Why does science not improve as fast as it used to

Pointing fingers to anything other than those who do science isn’t gonna fix anything, but maybe it’s the only accurate option we have.

Maybe it’s about the capabilities of humans in general.

Hear me out. Humans have five senses: hearing, touch, vision, smell, taste.

Hearing and touching are all mechanical. What’s mechanics but atoms and molecules in motion? Human scale of mechanical perception is frequencies ranging between 20Hz-20kHz and distance ranging one order of magnitude around centimeter in both directions.

Vision is optical. Human scale regarding optical phenomena is 380-760nm.

Smell and taste is chemical.

So essentially the topics of AMO physics and electromagnetics cover our entire interface modalities with the universe.

I mean Newton did some great stuff, so did Einstein and many other famous scientists.

But.. They already had an attack (https://www.cs.virginia.edu/~robins/YouAndYourResearch.html) on the problem they were dealing with, simply because they were born in the right era. That is, they were in the right place (mentally) in the right time.

Right now for most big problems we don’t have an attack. All our modalities are saturated with tons of related phenomena observed. And our models all do a good job explaining those phenomena, even though they don’t give us much of a clue about how they fundamentally work.

An example: why can’t we invent teleportation (portals) or time travel? The answer is that we don’t even have an attack on those problems. We don’t understand the system we’re in (the universe) enough to even begin working on such problems.

And that’s what’s happening with science, too.

All of science is similar to exploring a video game. Now imagine you’re in a video game. When you first begin, you can explore the video game simply by walking around the map.

Then you jump against the walls, trying to find out how the collision detection works. Then you do that too, and you need something else to explore the game.

Sadly for you, you are only able to walk around and jump in this game. What now? Surely there are internal design details of the game. But you have no access to them. You are merely an agent inside the game whose action space consists of walking and jumping.

This is why, even when the perfect, idealized version of AGI comes along, there won’t be an “intelligence explosion”. Because the expression of intelligence is bottlenecked by other factors. Like the action space of the AGI.

This is similar to the game analogy. Imagine if you had another agent in that same game, this time superintelligent. Well, the rules of the game (laws of physics) dictate that the superintelligent agent, too, can only walk around and jump. What could this superintelligent agent possibly do better than you in exploring the game? Nothing.

Well if the game had other objectives that rely on walking and jumping, the superintelligent agent could and would dominate you probably. Because now you give it the problems where it can express it’s intelligence. But when it comes to exploring the game? Nope.

So we really need other modalities to probe the universe to explore things further. The most prominent example to this is the LIGO centers that literally probe the streching and compressing of space-time.

Another thing to consider. There might be (and there usually are) phenomena that we think we understood pretty well, but there are actually a whole other connections that we could’ve made through models we could’ve used for deeper understanding of the game (the universe). These kind of things both help us inside the game to accomplish objectives, and also guide us to probe the game in novel ways which could possibly give us new phenomena.

The current bleeding-edge for science is nanotechnology, AMO physics, protein dynamics without a doubt. For these, we have to understand the complex systems, with heavy reliance on graph theory and geometry.

If AGI arrives soon, it’s gonna have to make huge leaps in these domains first, in order to understand the universe deeper. If not, we humans, will have to make huge leaps in these domains. Then using these leaps we will invent new tools, which then could guide us into probing the universe further for getting new phenomena.

But I don’t think that will happen quickly. So I think it’s safe to say that we’re currently in the era of engineering, rather than science.

Working on science rather than engineering today requires doing things that are similar conceptually to 17th century alchemy. It usually doesn’t work. It can, but nooooot really. (I mean if alchemists back then invented sophisticated nuclear reactors that doesn’t even exist today, they could’ve succeeded in converting cheap metals into gold. And in theory they could have invented them. But… you know they couldn’t.)

I also fully believe that for AGI to arrive, we need fundamentally new ways of looking at graph theory and systems that rely on it (most prominently neural networks, but possibly also physics). AGI isn’t gonna come by doing blind trial-error type of work, which dominates the current paradigm for AI research.

I find Wolfram Physics project so on point about all of these.

And recognize that my ideas about intelligence expression and bottleneck are in line with François Chollet’s. (He’s a great source of inspiration for me.)

François Chollet: Keras, Deep Learning, and the Progress of AI | Lex Fridman Podcast #38