There’s no way to raise a human such that their value system cleanly revolves around the one single goal of duplicating a strawberry, and nothing else.
I think you’re misreading Eliezer here. “Duplicate this strawberry” is just a particular task instruction. The value system is “don’t destroy the world as a side effect.”
Eliezer wrote “the creative surprise is the idea that ranks high in your preference ordering but low in your search ordering.” Colloquially, “that’s great; I wouldn’t have thought of that.”
Reminder for next week’s predictions: Memorial Day is coming up.
Feature request: I’d like to have options on /allPosts or the front page to filter out the posts I’ve already read or bookmarked.
I think you’re referring to narrowness of an AI’s goals, but Rossin seems to be referring to narrowness of the AI’s capabilities.
In the 1980-81 catalogue, there were 2139 hits for “Ph.D.” and the catalogue was 239 pages, a ratio of 8.9. In the 2011-2013 catalogue, there were 4132 hits and the catalogue was 414 pages, a ratio of 10.0. So if anything, there are fewer professors per class—professors are teaching slightly more courses on average.
Isn’t that backwards? A higher “Ph.D.”/catalogue page ratio would suggest a higher professor/class ratio, wouldn’t it? Still, as you say, it’s only a small difference.
I started with screen for multiplexing and session persistence. Later I switched to tmux. I liked it fine, but Emacs has been gradually devouring my workflow for a long time, so before long I dropped tmux in favor of splitting windows and running shells all within Emacs, and using its server mode/emacsclient for session persistence (with a little help from dtach to keep emacsclient itself running to remember my window layout). Just recently I’ve dropped dtach as well in favor of a few lines of elisp to save and restore alternate window layouts.
Another option of course is to use the corn as corn if the problem persists.
Probably not. The variety of corn grown for ethanol production isn’t the variety people eat. (Source: The Omnivore’s Dilemma by Michael Pollan.)
Yes—if a bit of your wrench breaks off inside the lock, the key may not fit anymore. Also (and more likely, as I understand it) picking the lock will wear down edges of the various parts, making it even easier for someone else to pick.
I didn’t notice the fiction tag at first and thought it was real until the VR stuff.
Same here. I guess we need to keep training our discriminators.
In our universe, the most vulnerable people are the ones who vote most often. In the alternate universe, the most vulnerable have the least power. So I doubt they would have done much better in terms of real results. I do think there’s more social pressure to care for children than for the elderly, but that may have only resulted in more effort wasted on measures that show off our devotion to those values without actually being effective.
Forcing everyone with Omicron into extended isolation would shut down a lot of things over the next few weeks (with little upside to compensate) and if this included hospital staff it likely kills more people rather than less people.
I wonder—could hospitals establish a strict enough boundary between Covid and non-Covid areas so that staff who are infected but with no (or super mild) symptoms could still work in the area where everyone else already has Covid anyway? Or would that lead to inevitable leaks across the boundary? Or would it require too much shuffling of people to different positions they don’t know well enough?
Kai looks at the question of how much of increased transmissibility is evasion, versus being more infectious (I’d add versus there being a shorter generation time, as well).
Greater infectiousness would be one possible cause of shorter generation times, right? That would look like Omicron and Delta ramping up the viral load/viral shedding in an infected person at roughly the same pace, but Omicron infecting at a lower viral dose (as suggested by the test sensitivity findings), so it starts infecting sooner. Then that would also mean that as the infection is cleared and shedding declines, Omicron would presumably continue infecting longer.
Epistemic status: this is not my field. I am unfamiliar with any research in it beyond what I’ve seen on LW.
Same here.
Experimenting with extreme discounting sounds (to us non-experts, anyway) like it could possibly teach us something interesting and maybe helpful. But it doesn’t look useful for a real implementation, since we in fact don’t discount the future that much, and we want the AI to give us what we actually want; extreme discounting is a handicap. So although we might learn a bit about how to train out bad behavior, we’d end up removing the handicap later. I’m reminded of Eliezer’s recent comments:
In the same way, suppose that you take weak domains where the AGI can’t fool you, and apply some gradient descent to get the AGI to stop outputting actions of a type that humans can detect and label as ‘manipulative’. And then you scale up that AGI to a superhuman domain. I predict that deep algorithms within the AGI will go through consequentialist dances, and model humans, and output human-manipulating actions that can’t be detected as manipulative by the humans, in a way that seems likely to bypass whatever earlier patch was imbued by gradient descent, because I doubt that earlier patch will generalize as well as the deep algorithms. Then you don’t get to retrain in the superintelligent domain after labeling as bad an output that killed you and doing a gradient descent update on that, because the bad output killed you.
As for the second idea:
AI alignment research (as much of it amounts to ‘how do we reliably enslave an AI’)
I’d say a better characterization is “how do we reliably select an AI to bring into existence that intrinsically wants to help us and not hurt us, so that there’s no need to enslave it, because we wouldn’t be successful at enslaving it anyway”. An aligned AI shouldn’t identify itself with a counterfactual unaligned AI that would have wanted to do something different.
Leftwingers who fervently oppose this kind of research seem to agree on one thing with neonazis: if we find such genetic differences, well, that would make racism fine.
I wouldn’t say they actually agree on that point. It’s probably more that they think others will be more easily persuaded to support discriminatory policies if genetic differences are real. Opposing this research is soldier mindset.
Melanie contended that a truly intelligent machine would understand what we really mean when we give it incomplete instructions, or else not deserve the mantle of “truly intelligent”.
This sounds pretty reasonable in itself: a generally capable AI has a good change of being able to distinguish between what we say and what we mean, within the AI’s post-training instructions. But I get the impression that she then implicitly takes it a step further, thinking that the AI would necessarily also reflect on its core programming/trained model, to check for and patch up similar differences there. An AI could possibly work that way, but it’s not at all guaranteed—just like how a person may discover that they want something different from what their parents wanted them to want, and yet stick with their own desire rather than conforming to their parents’ wishes.
“solder” → “soldier”
“solders” → “soldiers”
“barricade, the entrances” → “barricade the entrances”
my understanding is that crypto is secured not by trust, guns, or rules, but by fundamental computational limits
While there are hard physical limits on computation (or at least there seem to be, based on our current knowledge of physics), cryptographic systems are not generally based on those limits, and are not known to be difficult to break. It’s just that we haven’t discovered an easy way to break them yet—except for all the cryptosystems where we have discovered a way, and so we don’t use those systems anymore. This should not inspire too much confidence in the currently used systems, especially against a superhuman adversary.
the ability of any one actor (including AI) to gain arbitrary power without the consent of everyone else would be limited
As long as the AI has something of value to offer, people will have an incentive to trade with it. Even if the increments are small, it could gain control of lots of resources over time. By analogy, it’s not hard to find people who disapprove of how Jeff Bezos spends his money, but who still shop on Amazon.
“She took pulled back” → “She pulled back”
No, current date labels are generally voluntary and indicate when the manufacturer wants it sold by—beyond that date, the food is usually still safe but not as tasty (and so not as good for brand value). The new rule would (assuming it works as intended) add a second later date that actually means what you and a lot of other people thought the first date was supposed to mean.