Because our evolutionary environment doesn't contain cars, poetry, calculus, Star Craft, hamburgers, touch screen computers, or doors, and yet we are able to learn these things with (relative to a computer) very few examples.

Most of the effort of evolution was making cells work at all, and even then it's a bit weird, e.g. no plant or animal produces vitamin B12 and we all get this from some bacteria and archaea.

And evolution is kinda hard to time right: bacteria can reproduce in minutes, humans in decades, but only mutations that survive reproduction can be passed on. This makes it even starker as a difference: bacteria had order of 1e13 generations to become multicellular, while human DNA had about 40,000 generations to cope with fire, 220 generations for evolution to do anything with the invention of the wheel, and one generation to cope with the invention of Minecraft.

The analogy here would be: DNA is to our brains like a VN replicator bootstrapping a computer all the way up to a bare-metal-no-OS untrained model, and perhaps a few crude "hard coded" modules like a smiling-face-detector. It's a lot, but it's also missing a lot. If biology used the models and training processes that are state of the art in ML, it would take around a millennia to talk like a child and still fail the Sally-Anne test, and million years or so to pass a degree.

▲

fasterik 2 hours ago | parent [-]

I think you're underestimating how much knowledge about the world is encoded in human DNA, especially in the structure of the human brain at birth. It also depends how we count the "operations" used to train a human adult, even if we ignore the evolutionary history.

I'm still going to deny the premise of your argument, becasue I think we should define intelligence in terms of capabilities. If a system can discover a cure for cancer or solve P vs. NP, it doesn't matter how many FLOPs it took to train.

▲

ben_w an hour ago | parent [-]

I can literally point to how much information is encoded in our DNA, because it's four bases (so 2 bits per base pair) and ~3.1 billion base pairs. 6.2 gigabits total, or slightly less than 1 gigabyte.

A 1 gigabyte LLM isn't going to impress anyone with what it can do.

About 99% (depends who you ask) of our DNA is shared with our nearest primates. Like us, they can learn to use touch screens, but also like us they won't find touch screens in their natural environment. Dogs can be taught to drive cars (just about), but again, not natural environment.

> I'm still going to deny the premise of your argument, becasue I think we should define intelligence in terms of capabilities. If a system can discover a cure for cancer or solve P vs. NP, it doesn't matter how many FLOPs it took to train.

We can define it in either way. I think both are valid, because plenty of people mean each of these two things when discussing AI in particular. As I referenced in the other branch, these submarines sure can swim fast.

But at the same time, they have a lot of gaps. This is because some experience needs the real world: just as nine women can't make a baby in one month, a transistor running a million times faster than a synapse can't make a month-long cancer experiment happen in 2.6 seconds.

This dependency on data, and that state of the art ML is bad in specifically this way, is why Tesla's self-driving cars, despite having had around a trillion miles of real-world experience today, still come with steering wheels (even at least some of the Cybercabs, despite the big thing of this model supposedly being not needing them, though with Musk and his promises you should only count the Cybercabs when they actually ship and not just press releases).

▲

fasterik an hour ago | parent [-]

Note I used the word knowledge, not information. A random string can also contain 1 gigabyte of information.

Imagine an alien that matches your abilities across every domain, but has a 10 billion year training period, something many orders of magnitude more expensive than an LLM. I simply don't believe that alien is less intelligent than you.

We also don't expect humans to be competent in every domain. Most humans suck at most things. We will usually call someone intelligent if they excel at solving problems in one or two narrow domains.

▲

ben_w an hour ago | parent [-]

Information is an upper bound on knowledge.

> 10 billion year training period, something many orders of magnitude more expensive than an LLM.

I'm saying both definitions are valid definitions, they both point to important and different things: skill now, vs. how hard it is to get new skills. Some would describe it as "crystallised intelligence vs fluid intelligence".

I think it's important that any arguments are over the thing in dispute, not the label for that thing. Don't mistake the map for the territory.

Anyone who says "AI is stupid" by the first definition, what it can do, I think is making an error: they are already wildly super-human in at least some areas, if not generally.

Anyone who says "AI is stupid" by the second definition, how many examples they need, I agree with: there is a lot they are not currently able to learn even though it is easy for us, because the data they would need to do the learning on does not exist at the scale they need.

Also note: examples, not years. An alien intelligence whose synapses trigger 10 times faster or slower than mine (or ten million times faster or slower than mine), but who gets as much as I do out of each book or conversation, is my equal by the second definition.

	▲	fasterik 11 minutes ago \| parent [-]
		I wouldn't say that information is an upper bound on knowledge because we don't measure knowledge in bits. The number of possible sequences of N bits is 2^N and knowledge involves selecting the sequences that are useful in some way. I don't know how to quantify it, but in principle it could be much larger than N. I don't think I agree with your characterization of the second definition. Time scales matter. It's not much use to be able to solve human-scale problems if it takes millennia. And it only takes months to train an LLM to the level that it can solve cutting-edge math problems.