AI is making the same progression life does
In his brilliant book Life 3.0 Max Tegmark divides life into three levels. Life 1.0 is life that has both its hardware (its body) and its software (how it operates) fully specified in its DNA. Importantly, this means it can’t learn things throughout its life. An example of Life 1.0 are bacteria: every skill a bacterium has (such as moving towards higher concentrations of sugar), it has had from its conception. It can’t learn to, say, move away from sugar. Humans, and with us a lot of other animals, can learn new skills. In fact, when humans are born they can barely do anything besides breathing and sucking milk. This is why humans are an example of Life 2.0: life that can change its software, but not its hardware. Humans can learn all kinds of new skills and knowledge, but can’t change their own fundamental nature: we can’t remake ourselves to be able to age to a million years, or to be able to remember all of Wikipedia, to borrow Tegmark’s examples. Yet. Companies like Neuralink (which is developing brain-computer interfaces) are taking the first steps into the final level: Life 3.0, which, as the reader probably expects, is capable of both changing its software and its hardware. Reaching this level means becoming free of the restraints evolution has put on us.
Interestingly, it suddenly hit me today that Artificial Intelligence (AI) is dividable into similar levels as biological life. Let me explain.
Artificial Intelligence 1.0
Where today’s hype around AI is almost entirely about Machine Learning (and more specifically Deep Learning), earlier AI systems include the chess-playing Deep Blue. The first computer to ever beat a World champion under regular time constraints, I classify Deep Blue as AI 1.0, by analogy with Life 1.0. Why? Because as impressive as the system was, everything Deep Blue knew and would ever know it already knew from its start. The rules of chess, knowledge of chess openings, its search technique to find good moves — all of them were programmed into it by its developers. It couldn’t learn anything new, let alone change its hardware.
Everything Deep Blue knew and would ever know it already knew from its start.
Artificial Intelligence 2.0
AlphaZero is, to me at least, the most impressive AI ever built. Not only is it capable of playing multiple board games (chess, go and shogi) instead of “just” one, it can reach a superhuman level of play in each of these games purely by teaching itself. That’s right: AlphaZero wasn’t given any playing techniques or even records of earlier games played by expert humans to learn from. It was given the rules of the games, and by playing a huge number of games by itself, it reached extreme levels of expertise. This is what distinguishes AlphaZero from Deep Blue: its ability to learn how to play. AlphaZero is therefore AI 2.0: capable of changing its “software” (its skills), but not its hardware.
AlphaZero can reach a superhuman level of play in multiple games by teaching itself.
Artificial Intelligence 3.0
AI 3.0 would be AI that, like Life 3.0, can change both its “hardware” and its “software”. As you might have noticed, “software” here means skills and knowledge, like it does with biological life, not actual source code of the AI. Actually, in this context, the source code would be part of the “hardware”, the “body” of the AI: an AI capable of rewriting its own source code for its own purposes would be, for a large part, master of its own destiny. Of course, being able to change the actual computer the AI is running on (for example, adding extra processors to give itself more computing power) would give the AI even more freedom, and I’d say an AI capable of both is AI 3.0.
AI 3.0 might give rise to an intelligence explosion.
The World has yet to witness it’s first installment of AI 3.0, but when it does, the World will never be the same again. AI 3.0 might give rise to an intelligence explosion (see Seed AI: History, Philosophy and State of the Art), where the AI rewrites its source code and eventually acquires more and more computing resources to become more and more intelligent. The influence AI 3.0 will have on the World will be astonishing, either in a good or in a bad way.