The wish to construct machines that can perform tasks just like us or better than us is as old as our ability to reason about the world and question how things work.
They were made of gold but looked like real girls and could not only speak and use their limbs but were also endowed with intelligence and had learned their skills from the immortal gods. While they scurried around to support their lord, Hephaestus moved unsteadily to where Thetis was seated.1
Building robots was the job of gods. They breathed life into machines. Homer was telling us that the ability to create thinking machines could bestow on us god-like status.
The challenge we had back then, and still have, is understanding exactly how we might go about building such machines. While we could imagine their existence, we didn’t have the tools or methods that would allow us to chart a path to an actual machine. It is no wonder that very often when solutions were imagined, they included secret potions and alchemy that would magically breathe life into Frankenstein-like figures not entirely under our control.
Step by step, though, we have put some of the pieces of the puzzle together. At the mechanical level, humans managed to build very convincing automatons. Through clever tricks, the creators of these automatons even fooled people into believing they were magically endowed with intelligent thought. From the ancient Greeks to the Han dynasty in China, the water-operated automatons of al-Jazarī in Mesopotamia, and Leonardo da Vinci’s knight, we have always tried to figure out how to get mechanical objects to move like real-life objects while tricking the observer into thinking there is an intelligent force within them causing action.
From a reasoning perspective, we went from the beginnings of formal reasoning, such as Aristotle’s syllogisms, to understanding how we can describe complex processes through algorithms (the work of al-Khwārizmī around 820 A.D.), through to Boole’s The Laws of Thought, which gave us formal mathematical rigor.
Nevertheless, we were still woefully ill-equipped to create complex systems. While lay people could be fooled through smoke and mirrors, the practitioners of the field knew that their systems where nowhere near close to the level of complexity of human (or any form of natural) intelligence.
As advances marched on and we got to the 19th century, the field of computer science started taking shape. With Charles Babbage’s Analytical Engine and Ada Lovelace’s work on programming, the outlines of a path from imagination to realization started emerging. Ada Lovelace speculated that the Analytical Engine, with her programming, “might compose elaborate and scientific pieces of music of any degree of complexity or extent.”
By the end of the Second World War, we had progress in computing that was the result of efforts to build large code-breaking machines for the war, and the theoretical advances made by Alan Turing. A few years later (1950), Alan Turing even provided us with a test to apply to machines that claim to be able to think. The test calls for a human being to hold a conversation with a machine while a third observer is able to follow the conversation (by seeing what was said printed out). If the third observer cannot distinguish between the human and the machine, the machine passes the test.2
The path toward artificial intelligence was getting increasingly clearer.
The Birth of a New Field of Study
On August 31, 1955 a group of researchers in the United States produced a brief document3 asking for funding for a summer research project. The opening paragraph is a testament to the unbounded optimism of humans and a lesson in what the phrase “hindsight is everything” means.
We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire. The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it. An attempt will be made to find how to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves. We think that a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.
The most striking phrase, to a 21st century reader, is the very last one. It claims that a significant advance can be made after just a single summer of work. Keep in mind that this was with 1956 computing technology. The first integrated transistor was 4 years away. Computers occupied entire rooms and logic modules used vacuum tubes.
Now, to be fair, the group of scientists in question was exceptional. The proposal was co-signed by four very influential computer scientists. Claude Shannon, the father of information theory; Marvin Minsky, one of the first to build a randomly wired neural network; Nathaniel Rochester, the lead designer of the IBM 701, the first general-purpose, mass produced computer; and John McCarthy, widely credited with coining the term “artificial intelligence” and the creator of the Lisp programming language. If any group of people stood a chance of making a significant breakthrough in 1956, this was certainly it.
This meeting in Dartmouth is generally credited as the first conference on AI. It shaped academic thought around the range of issues to be dealt with in order to create machines that exhibit intelligence. It also produced a lot of excitement and buzz and led to several years of funding for AI, largely based on the trust people had in people like Minsky to pull the proverbial rabbit out of the hat.
Those golden (in an almost literal sense) years laid the foundations of the field of AI. From 1955 to the early 1970s, a whole range of subfields was created, and problems and their challenges stated. From different ways to reason to a start at natural language understanding, the first neural networks, and so much more, the field was booming.
As ever, however, buzz and excitement, unaccompanied by the results that people expected or predicted, led to what was known as the first AI winter. People, initially bedazzled by hyperbolic claims, were eventually disillusioned with AI and lost hope in its ability to produce results. AI was sidelined, funding was reduced, and focus turned to other issues.
It’s important to note that research in AI did not stop over this period. It may have been less visible to the public eye, and overall resources were reduced, but interest remained. Research grants used the term “AI” less often, but they were still trying to solve the same problems.
A Practical Application
In the 1980s there was a resurgence. AI had found a problem that it could solve for businesses in a manner that was widely applicable and where returns were clear. By that time, several research groups, but in particular the Stanford Heuristic Programming Project, came to the realization that instead of trying to create general problem solvers, they could instead focus on constrained domains that required expert knowledge. These expert systems used all the research that happened in the previous decades about how to codify knowledge and reason about it, and focused it on specific use cases within restricted domains.
Expert systems can broadly be described as the combination of codifying knowledge as a set of rules and creating an inference engine that can take in a description of the state of things and derive some conclusions. Roughly, you would have a collection of rules such as “If temperature is below 60°F, wear a warm coat” and relationships such as “A jacket is a type of coat.” Combining large numbers of such rules and relationships, expert systems can capture an expert’s highly complicated knowledge in a specific field and act as an aid to augment human-based expertise. By the mid-1980s there was a whole industry of companies focused on supplying technology to run expert systems. These software behemoths used programming languages such as Lisp to encapsulated knowledge, and used dedicated machines to be able to crunch through the rules.
The iconic expert system example from the 1980s was called XCON (the eXpert CONfigurer), built for Digital Equipment Corporation. XCON’s job was to assist in the ordering of DEC computers. The inputs were the client’s requirements and the outputs were the necessary computer system components. XCON was built based on the knowledge of DEC’s experts and had around 2,500 rules. It ended up saving up to 25 million dollars a year because it processed orders with a higher level of accuracy (between 95% and 98%) than technicians, thereby reducing the number of free components DEC had to send out following a mistaken configuration.
With expert systems, entrepreneurs saw an opportunity; AI had found its killer feature, and business was booming again. Once more though, hype got the better of everyone. Big claims were made and, while a lot of useful software was built and was effectively being used, the expectations were set too high. As a result, that industry crashed and several companies disappeared. AI was in the doghouse once more.
What is a less-often mentioned feature of this second AI winter is that a lot of the industry (i.e., a lot of the money involved) was focused on building specialized machines able to run expert systems. It’s important to keep in mind that the PC industry was still confined to hobbyists at this time. What the expert systems industry did not foresee was the rise of PCs. They became popular and were recognized as valid and useful business machines, making the expensive expert systems machines seem less attractive.
Quite simply, there was no need for a large industry building actual computers to run expert systems, because people could get a lot done with normal and cheaper PCs. Business could do digital transformation in a much more agile way by introducing PCs.4 Giving everyone a word processor and a flexible number cruncher and letting them figure out how to make good use of them seemed like a much better way to invest money at the time. Part of the story, therefore, is that the technological landscape changed, and the AI industry failed to adapt quickly enough.
Businesses use technology to gain a competitive advantage. At the time, investing in large and complicated AI systems was risky. It made much more sense to bring the entire company up to date with computing technology and empower many more people within the company to benefit from general computing improvements.
Understanding these ebbs and flows is fundamental in learning to distinguish noise from signal. All through those 40 years (from the early 1960s to the early 2000s) capable people built capable software that solved real problems using AI. They also advanced the science of how we could solve complex problems with computers. The battle was not really fought or measured in that respect. The AI winters or springs were measured in the ability of people to attract funding (which induces people into making lofty claims) and the competition AI had from other types of technologies vying for investment. By the early 1990s AI generally had a tarnished name, and there were so many new, exciting, and much more immediately applicable technologies to invest in (like a little something called the World Wide Web).
AI Goes into Hiding
As a result of all of this, AI researchers devised a different survival strategy. Instead of saying they were working on AI, they would focus on talking about the specific subfield they worked in. A whole range of different fields emerged, from multiagent systems (the field I worked on as a PhD student and researcher) to knowledge management, business rules management, machine learning, planning, and much more. What we were all doing was trying to build machines that solve problems in a better way. We just stopped emphasizing the AI part that much. Even the researchers who were directly focused on the hardest task of them all, using computers to understand or recreate human behavior, labeled it as “cognitive computing” or “cognitive sciences” rather than calling it research in artificial intelligence.
In the background, gatherings such as the International Joint Conference on Artificial Intelligence (running every 2 years from 1969 to 2015 and now, incidentally, running yearly) remained. The people who were reticent to talk about AI on funding proposals still attended. Everyone knew they were working on AI, but they would not necessarily hype it as much.
My own personal path to AI reminds me of just how much AI was out of favor. As an undergraduate at Warwick University, in 1999, I had to get “special” permission from the Computer Science department to attend a cognitive computing course because it was taught in the Psychology department! I ended up loading my undergrad studies with so many AI-related courses that I was called in to justify it as still a valid degree in Computer Science. Luckily, all the AI academics-in-hiding had my back and the argument was won. Fast-forward to 2019 and universities are rushing to introduce Artificial Intelligence degrees while academics are reclaiming titles such as Professor of Artificial Intelligence, when only 5 years ago they were “plain” Professors of Computer Science.
What all this helps illustrate and what is essential to keep in mind when you hear about AI having gone through cycles before and AI being just hype, is that AI is an extremely broad term. It is such a broad term that it is arguably close to useless—unless, that is, it captures the imagination of investors and the popular press. The problems it is trying to solve, however, can’t go away. They are real scientific questions or real challenges that humanity has to tackle. Separating the noise of the press and the hype that gets investors excited from what matters is crucial in approaching AI. In the next section, I will explain why I think we have reached a point where talking about cycles of AI is not very useful, and in the next chapter we will get a better handle on what AI actually means.
The New Eternal Spring
Since the late 2000s to now, 2019, excitement about AI feels like it has reached fever pitch and will need to settle down a bit. At the same time, a lot of other things have been happening in the world of technology and outside of it. These other elements have fundamentally changed the landscape once more.
This time around, unlike the 1980s, I believe that the way the landscape is changing is in favor of the long-term growth of AI and the growing industry of AI-related technologies. The changes are not only going to help spread AI, they are going to make AI a necessity. AI will stop being a novelty application that is occasionally introduced. Instead, it will be one of the fundamental pillars woven into the fabric of everything that we do.
We may call it different names and investors may lose interest once more, but the tools and the solutions to problems they provide will remain. We can already see this happening explicitly with smartphones, where dedicated chips for AI calculations are introduced. If those were switched off, we wouldn’t be able to even open the phones (face recognition) or type on them (predictive typing).
- 1.
We have an unprecedented increase in the amount of data we handle, and the need to carefully curate it and base decisions on it.
- 2.
Processing power, which has followed Moore’s law for the past 50 years, is making it possible to apply complex algorithms to data. The performance of these algorithms for well-defined domains is equaling or surpassing human abilities. At the very least, these make the tools useful companions, augmenting human capabilities.
- 3.
Cloud computing is making both storage and processing power widely available with an added layer of sophistication, so that anyone can access and make use of complex AI tools. Open source tooling is doing the same in terms of the lower level capabilities. AI is being democratized in a way that was never possible before.
Flooded in Data
The first fundamental shift is the extent to which what we can do with data has changed. Data has always been produced. The question is whether it could be captured, stored, retrieved, and analyzed. In that respect, the past years have introduced magnitudes of change. We can now, very cost effectively, store huge amounts of data (although most of it is still unstructured data), and we can process huge chunks of the data that we do store.
80,000,000 MRIs taken every year
600,000,000 forms of malware introduced daily
2,200,000,000 locations generating hyperlocal weather forecasts every 15 minutes
Over 115,000,000,000 emails sent in the day already
Over 35,000,000 photos uploaded to Instagram
Over 3,000,000,000 searches on Google already
For once it is no exaggeration to use terms like “flooded” or “inundated” to describe the situation with data. According to research done by Raconteur, we are likely to reach 463 exabytes5 of data produced daily by 2025.6 That is going to be driven to a large extent by the increase in devices that are connected to the Internet of Things (IoT) in wearables, smart devices around the home, and in industry.
The AI opportunity is simply that humans can in no way, shape, or form expect to analyze even a tiny fraction of this data without automation. Applying large-scale data analysis to it is our only hope. Taking this a step further, we can also expect that unless we embed automated decision-making into systems, we will not be able to cope with the growth in demand for decision making and action.
To consider some of the staggering needs humanity has, here is one example from education. According to goals adopted by the UN General Assembly, in order to meet a 2030 education goal of providing every child with primary and secondary education, an additional 69 million teachers will be needed across the world.7 As of 2018, there were 64 million teachers around the world8 (this is ignoring quality issues such as level of training of the teachers themselves). We would need to double the number of existing teachers worldwide in 12 years. To train enough teachers to meet our goals, we need a huge amount of support from technology to scale teacher training, scale pupil education, and more accurately measure results. All of this will depend on effectively managing data and automating processes, to allow human resources to focus on what they can do best—person-to-person interactions.
In China, a company called Squirrel AI has managed in 2 years to build over 700 schools for K–12 education, catering to over 1 million students.9 The entire system depends on sophisticated algorithms that are able to adapt a teaching program to a student’s individual needs. While it is healthy to be skeptical about how such systems will really impact children and education, we have to, at the same time, accept that the exploration of such solutions is the only way to manage a growing population and growing needs for education.
Raw Processing Power and Better Algorithms
What do first-person shooter games and AI have in common?
Believe it or not, without the former we might not have had quite the resurgence of the latter. In 2009, at the International Conference on Machine Learning, Raina et al.10 presented a paper on the use of graphics processors applied to large-scale deep belief networks. Raina’s research group at Stanford, led at the time by the now well-known Andrew Ng,11 figured out how to effectively use graphical processing units (GPUs), specifically through NVidia’s CUDA programming model, to dramatically decrease the time it took to train machine learning models by exploiting the parallelization afforded by GPUs. By that time, GPUs were so efficient and very cost-effective because of the increase in sales of GPUs to power the needs of the computer gaming industry for, point in question, better graphics for first-person shooter games.
The use of GPUs, coupled with advances in algorithm design (in particular from Geoffrey Hinton’s group at the University of Toronto), led to a step change in the quality of results and the effort it took to arrive at those results.
In many ways, these two advances have released the genie out of the bottle. AI as an industry may suffer at some point, investors may complain, and we may not get all the things that are being promised. The techniques developed so far, however, will always be available, and because of their wide applicability they will always be used to solve problems—even more so as we see access to AI democratized through cloud computing and open source software.
The Democratization of AI
Up until 2015, actually using AI technologies was generally hard. Unless you had AI experts in your team, you couldn’t realistically attempt to apply any of these technologies. A lot has changed since then.
In November 2015, TensorFlow12 (Google’s framework for building machine learning applications) was released as an open source tool. Since then, the open source space of AI tooling has exploded. Online courses abound, both paid for and free, and the main blocker to learning about AI techniques is your time availability and access to bandwidth.
In addition, by 2019 we have an unprecedented level of access to sophisticated machine learning services through simplified application programming interfaces. Amazon, Microsoft, Google, IBM, and many others offer access to their tools, allowing organizations to upload data, train models, and deploy solutions within their applications.
Microsoft, in particular, talks directly about democratizing AI.13 The aim is to make it available to all software engineers in the same way database technologies or raw computing power is available via cloud-based solutions.
The wide availability of these technologies as cloud solutions doesn’t just reduce the level of expertise required to implement an AI-powered application. It also reduces the amount of time it takes to go from idea to prototype to production-level roll-out. As we will see later on, the ability to iterate, experiment, and learn while doing is just as important as having the technologies readily available.
Moving Forward
In the last section we laid out the case for why AI cannot simply disappear. Please note that this is not about the economics of the AI industry. It is not about the venture capitalists, startups, or large corporation and government politics of who gets more attention and funding. The argument is about the fundamentals of how the digital world is evolving and the needs of the analog world. The problems are getting bigger, and we are running out of ways to scale systems unless we introduce automation.
Whether it is agriculture, education, health, or work in the office, we cannot just keep working longer and stressing out more. Something has to give and something has to change. My hope is that what is going to change is more efficient use of technology to allow us to focus on what is more important. Yes, along the way we will go through booms and busts. However, do not confuse press coverage or funding news about unicorn startups with the fundamentals. The goal of AI should not be to make us all more like machines or to create unicorn startups.
The goal of AI, as that of any technology, should be to lift us up from our current condition and give us back the time to explore what it means to be human.