Understanding Intelligence
Intelligence as a Weapon
Since we are, or at least seem to be, intelligences, it is difficult for most people to really grasp the power of intelligence. Zooming out, you can imagine various pieces on a board. Those with more limited options to move have more limited intelligence, such as pawns. While those with a higher range of possible moves have deeper, more general forms of intelligence. The number of directions they can move can be thought of as the degree of general intelligence they possess. While how far they can reach in a single move can be thought of how domain deep, or how powerful, their intelligence is. Queens can move in multiple directions, and can move much farther than pawns.
Now think of this board on a multi-dimensional matrix. There are many different layers on the board, with many different option spaces on each. The total number of option spaces is too vast to comprehend. But what could be seen is that while other species had mostly limited directional power and range in their moves at each turn, one species started showing different characteristics. Over a few iterations of the game, their range got further, and their directions on multiple spaces got broader. Visualizing the possible option spaces, any species would be far, far, far away from reaching something like nuclear weapons. But in a few turns, the human pieces executed a set of moves, using their increased direction and range, to reach a part of the board no other species could come close to reaching.
We don’t know the full range of possibilities in that option space, the board is too vast. We don’t know what else is possible on that board. What we can see is that pieces with greater ability to move on the board can execute a set of moves over a short number of steps to reach parts of the board previously unimagined.
Artificial Narrow Intelligence can be thought of as a piece with a much higher range in how far it can move at each turn, but with very little in terms of directional options. Deep, but not broad. AlphaGo is such a network. However, Alpha Go Zero could play Go without observing any human. And Mu Zero can play Chess, Go, and Atari systems, better than humans, and without being taught. The piece that had little directional flexibility is starting to have more moves in the option space. It’s getting broader, and deeper.
GPT models are very versatile when it comes to language, and at the beginning, did not possess a very deep understanding of them. They were broad, but not deep. They had many directional possibilities, but did not have much range to move at each turn. However, we are seeing with GPT3, and now GPT4, that it is gaining broader directional power, and greater range with each move. It is getting deeper, and more broad.
Fragility of the board
If the board was severely sturdy, and it seemed difficult for any group of pieces to drastically alter the game, this might not be cause for great alarm. But what we saw, with quick moves such as the ones humans made to get to the Nuclear space on the board, is that it is actually very easy for a set of pieces with high directional power and range to execute a series of moves that can quickly end the game for most of the other pieces. It shockingly only took a few steps to get from the Steam Power space to the Fission Power space. And now most moves by most pieces on the game are irrelevant, because their entire continuation on the board depends solely on where the human pieces move next. The board itself is fragile.
Zooming out, with the advent of Artificial General Intelligence, you might see something strange happening on the board. New pieces emerging, somewhat in the domain space of the human pieces, but increasingly moving outside their own possible moves at any given turn. They sometimes seem to only follow the human pieces, but not always, sometimes seeming to move on their own. What can be seen is that if the trajectory continues, these new pieces will quickly gain greater directional power in the game than the humans, and greater range to move on the board. And with each move, these new pieces are finding new option spaces on the board that previously only humans, or no player, could find.
