Hacker News new | ask | show | jobs
by kragen 257 days ago
Eliezer’s short story “That Alien Message” providing a convincing argument that humans are cognitively limited, not data-limited, through the device of a fictional world where people think faster: https://www.lesswrong.com/posts/5wMcKNAwB6X4mp9og/that-alien...

> Yes. There is. The theoretical limit is that every time you see 1 additional bit, it cannot be expected to eliminate more than half of the remaining hypotheses (half the remaining probability mass, rather). And that a redundant message, cannot convey more information than the compressed version of itself. Nor can a bit convey any information about a quantity, with which it has correlation exactly zero, across the probable worlds you imagine.

> But nothing I've depicted this human civilization doing, even begins to approach the theoretical limits set by the formalism of Solomonoff induction.

This is also a commonplace in behavioral economics; the whole foundation of the field is that people in general don't think hard enough to fully exploit the information available to them, because they don't have the time or the energy.

——

Of course, that doesn't mean that great intelligence could figure out warp drives. Maybe warp drives are actually physically impossible! https://en.wikipedia.org/wiki/Warp_drive says:

> A warp drive or a drive enabling space warp is a fictional superluminal (faster than the speed of light) spacecraft propulsion system in many science fiction works, most notably Star Trek,[1] and a subject of ongoing real-life physics research. (...)

> The creation of such a bubble requires exotic matter—substances with negative energy density (a violation of the Weak Energy Condition). Casimir effect experiments have hinted at the existence of negative energy in quantum fields, but practical production at the required scale remains speculative.

——

Cancer, however, is clearly curable, and indeed often cured nowadays. It wouldn't be terribly surprising if we already had enough data to figure out how to solve it the rest of the time. We already have complete genomes for many species, AlphaFold has solved the protein-folding problem, research oncology studies routinely sequence tumors nowadays, and IHEC says they already have "comprehensive sets of reference epigenomes", so with enough computational power, or more efficient simulation algorithms, we could probably simulate an entire human body much faster than real time with enough fidelity to simulate cancer, thus enabling us to test candidate drug molecules against a particular cancer instantly.

Also, of course, once you can build reliable nanobots, you can just program them to kill a particular kind of cancer cell, then inject them.

Understanding this does not require believing that "intelligence that can solve every problem when it reaches a sufficient level, regardless of what data and resources it has to work with", which I think is a strawman you have made up. It doesn't even require believing that sufficient intelligence can solve every problem if it has sufficient data and resources to work with. It only requires understanding that being able to do the same thing regular humans do, but much faster, would be sufficient to cure cancer.

——

There does seem to be an open question about how general intelligence is. We know that there isn't much difference in intelligence between people; 90+% of the human population can learn to write a computer program, make a pit-fired pot from clay, haggle in a bazaar, paint a realistic portrait, speak Chinese, fix a broken pipe, interrogate a suspect and notice when he contradicts himself, fletch an arrow, make a convincing argument in courts, program a VCR, write poetry, solve a Rubik's cube, make a béchamel sauce, weave a cloth, sing a five-minute lullaby, sew a seam, or machine a screw thread on a lathe. (They might not be able to learn all of them, because it depends on what they spend time on.)

And, as far as we know, no other animal species can do any of those things: not chimpanzees, not dolphins, not octopodes, not African grey parrots. And most of them aren't instinctive activities even in humans—many didn't exist 1000 years ago, and some didn't exist even 100 years ago.

So humans clearly have some fairly flexible facility that these other species lack. "Intelligence" is the usual name for that facility.

But it's not perfectly general. For example, it involves some degree of ability to imagine three-dimensional space. Some of the humans can also reason about four- or five-dimensional spaces, but this is a much slower and more difficult process, far out of proportion to the underlying mathematical difficulty of the problem. And it's plausible that this is beyond the cognitive ability of large parts of the population. And maybe there are other problems that some other sort of intelligence would find easy, but which the humans don't even notice because it's incomprehensible to them.

2 comments

Regarding "Alien Message", I don't find that story particularly convincing. I think it's muddled and contrived.

The basic issue is that we have to deduce stuff about the world we live in, using resources from the world we live in. In the story, the data bandwidth is contrived to be insanely smaller than the compute bandwidth, but that's not realistic. In reality, we are surrounded by chaotic physical systems that operate on raw hardware. They are, in fact, quite fast, and probably impossible to simulate efficiently. For instance, we can obviously never build a computer that can simulate the behavior of its own circuitry, using said circuitry, faster than it operates. But I think there's a lot of physical systems that are just like that.

Being data-limited means that we get data slower than we can analyze and process it. It is certainly possible to improve our ability to analyze data, but I don't think we can assume that the best physically realizable intelligence would overcome data limitation, nor that it would be cost-effective in the first place, compared to simply gathering more data and experimenting more.

> Regarding "Alien Message", I don't find that story particularly convincing. I think it's muddled and contrived.

Well, yes. it's from Eliezer Yudkowsky. The kind of people who who generally find him persuasive, will do so. Those who don't find him convincing or even find him somewhat of a crank, like the other self-proclaimed "rationalists", will do do. "muddled" is correct, he lacks rigour in everything, but certainly brings the word count.

You're the guy who in https://news.ycombinator.com/item?id=45517647 I demonstrated was a physics crank: unskilled and unaware of it, dismissing the Alcubierre metric as "fantasy, nigh nonsensical and self-contradictory", unlike actual physicists. And, when I presented the evidence that that's not what actual physicists say about it, you responded by heaping personal abuse on me. Perhaps you posted this comment later as an additional form of ego defense, since it implicitly calls me a crank, by implying that I'm a "rationalist"?
Those are odd claims, but they don't interest me. You have not and are not demonstrating anything outside of your own fixations. Project much?
You seem to be agreeing with the story's thesis, rather than disagreeing. The story claims that we get an enormous amount of data from which we could compute much more than we do. You, too, are claiming that we get an enormous amount of data from which we could compute much more than we do. If that's true, then we aren't limited by our data, which is what I meant by "data-limited"—although you seem to mean the opposite, "we get data slower than we can analyze and process it", in which we are limited not by the data but by the processing. This tends to rebut the claim above, "If you had AGI tomorrow and asked it to cure cancer, it would just ask for more experimental data and resources."

It may very well be true that you could cure cancer even faster or more cheaply with more experimental data, but that's irrelevant to the claim that more experimental data is necessary.

It may also be the case that there's no "shortcut" to simulating a human body well enough to test drugs against a simulated tumor faster than real time—that is, that you need to have enough memory to track every simulated atom. (The success of AlphaFold suggests that this is not the case, as does the ability of humans to survive things like electric shocks, but let's be conservative.) But a human body only contains on the order of 10²⁴ atoms, so you can just build a computer with 10²⁸ words of memory, and processing power to match. It might be millions of times larger than a human body, but that's okay; there's plenty of mass out there to turn into computronium. It doesn't make it physically unrealizable.

Relatedly, you may be interested in seeing Mr. Rogers confronting the paperclip maximizer: https://www.youtube.com/watch?v=T-zJ1spML5c

It's not a strawman, it's a thought experiment: if the premise of AGI is that a superintelligence could do all these amazing things, what could it do today if it existed but only had its superintelligence? My suggestion is that even something a billion times more intelligent than a human being might not be able to cure cancer with the information it has available today. Yes it could build simulations and throw a lot of computing power at these problems, but is the bottleneck intelligence or computing power to run the algorithms and simulations? You're conflating the two, no one disagrees that one billion times more computing power could solve big problems, the disagreement is whether one billion times more intelligence has any meaningful value which was the point of isolating that variable in my thought experiment.
It's fair that I'm conflating raw computational power with strategic usage of that power. And it is at least theoretically conceivable that brute force computational power is not something that could be replaced by clever algorithms.

But if you agree that with 10²⁸ more times more computational power we could almost surely cure cancer without gathering much more data, then you agree that we have enough empirical data and just need to analyze it better. We're sort of arguing about the details of what kinds of approaches to analyzing the data better would work best.

I'll continue that argument about details a bit more here. So far, even with merely human intelligence, hard computational problems like car crash simulation, protein folding, and mixed integer-linear programming (optimization) have continued to gain even more efficiency from algorithmic improvements than from hardware improvements.

According to our current understanding of complexity theory, we should expect this to continue to be the case. An enormous class of practically important problems are known to be NP-complete, so unless P = NP, they take exponential time: solving a problem of size N requires k**N steps. Hardware advances and bigger compute budgets allow us to do more steps, while algorithmic improvements reduce k.

To be concrete, let's say k = 1.02, we have a data center full of 4096 1-teraflops GPUs, and we can afford to wait a month (2.6 megaseconds) for an answer. So we can apply about 10²² operations to the problem, which lets us solve problems up to about size N = 2600. Now suppose we get more budget and build out 1000 such data centers, so we can apply 10²⁵ ops, but without improving our algorithms. This allows us to handle N = 2900.

But suppose that instead we improve the heuristics in our algorithm to reduce k from 1.02 to 1.01. Suddenly we can handle N = 5100, twice as big.

We can easily calculate how many data centers we would need to reach the same problem size without the more intelligent algorithm. It's about 6 × 10²¹ data centers.

For NP-complete problems, unless P = NP, brute-force computing power lets you solve logarithmically larger problems, while intelligence lets you solve linearly larger problems, equivalent to an exponentially larger amount of computation.