We’ve built pyramids. We learned to harness electricity. We understand how our universe formed after the Big Bang (mostly). And of course, the cliché example, we’ve put a man on the moon. Yet, in spite of all these accomplishments, the human brain remains far outside the understanding of modern science and is, by default, the most complex object in the known universe—or at least our understanding of it.
Given this reality, it shouldn't be altogether shocking that we haven't yet built an artificial intelligence (AI) on par with humans. An AI like Data (Star Trek), Rachael (Blade Runner), and David (Prometheus), or non-humanoid AI like Samantha (Her) and TARS (Interstellar), these are all examples of the next great milestone in AI development: artificial general intelligence (AGI, sometimes also referred to as HLMI or Human Level Machine Intelligence).
In other words, the challenge AI researchers are facing is: How can we build an artificial mind comparable to our own when we don't even have a full understanding of how our own mind works?
We’ll explore this question, along with how humans will stack up against future AGIs, and finally, how society will change the day after the first AGI is announced to the world.
What is an artificial general intelligence?
Design an AI that can beat the top-ranked players in Chess, Jeopardy, and Go, easy (Deep Blue, Watson, and AlphaGO respectively). Design an AI that can serve you answers to any question, suggest items you might want to buy, or manage a fleet of rideshare taxis—entire multi-billion dollar companies are built around them (Google, Amazon, Uber). Even an AI that can drive you from one side of the country to the other ... well, we're working on it.
But ask an AI to read a children’s book and understand the content, meaning or morals it’s trying to teach, or ask an AI tell the difference between a picture of a cat and a zebra, and you’ll end up causing more than a few short circuits.
Nature spent millions of years evolving a computing device (brains) that excels at processing, understanding, learning, and then acting upon new situations and within new environments. Compare that with the last half century of computer science that focused on creating computing devices that were tailored to the singular tasks they were designed for.
In other words, the human-computer is a generalist, while the artificial computer is a specialist.
The goal of creating an AGI is to create an AI that can think and learn more like a human, through experience rather than through direct programming.
In the real world, this would mean a future AGI learning how to read, write, and tell a joke, or walk, run and ride a bike largely on its own, by way of its own experience in the world (using whatever body or sensory organs/devices we give it), and through its own interaction other AI and other humans.
What it will take to build an artificial general intelligence
While technically difficult, creating an AGI must be possible. If fact, there's a deeply held property within the laws of physics—the universality of computation—that basically says everything a physical object can do, a sufficiently powerful, general-purpose computer should, in principle, be able to copy/simulate.
And yet, it’s tricky.
Thankfully, there are a lot of clever AI researchers on the case (not to mention lots of corporate, government and military funding supporting them), and so far, they’ve identified three key ingredients they feel are necessary to solve in order bring an AGI into our world.
Big data. The most common approach to AI development involves a technique called deep learning—a specific type of machine learning system that works by slurping up giant amounts of data, crunching said data in a network of simulated neurons (modeled after the human brain), and then use the findings to program its own insights. For more details about deep learning, read this.
For example, in 2017, Google fed its AI thousands of images of cats that its deep learning system used to learn not only how to identify a cat, but differentiate between different cat breeds. Not long after, they announced the impending release of Google Lens, a new search app that lets users take a picture of anything and Google will not only tell you what it is, but offer some useful contextual content describing it—handy when travelling and you want to learn more about a specific tourist attraction. But here too, Google Lens wouldn't be possible without the billions of images currently listed in its image search engine.
And yet, this big data and deep learning combo still isn’t enough to bring about an AGI.
Better algorithms. Over the past decade, a Google subsidiary and leader in the AI space, DeepMind, made a splash by combining the strengths of deep learning with reinforcement learning—a complimentary machine learning approach that aims to teach AI how to take actions in new environments to achieve a set goal.
Thanks to this hybrid tactic, DeepMind’s premiere AI, AlphaGo, not only taught itself how to play AlphaGo by downloading the rules and studying the strategies of master human players, but after playing against itself millions of times was then able to beat the best AlphaGo players using moves and strategies never before seen in the game.
Likewise, DeepMind's Atari software experiment involved giving an AI a camera to see a typical game screen, programming it with the ability to input game orders (like joystick buttons), and giving it the singular goal to increase its score. The result? Within days, it taught itself how to play and how to master dozens of classic arcade games.
But as exciting as these early successes are, there remain some key challenges to solve.
For one, AI researchers are working on teaching AI a trick called ‘chunking’ that human and animal brains are exceptionally good at. Put simply, when you decide to go out to buy groceries, you’re able to visualize your end goal (buying an avocado) and a rough plan as to how you’d do it (leave the house, visit the grocery store, buy the avocado, return home). What you don’t do is plan every breath, every step, every possible contingency on your way there. Instead, you have a concept (chunk) in your mind of where you want to go and adapt your trip to whatever situation that comes up.
As common as it may feel to you, this ability is one of the key advantages human brains still have over AI—it's the adaptability to set a goal and pursue it without knowing every detail in advance and in spite of any obstacle or environmental change we might encounter. This skill would enable AGIs to learn more efficiently, without the need for the big data mentioned above.
Another challenge is the ability to not just read a book but understand the meaning or context behind it. Long term, the goal here is for an AI to read a newspaper article and be able to accurately answer a range of questions about what it read, kind of like writing a book report. This ability will transform an AI from simply a calculator that crunches numbers to an entity that crunches meaning.
Overall, further advancements to a self-learning algorithm that can mimic the human brain will play a key role in the eventual creation of an AGI, but alongside this work, the AI community also needs better hardware.
Better hardware. Using the current approaches explained above, an AGI will only become possible after we seriously boost the computing power available to run it.
For context, if we took the human brain's ability to think and converted it into computational terms, then the rough estimate of an average human's mental capacity is one exaflop, which is equivalent to 1,000 petaflops (‘Flop' stands for floating-point operations per second and measures the speed of computation).
In comparison, by the end of 2018, the world’s most powerful supercomputer, Japan’s AI Bridging Cloud will hum at 130 petaflops, far short of one exaflop.
As outlined in our supercomputers chapter in our Future of Computers series, both the US and China are working to build their own exaflop supercomputers by 2022, but even if they’re successful, that still might not be enough.
These supercomputers operate on several dozen megawatts of power, take up several hundred square meters of space, and cost several hundred million to build. A human brain uses just 20 watts of power, fits inside a skull roughly 50 cm in circumference, and there are seven billion of us (2018). In other words, if we want to make AGIs as commonplace as humans, we’ll need to learn how to create them way more economically.
To that end, AI researchers are beginning to consider powering future AIs with quantum computers. Described in more detail in the quantum computers chapter in our Future of Computers series, these computers work in a fundamentally different way than the computers we've been building for the last half-century. Once perfected by the 2030s, a single quantum computer will out-compute every supercomputer currently operating in 2018, globally, put together. They will also be much smaller and use far less energy than current supercomputers.
How would an artificial general intelligence be superior to a human?
Let’s assume that every challenge listed above gets figured out, that AI researchers find success in creating the first AGI. How will an AGI mind be different than our own?
To answer this kind of question, we need to classify AGI minds into three categories, those that live within a robot body (Data from Star Trek), those that have a physical form but are connected wirelessly to the internet/cloud (Agent Smith from The Matrix) and those without a physical form that live entirely in a computer or online (Samantha from Her).
To start off, AGIs inside a robotic body isolated from the web will compete on par with human minds, but with select advantages:
- Memory: Depending on the design of the AGI's robotic form, their short-term memory and memory of key information will definitely be superior to humans. But at the end of the day, there is a physical limit to how much hard drive space you can pack into robot, assuming we design them to look like humans. For this reason, AGIs' long-term memory will act very much like that of humans, actively forgetting information and memories that's deemed unnecessary for its future functioning (in order to free up 'disk space').
- Speed: The performance of neurons inside the human brain max out at roughly 200 hertz, whereas modern microprocessors run at the gigahertz level, so millions of times faster than neurons. This means compared to humans, future AGIs will process info and make decisions faster than humans. Mind you, this doesn't necessarily mean this AGI will make smarter or more correct decisions than humans, just that they can come to conclusions faster.
- Performance: Simply put, the human brain gets tired if it operates too long without rest or sleep, and when it does, its memory and its ability to learn and reason gets impaired. Meanwhile, for AGIs, assuming they get recharged (electricity) regularly, they won’t have that weakness.
- Upgradability: For a human, learning a new habit can take weeks of practice, learning a new skill can take months, and learning a new profession can take years. For an AGI, they will have the ability to learn both by experience (like humans) and by direct data upload, similar to how you regularly update your computer’s OS. These updates can apply to knowledge upgrades (new skills) or performance upgrades to the AGIs physical form.
Next, let’s look at AGIs that have a physical form, but are also connected wirelessly to the internet/cloud. The differences we can see with this level when compared to non-connected AGIs include:
- Memory: These AGIs will have all the short-term advantages that the previous AGI class has, except that they will also benefit from perfect long-term memory since they can upload those memories to the cloud to access when needed. Obviously, this memory won’t be accessible in areas of low connectivity, but that will become less of a concern during the 2020s and 2030s when more of the world comes online. Read more in chapter one of our Future of the Internet series.
- Speed: Depending on the type of obstacle this AGI faces, they can access the larger computing power of the cloud to help them solve it.
- Performance: No difference when compared to unconnected AGIs.
- Upgradability: The only difference between with this AGI as it relates to upgradability is that they can access upgrades in real time, wirelessly, instead of having to visit and plug into an upgrade depot.
- Collective: Humans became the Earth's dominant species not because we were the biggest or strongest animal, but because we learned how to communicate and collaborate in various ways to achieve collective goals, from hunting down a Woolly Mammoth to building the International Space Station. A team of AGIs would take this collaboration to the next level. Given all the cognitive advantages listed out above and then combine that with the ability to communicate wirelessly, both in person and across long distances, a future AGI team/hive mind could theoretically tackle projects far more efficiently than a team of humans.
Finally, the last type of AGI is the version without a physical form, one that operates inside a computer, and has access to the full computing power and online resources that its creators provide it with. In sci-fi shows and books, these AGIs usually take the form of expert virtual assistants/friends or the spunky operating system of a spaceship. But compared to the other two categories of AGI, this AI will differ in the following ways;
- Speed: Unlimited (or, at least to the limits of the hardware it has access to).
- Memory: Unlimited
- Performance: Increase in decision making quality thanks given its access to supercomputing centers.
- Upgradability: Absolute, in real time, and with an unlimited selection of cognitive upgrades. Of course, since this AGI category doesn’t have a physical robot form, it won’t have a need for the physical upgrades available unless those upgrades are to the supercomputers its operating in.
- Collective: Similar to the previous AGI category, this bodiless AGI will collaborate effectively with its AGI colleagues. However, given its more direct access to unlimited computing power and access to online resources, these AGIs will usually take leadership roles in an overall AGI collective.
When will humanity create the first artificial general intelligence?
There is no set date for when the AI research community believes they will invent a legitimate AGI. However, a 2013 survey of 550 of the world's top AI researchers, conducted by leading AI research thinkers Nick Bostrom and Vincent C. Müller, averaged out the range of opinions to three possible years:
- Median optimistic year (10% likelihood): 2022
- Median realistic year (50% likelihood): 2040
- Median pessimistic year (90% likelihood): 2075
How precise these forecasts are doesn’t really matter. What does matter is that the vast majority of the AI research community believes we will invent an AGI within our lifetimes and relatively early in this century.
How creating an artificial general intelligence will change humanity
We explore the impact of these new AI in detail throughout the very last chapter of this series. That said, for this chapter, we'll say that the creation of an AGI will be very similar to the societal reaction we'll experience should humans find life on Mars.
One camp won’t understand the significance and will carry on thinking that scientists are making a big deal about creating yet another more powerful computer.
Another camp, likely comprised of Luddites and religious minded individuals, will fear this AGI, thinking it's an abomination that it will try to exterminate humanity SkyNet-style. This camp will actively advocate to delete/destroy AGIs in all their forms.
On the flip side, the third camp will view this creation as a modern spiritual event. In all the ways that matter, this AGI will be a new form of life, one that thinks differently than we do and whose goals are different than our own. Once the creation of an AGI is announced, humans will no longer be sharing the Earth with just animals, but also alongside a new class of artificial beings whose intelligence is on par or superior to our own.
The fourth camp will include business interests who will investigate how they can use AGIs to address various business needs, such as filling gaps in the labor market and accelerating the development of new goods and services.
Next, we have representatives from all levels of government who will trip over themselves trying to make sense of how to regulate AGIs. This is the level where all the moralizing and philosophical debates will come to a head, specifically around whether to treat these AGIs as property or as persons.
And finally, the last camp will be the military and national security agencies. In truth, there's a good chance the public announcement of the first AGI may be delayed by months to years due to this camp alone. Why? Because the invention of an AGI, will in short order lead to the creation of an artificial superintelligence (ASI), one that will represent a massive geopolitical threat and an opportunity far surpassing the invention of the nuclear bomb.
For this reason, the next few chapters will focus entirely on the topic of ASIs and whether humanity will survive after its invention.
(Overly dramatic way to end a chapter? You betcha.)