I'd blame the MIT press release organ for being clickbait, but the paper isn't much better. It's almost entirely flash with very little substance. This is not to say there's no math - the math just doesn't much apply to the real world. For example, the idea that deep neural networks work well because they recreate the hierarchical generative process for the data is a common misconception.
And then from this starting point you want to start speculating?
Can you explain why that's a misconception? Or at least point me to a source that explains it?
I've started working with neural networks lately and I don't know too much yet, but the idea that they recreate the generative process behind a system, at least implicitly, seems almost obvious. If I train a neural network on a simple linear function, the weights on the network will probably change to reflect the coefficients of that function. Does this not generalize?
It fits with the idea of the universe having an orderly underlying structure. The simulation hypothesis is just one way that can be true. Physics being true is another, simpler explanation.
When I started hearing about the latest wave of results from neural networks, I thought to myself that Eliezer was probably wrong to bet against them. Should MIRI rethink its approach to friendliness?
Neural networks may very well turn out to be the easiest way to create a general intelligence, but whether they're the easiest way to create a friendly general intelligence is another question altogether.
Many humans fear AI, indicating that other intelligent species might also fear it and might consider not developing AI. I think the great filter is more likely something that no one fears but most industrialized species do. For example, if (in a universe similar to ours) because of a weird quirk of physics that you don't discover until after you solved string theory, building a radio destroys your planet then radios would be the cause of the great filter.
Many civilizations may fear AI, but maybe there's a super-complicated but persuasive proof of friendliness that convinces most AI researchers, but has a well-hidden flaw. That's probably a similar thing to what you're saying about unpredictable physics though, and the universe might look the same to us in either case.
Any real solution of Fermi paradox must work in ALL instances. If we have 100 000 AIs in past light cone, it seems unplausible that all of them will fail the same trap before creating vNP.
Most of them will have stable form of intelligence like local "humans" which will able to navigate starships even after AI fails. So it will be like old school star navigation without AI. We will return to the world there strong AI is impossible and space is colonised by humanoid colonists. Nice plot, but where are way?
Another solution to FP is that most of new AIs fail to superAI predator which sends virus-like messages via some kind of space radio. The message is complex enough that only AI could find and read it.
Not necessarily all instances. Just enough instances to allow our observations to not be incredibly unlikely. I wouldn't be too surprised if out of a sample of 100 000 AIs none of them managed to produce successful vNP before crashing. In addition to the previous points the vNP would have to leave the solar system fast enough to avoid the AI's "crash radius" of destruction.
Regarding your second point, if it turns out that most organic races can't produce a stable AI, then I doubt an insane AI would be able to make a sane intelligence. Even if it had the knowledge to, its own unstable value system could cause the VNP to have a really unstable value system too.
It might be the case that the space of self-modifying unstable AI has attractor zones that cause unstable AI of different designs to converge on similar behaviors, none of which produce VNP before crashing.
Your last point is an interesting idea though.
While AI may halt by many reasons, which I listed in my map of AI failure levels (http://lesswrong.com/lw/mgf/a_map_agi_failures_modes_and_levels/), it does not explain Fermi paradox. ((
Because before crash an AI may create non-evolving von Neumann probes. They will be safe against philosophical crisises, and will continue "to eat" universe. If many AIs crashed in the universe before us, one surely created self-replicating space probes. Why we don't see them?
That's a good point. Possible solutions:
AI just don't create them in the first place. Most utility functions don't need non-evolving von Neumann probes, and instead the AI itself leads the expansion.
AI crash before creating von Neumann probes. There are lots of destructive technologies an AI could get to before being able to build such probes. An unstable AI that isn't in the attractor zone of self-correcting fooms would probably become more and more unstable with each modification, meaning that the more powerful it becomes the more likely it is to destroy itself. von Neumann probes may simply be far beyond this point.
Any von Neumann probes that could successfully colonize the universe would have to have enough intelligence to risk falling into the same trap as their parent AI.
It would only take one exception, but the second and third possibilities are probably strong enough to handle it. A successful von Neumann probe would be really advanced, while an increasingly insane AI might get ahold of destructive nanotech and nukes and all kinds of things before then.
The Fermi paradox has been discussed here a lot, and many times it has been argued that AI cannot be the great filter, because we would observe the paperclipping of the AI just as much as we would observe alien civilizations. I don't think we should totally rule this out though.
It may very well be the case that most unfriendly AI are unstable in various ways. For instance imagine an AI that has a utility function that changes whenever it looks at it. Or an AI that can't resolve the ontological crisis and so fails when it learns more about the world. Or maybe an AI that has a utility function that contradicts itself. There seem to be lots of ways that an AI can have bugs other than simply having goals that aren't aligned with our values.
Of course most of these AI would simply crash, or flop around and not do anything. A small subset of them might foom and stabilize as it does so. Most AI developers would try to move their AI from the former to the latter, and in doing so may pass through a space of AI that can foom to a significant degree without totally stabilizing. Such an AI might become very powerful, but exhibit "insane" behaviors that cause it to destroy itself and its parent civilization.
It might seem unlikely that an "insane" AI could manage to foom, but remember that we ourselves are examples of systems that can use general reasoning to gain power while still having serious flaws.
This would prevent us from observing either alien civilizations or paperclipping, and is appealing as a solution to the Fermi paradox because any advancing civilization would likely begin developing AI. Other threats that could arise after the emergence of civilization probably require the civilization to exhibit behaviors that not all civilizations would. Just because we threatened each other with nuclear annihilation doesn't mean all civilizations would, and it only takes one exception. But AI development is a natural step in the path of progress and very tricky. No matter how a civilization behaves, it could still get AI wrong.
If this does work as a solution, it would imply that friendliness is super hard. Most of the destroyed civilizations probably thought they had it figured out when they first flipped the switch.
Not necessarily - See Hilbert's Paradox.
https://en.wikipedia.org/wiki/Hilbert%27s_paradox_of_the_Grand_Hotel
Infinity is really confusing.
Doesn't the word "ALL" make your statement self-contradictory?
My statement itself isn't something I believe with certainty, but adding that qualifier to everything I say would be a pointless hassle, especially for things that I believe with a near-enough certainty that my mind feels it is certain. The part with the "ALL" is itself a part of the statement I believe with near certainty, not a qualifier of the statement I believe. Sorry I didn't make that clearer.
View more: Next
= f037147d6e6c911a85753b9abdedda8d){kind=logo}
= 27d567bc2d48d9f40e9ccc20ea370b15)
= b715d02e85f7b3b4ae65ca3d3e450868)
Subscribe to RSS Feed
Without commenting on whether this presentation matches the original metaethics sequence (with which I disagree), this summary argument seems both unsupported and unfalsifiable.
I don't think it being unfalsifiable is a problem. I think this is more of a definition than a derivation. Morality is a fuzzy concept that we have intuitions about, and we like to formalize these sorts of things into definitions. This can't be disproven any more than the definition of a triangle can be disproven.
What needs to be done instead is show the definition to be incoherent or that it doesn't match our intuition.