinations of them are what is most exciting. Human-connected devices and machine-connected IoT devices enable exponentially more data at the edge. The scale of core, infrastructural (cloud) computational capabilities enables us to capture and analyze all that information. And this in turn sets the stage for AI and machine learning tools to analyze and capture new insights.

This new chapter is motivating tech developers and users to apply advanced neural nets and deep learning tools to their most intractable problems. Most companies believe that implementing advanced AI solutions will lead to significant efficiencies, growth and competitive differentiation. However, matching new tools to high value applications and use cases will challenge many industry participants.

We all know that AI tools are trained on large data sets, but most people do not grasp that AI applications require thousands or even hundreds of thousands times more data than a human would need to solve an equivalent problem. If you examine applications where machine learning is successful, it quickly becomes apparent that they are in domains where acquiring lots of data is relatively easy—think facial or speech recognition, where technology developers have vast troves of data they can access.

Data-driven apps are the core value creation mechanisms within the Smart Systems and the IoT. But the B2B world that comprises so much of the IoT doesn’t have the same unified sources or monolithic usage tracking and analytics that the consumer world utilizes to make money. Based on our consulting work, we estimate that B2B development projects lack as much as half the data needed to inform new application values and fulfill on artificial intelligence and machine learning opportunities.

An additional challenge is the fact that most machine learning systems today run “narrow purpose” applications that can do only a single type of learning. Current neural networks cannot be trained to run multiple parallel applications, such as identifying images and playing video games, or predictively diagnosing machine failures and listening to and identifying music, all at the same time.

Finally, the impacts of new AI tools will be higher and more straightforward to achieve where the user’s propensity to experiment with new tools and methods is also higher.

WHERE TO LOOK FOR AI/ML ROI

AI and machine learning are being turbo charged. An explosion of AI/ML tools is lowering the barrier to entry to high-end data science. Historically, developing AI/ML applications and use cases involved data teams doing much “heavy lifting” to design and deploy complex custom models. Today, new data tools are gaining wider adoption. Standardized schemas for data ingestion and transformation are setting the stage for many more companies to incorporate AI/ML into their products.

Based on our research and consulting, we believe the most significant AI-related ROI will occur:

• Where users traditionally have understood the value of monitoring and collecting data
• In markets and opportunities that are typically more “mission critical” in nature
• In domains that contain diverse equipment and systems that have over many years been subject to continuous and multiple measurements
• In domains where the capital cost of the equipment tends to be high and
• In domains where the economic impact of the equipment and systems “in-use” is high, and
• Where fractional improvements or cost optimization will yield significant business results

Additionally, domains and applications where customer data is actively collected allows those data sets to be combined with adjacent data to address new opportunities such as customer prediction analytics. New opportunities are emerging all over the economic landscape to fuse very large data sets with new sensor readings and measurements. Logistics companies will use external data to predict disruptions in their retail customers’ supply chains. Smart city emergency response services will get to accidents or disaster sites and on to hospitals as quickly as possible leveraging imaging data, maps, traffic flow data and much more. Combining weather forecasts and legacy weather models with sensors mounted on vehicles that contribute real time road-condition measurements will help large fleet owners optimize the use of their trucks.

In order for developers to provide appropriate tools and services, and for users to effectively justify the significant investment in AI/ML capabilities, identifying and aligning the tools with the intended use cases and applications will be critical.