Our brains may have more in common with large language models than we may want to believe.
"...Artificial intelligence is not human, no matter how much it approximates being human. Regardless of how well it conveys intelligence, personality, and creativity, it is not a sentient, perceiving being.
Yet it is important to appreciate that much of the work we credit to our brain isn’t really sentient or part of our perception. Most of our brain’s activity is subconscious, including what we would often consider creativity. Any artist will tell you they often feel a flash of insight that leads to the creative act.
Similarly, how many times have you been told to “sleep on the problem”? I myself am a master of this art. In college, when I faced seemingly intractable math problems, I would engage with them for a few minutes and then delegate them to my subconscious. I would tell my brain to essentially come up with the answers and tell “me” when it was done. Most of the time, I had the answers by the next morning without having to consciously struggle with them. I’m not alone in doing this. Many people find it a useful way to approach difficult problems.
I now do the same thing when I face a tough problem while leading Khan Academy. I have faith that my brain, or someone else’s, will come up with a creative solution by morning. What are our brains doing subconsciously while our consciousness waits for an answer? Clearly, when you “sleep on a problem,” some part of your brain continues to work even though “you” aren’t aware of it. Neurons activate, which then activate other neurons depending on the strength of the synapses between them. This happens trillions of times overnight, a process mechanically analogous to what happens in a large language model. When a plausible solution presents itself, the subconscious then surfaces it to the conscious as a flash of insight.
Meditation gives us direct experience with this. Close your eyes for a few minutes and observe your own thoughts. They really begin to feel very much like the output of a large language model—or several competing models—whose latest output gets fed as input for the next iteration of output. With a bit of practice, your conscious mind can temporarily disassociate itself from these thoughts until you experience stillness or “no thought.” You’ll begin to see your thoughts for what they are and aren’t. They aren’t you.
Think about a flow state that most experts in their craft can attain after the often-noted ten thousand hours of practice (which is analogous to pretraining for generative AI models). They will often say that their greatest creativity or actions occur when they do not allow themselves to be conscious of what they are doing. The best way to ruin their performance, or creativity, is to consciously think about what is happening. Great orators will tell you that it feels like their brain is doing the talking while their conscious selves are just there to observe the output. After making thousands of videos, I often feel this way when I press record. I won’t claim that what experts’ well-trained brains are doing when they create is identical to what large language models do, but it seems awfully similar."
"Brave New Words: How AI Will Revolutionize Education (and Why That's a Good Thing)" by Salman Khan: https://a.co/i8LPsmB
No comments:
Post a Comment