A few days ago I listened to an interview between Dwarkesh Patel and Ilya Sutskever, OpenAI Chief Scientist, about ChatGPT and Artificial Intelligence. The following snippet got me thinking:
DP: Robotics. Was it the right step for OpenAI to leave robotics behind?
IS: Yeah, it was. Back then it really wasn’t possible to continue working in robotics because there was so little data. There was no path to data on robotics. You really need to build many thousands, tens of thousands, hundreds of thousands of robots, and somehow collect data from them and find a gradual path where the robots are doing something slightly more useful.
You could imagine it’s this gradual path of improvement, where you build more robots, they do more things, you collect more data, and so on. You need to be really willing to solve all the physical and logistical problems of dealing with them. It’s not the same as software at all. Link: youtu.be/Yf1o0TQzry8?t=771
I was thinking, “Right. If you screw some pieces of metal together, add some DC motors and an electronic brain, there’s not much to learn of it. Where should the data come from?”
On the other hand, take any organic animal, for example a human. Take our unique, first-person perspective. Our bodies have been going through millions of years of evolution; and from a phylogenetic perspective through billions of individuals lives and trillions and trillions of experiences and stories. We have been diversifying, branching and pruning. Just like all organic beings on this planet we have been pretty busy. The results of these millions of years of development are right there, at our fingertips, literally. I could touch my pointer finger to my nose, and go, “Oh, that’s my nose. It’s quite pointed I have to say.” I could look at a cat and tell her, “You’re so fine, but your nose is much smaller than mine.” And the cat would be quite bored of my tittle-tattle and turn her cute little head away.
Our bodies, which resulted from millions of years of data processing, are right there, at our disposal for learning. When we move a hand, an arm, the head, when we do something useful, or even when we’re just fooling around, it can be a rich experience. But whatever we make of it, it takes a well functioning brain to perform any movement well, and to perceive it well as well. We can spend thousands of hours to learn ourselves and still, who can’t improve further in playing the piano, or in singing, or in cooking? Or even in chewing without biting one’s cheek once in a while. There’s no limit to learning and improvement and failure, it’s our nature.
And then society! Oh so much to learn. How we interact with others, how we learn to become part of a community with our work, but also in thinking, believing, in speaking the same language with the same accent. There’s so much to do, to experience in play, in work, in sink or swim, in life or death; there’s so much data generated.
And lastly, our self-directed learning. It’s our dignity, our starting point of new adventures, our lever, saviour and maybe even our downfall.
I just listed three aspects of learning, or data processing, or token generation, if we would call it that. Which reminds me of the very beginning of Moshé Feldenkrais’s 1972 book Awareness Through Movement. It goes like this:
We act in accordance with our self-image. This self-image—which, in turn, governs our every act—is conditioned in varying degree by three factors: heritage, education, and self-education.
In this first paragraph Moshé Feldenkrais also mentions three things that define and shape who we are and how we do things: heritage, education and self-education. For ChatGPT those three things probably would be the model, the pre-training, and fine-tuning.
It certainly could be a worthwhile exercise to re-read Moshé Feldenkrais’s book from this angle, a fresh look at how and why we learn, and what makes us feel that we’re doing something slightly more useful.