Макс Тегмарк - Life 3.0 - Being Human in the Age of Artificial Intelligence

THE BOTTOM LINE:

• Intelligence, defined as ability to accomplish complex goals, can’t be measured by a single IQ, only by an ability spectrum across all goals.

• Today’s artificial intelligence tends to be narrow, with each system able to accomplish only very specific goals, while human intelligence is remarkably broad .

• Memory, computation, learning and intelligence have an abstract, intangible and ethereal feel to them because they’re substrate-independent: able to take on a life of their own that doesn’t depend on or reflect the details of their underlying material substrate.

• Any chunk of matter can be the substrate for memory as long as it has many different stable states.

• Any matter can be computronium, the substrate for computation, as long as it contains certain universal building blocks that can be combined to implement any function. NAND gates and neurons are two important examples of such universal “computational atoms.”

• A neural network is a powerful substrate for learning because, simply by obeying the laws of physics, it can rearrange itself to get better and better at implementing desired computations.

• Because of the striking simplicity of the laws of physics, we humans only care about a tiny fraction of all imaginable computational problems, and neural networks tend to be remarkably good at solving precisely this tiny fraction.

• Once technology gets twice as powerful, it can often be used to design and build technology that’s twice as powerful in turn, triggering repeated capability doubling in the spirit of Moore’s law. The cost of information technology has now halved roughly every two years for about a century, enabling the information age.

• If AI progress continues, then long before AI reaches human level for all skills, it will give us fascinating opportunities and challenges involving issues such as bugs, laws, weapons and jobs—which we’ll explore in the next chapter.

*1 To see this, imagine how you’d react if someone claimed that the ability to accomplish Olympic-level athletic feats could be quantified by a single number called the “athletic quotient,” or AQ for short, so that the Olympian with the highest AQ would win the gold medals in all the sports.

*2 Some people prefer “human-level AI” or “strong AI” as synonyms for AGI, but both are problematic. Even a pocket calculator is a human-level AI in the narrow sense. The antonym of “strong AI” is “weak AI,” but it feels odd to call narrow AI systems such as Deep Blue, Watson, and AlphaGo “weak.”

*3 NAND is short for NOT AND: An AND gate outputs 1 only if the first input is 1 and the second input is 1, so NAND outputs the exact opposite.

*4 I’m using “well-defined function” to mean what mathematicians and computer scientists call a “computable function,” i.e., a function that could be computed by some hypothetical computer with unlimited memory and time. Alan Turing and Alonzo Church famously proved that there are also functions that can be described but aren’t computable.

*5 In case you like math, two popular choices of this activation function are the so-called sigmoid function σ( x ) ≡ 1/(1 + e − x ) and the ramp function σ( x ) = max{0, x }, although it’s been proven that almost any function will suffice as long as it’s not linear (a straight line). Hopfield’s famous model uses σ( x ) = −1 if x < 0 and σ( x ) = 1 if x ≥ 0. If the neuron states are stored in a vector, then the network is updated by simply multiplying that vector by a matrix storing the synaptic couplings and then applying the function σ to all elements.

Chapter 3 The Near Future: Breakthroughs, Bugs, Laws, Weapons and Jobs

If we don’t change direction soon, we’ll end up where we’re going.

Irwin Corey

What does it mean to be human in the present day and age? For example, what is it that we really value about ourselves, that makes us different from other life forms and machines? What do other people value about us that makes some of them willing to offer us jobs? Whatever our answers are to these questions at any one time, it’s clear that the rise of technology must gradually change them.

Take me, for instance. As a scientist, I take pride in setting my own goals, in using creativity and intuition to tackle a broad range of unsolved problems, and in using language to share what I discover. Fortunately for me, society is willing to pay me to do this as a job. Centuries ago, I might instead, like many others, have built my identity around being a farmer or craftsman, but the growth of technology has since reduced such professions to a tiny fraction of the workforce. This means that it’s no longer possible for everyone to build their identity around farming or crafts.

Personally, it doesn’t bother me that today’s machines outclass me at manual skills such as digging and knitting, since these are neither hobbies of mine nor my sources of income or self-worth. Indeed, any delusions I may have held about my abilities in that regard were crushed at age eight, when my school forced me to take a knitting class which I nearly flunked, and I completed my project only thanks to a compassionate helper from fifth grade taking pity on me.

But as technology keeps improving, will the rise of AI eventually eclipse also those abilities that provide my current sense of self-worth and value on the job market? Stuart Russell told me that he and many of his fellow AI researchers had recently experienced a “holy shit!” moment, when they witnessed AI doing something they weren’t expecting to see for many years. In that spirit, please let me tell you about a few of my own HS moments, and how I see them as harbingers of human abilities soon to be overtaken.

Breakthroughs

Deep Reinforcement Learning Agents

I experienced one of my major jaw drops in 2014 while watching a video of a DeepMind AI system learning to play computer games. Specifically, the AI was playing Breakout (see figure 3.1), a classic Atari game I remember fondly from my teens. The goal is to maneuver a paddle so as to repeatedly bounce a ball off a brick wall; every time you hit a brick, it disappears and your score increases.

Figure 3.1: After learning to play the Atari game Breakout from scratch, using deep reinforcement learning to maximize the score, the DeepMind AI discovered the optimal strategy: drilling a hole through the leftmost part of the brick wall and letting the ball keep bouncing around behind it, amassing points very rapidly. I’ve drawn arrows showing the past trajectories of ball and paddle.

I’d written some computer games of my own back in the day, and was well aware that it wasn’t hard to write a program that could play Breakout—but this was not what the DeepMind team had done. Instead, they’d created a blank-slate AI that knew nothing about this game—or about any other games, or even about concepts such as games, paddles, bricks or balls. All their AI knew was that a long list of numbers got fed into it at regular intervals: the current score and a long list of numbers which we (but not the AI) would recognize as specifications of how different parts of the screen were colored. The AI was simply told to maximize the score by outputting, at regular intervals, numbers which we (but not the AI) would recognize as codes for which keys to press.

Initially, the AI played terribly: it cluelessly jiggled the paddle back and forth seemingly at random and missed the ball almost every time. After a while, it seemed to be getting the idea that moving the paddle toward the ball was a good idea, even though it still missed most of the time. But it kept improving with practice, and soon got better at the game than I’d ever been, infallibly returning the ball no matter how fast it approached. And then my jaw dropped: it figured out this amazing score-maximizing strategy of always aiming for the upper-left corner to drill a hole through the wall and let the ball get stuck bouncing between the back of the wall and the barrier behind it. This felt like a really intelligent thing to do. Indeed, Demis Hassabis later told me that the programmers on that DeepMind team didn’t know this trick until they learned it from the AI they’d built. I recommend watching a video of this for yourself at the link I’ve provided.1