Enbluesment of AI
Between enlightenment and endarkenment, there has to be a much more precise medium
Broth of questions
I hadn’t expected to get a philosophical nudge from something as simple as going through accounts payable for something as fun as the hidden taxes of a tax heaven.
This got me thinking. It reminded me of the best things in this world bearing a dual nature. Say, the Yin and Yang concept, to some extent. The perishable beauty of butterflies. Sunny days in northern countries. The terminal nature of human consciousness.
The idea of hidden taxes in a tax heaven was enough of a paradox to spark a deeper internal conversation spanning critical thinking, concepts of questions and answers, and, surprisingly, AI.
How?
I still don’t have a clear explanation of the mental bridge here, but bear with me for a second. There are two streams to it. The first one is the duality of life itself. Those tiny paradoxes you have every day. Those little things that inevitably raise questions for a mind squinting hard enough. Things like the question of life itself - would life make sense if it were eternal?
The second stream is some prejudgement I acknowledge that I have. I also acknowledge that I have no clue whether it’s true. So I would rather state it as a question.
Is it true that:
good writers are keen to answer questions
the best writers are damningly good at asking proper questions and refining them
This has crossed paths with a recent deep conversation I had. About the challenge of nurturing critical thinking in the age of AI. How do you make one think independently when there is always a ready, somewhat opinionated answer just a few fingertaps away?
And the final element in this broth of concepts:
I would assume that we have created the ultimate question-answering machines
Question-asking machines are still out there, unreplicated
Will they ever be?What is the exact quality of billions of neurons (and periphery) that makes us ask questions in a way that makes them impossible to answer sometimes?
How does it relate to critical thinking?
Post-Training
Question-answering machines. Why did it end up being that way? Intelligence does not always seem to be about that. Why have we converged on this exact mechanism? Why not something like the SF-hot idea of always-on intelligence?
Agents and coding harnesses complicate the picture, of course. They don’t just answer. They act, build and sometimes run for hours. But they still tend to begin with something from us - a question, a goal, a task. Question-answering is no longer the whole interface. It is still somewhere underneath it.
So, question-answering. My hand-woven theory is that it just kinda worked once, during the experimentation stage, when researchers faced the intriguing capability of an abyss holding the hidden distributions of a language.
I mean, imagine having a rather abstract bulb. You know it holds patterns you never knew existed in a language. Probably better than you do. Language in terms of both the math-like logic of reasoning and knowledge about things.
How do you sample from that bulb? How do you even approach it?
Steering it towards “here is the question, here is what the answer looks like” seems to be quite a logical way to select one of the possible multidimensional planes of sampling that makes it work.
Can it sample everything that is there? Is it the best way? What does “best” even mean here?
I am quite proud, though, that we have converged on some patterns that can one-shot quite a number of things these days. Even slightly “out of distribution”. Like reprogramming a robo hoover to give puss a ride.
The core tension I am trying to bring here is: “Is just answering questions sufficient for something to be considered intelligence?” Topologically speaking, the ultimate question-answering machine can be downsized all the way to a regular book. A book also holds answers. It is slower, static, and annoyingly dependent on the reader finding the right page. But the broad shape remains: the question comes from the human; the answer comes from the thing.
Critical thinking
This one is harder. It’s quite easy to think of a dystopian society where all of the questions are answered. So no one asks them anymore. Because why bother with yet another question if you know that GPT-Solar-Array-Powered-7777 will have an answer within just a minute or so?
What’s the point of even building new stuff if you know that ANYTHING you can potentially think of could be built with a few clicks? What’s the point of taking over a robo lawn mower for a cat ride if the concept itself, plus the understanding that it could be easily done (and that you know which prompt to write), gives enough dopamine?
I mean. I can even think of an Ayn Rand-style narrative written in 50 years or so. Where John Galt is the one who never stopped asking questions and challenging things, while the swamp majority stopped asking questions a long time ago.
There are layers to such imagery of how things could go wrong.
Is there, though, any light at the end of the tunnel that will bring us to a dramatically better picture?
How do we make question-asking an ultimate cultural value for future generations?
And more importantly - is my nasty approximation of question-asking a good enough proxy for critical thinking?
Enbluesment
Enbluesment is obviously a nonexistent word. I coined it through some melancholic writing I did earlier today.
It was the best thing I could think of in order to describe the feeling I have sometimes towards the world and life in general.
It’s something that could also be called “premature nostalgia for being alive” or “existential melancholy”.
It is a kind of burden of critical thinking. The burden of not seeing in black and white. The burden of seeing Yin and Yang blurred through the fabric of reality and perception.
It’s beautiful butterflies that flutter their vivid wings for way too short. Are they beautiful because their lives are short? Would they be as meaningful and culturally acknowledged if they lived for centuries?
Where does contrast matter? Where is it, rather, a scar highlighting the other side? Could such a scar be completely useless?
The crown assumption here would be that there is a flaw in life itself: the fact that life is quite finite. And the Yang part of it - that this might actually be the motivational power for self-expression, art, and science.
Is that the reason legacy matters? Is it the oldest way to live forever? Through art, through families and descendants, through laws of mathematics and physics? Through inventions and the lives of people whose trajectories got nudged upward through the random collisions of Brownian motion?
Why AI?
Enbluesment of AI is an even more made-up term than Enbluesment itself. However, there is some elegant meaning to it that I want to explain.
Remember the bulb I mentioned earlier?
Sampling from it at the correct angle could probably tell you that the concept of Enbluesment is definitely there, as well as hundreds of other not-yet-named things.
How do you sample them?
Given the framing of this post, I would just speculatively say - maybe we teach AI to ask questions? To seek disturbances in distributions and do what a human mind could do with them? Ask question after question and refine it until it clicks? Clicks by some internal metric, of course.
Maybe we reframe the paradigm to sample from this bulb in a different way. Some smart sampling aimed at not LOWERING the entropy of the next-token sequence - but rather cracking up entropy until we discover God. Or something similarly wonderful.
Could it be done via some fancy post-training?
Is it maybe too shallow of an idea spiced up with existential philosophy?
Probably
The thing I love about proper questions is that they don’t have answers. And their value is rather in the mental journey one shall partake in while thinking them through.
P.S. I know that it’s tax haven. Heaven is intentional. Cause there are havens. And there are Heavens.