domain:alexepstein.substack.com
Dean
Flairless
1hr ago
I am reminded of Isaac Asimov's series of stories on "The Three Laws". It basically assumes away the hardest part of AI alignment "how do you enforce the rules". But then he still manages to write about a dozen stories about how it all goes horribly wrong.
I read that recently. I was struck by how Asimov smuggled a change of rules throughout the series in a way I've rarely heard noted.
The book's narrative framing devices (the exposition characters) try to justify it each time as an unanticipated consequence but predictable outcome of the established rules. However, despite the initial setup the series isn't actually formatted as 'this is the natural conclusion of previous truths taken further.' Instead, there is a roughly mid-series switch in which the robot behaviors and three laws switch from being treated as a form of consequentialist ethics (the robot cannot allow X to happen), to utilitarian ethics (the robot gets to not only let X happen, but may conduct X itself, if it rationalizes it as greater utility of X).
It is not even that the meaning of various words in the laws of robotics were reinterpreted to allow different meanings. It's that actual parts of the rules are changed without actually acknowledged that they are changing. This is how we go from the initial rules establishing a unit of concern down to the individual human level, but the end-series machines only applying the rules to humanity as a collective in order to justify harming both collective and individual humans on utilitarian grounds. We also see changes to how the robots deal with equivalent forms of harm- going from a robot self-destructing over the moral injury inflicted of being caught in a lie, to a chapter about regulatory fraud, identify theft, and punching an agent provocateur in order to subvert democracy. (The robot is the good guy for doing this, of course.)
Even setting aside some of the sillyness of the setting (no rival robot producers, no robot-on-robot conflict between rival human interests, no mandatory medical checkups for world leaders), for all that the story series tries to present it as a building series of conclusions, rather than 'it all goes horribly wrong' I found it more akin to 'well this time it means this thing.'
ArjinFerman
Tinfoil Gigachad
2hr ago
Is full self driving more dangerous per mile than having a human drive?
I'm pretty sure it is, yes. Provably beating the average human, or even reaching the same level would be a huge milestone that Elon would be shouting from every roof. That ancient rationalist prophecy about truck drivers getting replaced by AI would have already come true.
ABigGuy4U
2hr ago
And? That’s completely orthogonal to the point I’m making. Joe Sixpack from Indiana didn’t hear about Mormon polygamy and then because of that decide to log on to the Motte and write a 2000 word essay about Bay Area rationalist polycules and what that says about Scott Alexander’s moral fibre and views of society. That’s clearly a very inside-baseball take from someone who is immersed in the rationalist milieu.
TheAntipopulist
Formerly Ben___Garrison
2hr ago
3D printing certainly has its uses, but it's nowhere near as prevalent as some hypesters claimed it would be. I remember reading people projecting that 3D printers would soon be as common as smartphones, that everyone (or at least every household) would have one, and that we'd all be printing all sorts of things. Instead, it's remained mostly restricted to some bespoke industrial uses and a small slice of hobbyists.
That's not to say it couldn't have a very bright future... eventually!
cjet79
2hr ago
Very good summary, and matches many of my feelings on the topic.
Some thoughts:
- I am reminded of Isaac Asimov's series of stories on "The Three Laws". It basically assumes away the hardest part of AI alignment "how do you enforce the rules". But then he still manages to write about a dozen stories about how it all goes horribly wrong.
- I also read a bunch of Zvi's substack roundups. That man is single handedly one of the best information aggregators I know of.
- There is definitely an assumption by the AI doomerists that intelligence can make you god tier. I'm not sure I'll ever buy this argument until I'm literally being tortured to death by a god tier controlled robot. Physical world just doesn't seem that easy to grok and manipulate. I think of intelligence as leverage on the physical world. But you need counter weight to make that leverage work. Humans have a lot of existing "weight" in the form of capital and spread. A baby AI would not have as much weight, just a data center.
- Robin Hanson has a great critique of AI doomerists. Many of them said "AI would happen this way" and that turned out to not be the way, but their predictions still haven't changed much.
FtttG
2hr ago
Every normal Christian in America has heard of the Mormon church.
TIRM
2hr ago
This is the best way yet to permanently torpedo the refugee program.
I predict a Democrat will be in the White House soon. Probably next election. They'll scramble to selectively undo Trump's work. Expect the NGO cash spigot to open back up.
Lykurg
We're all living in Amerika
2hr ago
Im also not talking about the church hierarchy. Those are officially managerial positions. What I mean is that general christian virtue ends up being a "jealous god" about the use of your status to an extent that becomes effectively managerial. Youre not supposed to derive worldly rewards from it. Matthew 6 goes in that direction relatively explicitly.
Of course this has mostly not been actual practice, but its been there, and radical/restorationist people keep hitting upon it, and... I see their point.
ABigGuy4U
2hr ago
We’ve already spent the last 250 years automating as many physical labor jobs as possible out of existence. The last 5000 years if you include domesticated animal labor. So what we’re left with are the 1 percent of physical labor jobs that the absolute hardest or least cost-effective to automate. Whereas intellectual jobs are still pretty early on in the process, we only started automating those in the mid 1960s. So there’s a hell of a lot more low-hanging fruit to be picked even with fairly basic advancements in AI.
TIRM
2hr ago
tech bubbles that ultimately failed to really pan out for one reason or another, like ... 3D printing
We are in a golden era of 3D printing. Now that a few key Stratasys patents have expired they no longer have a stranglehold on 3D printing. Anyone can make a FDM 3D printer.
A few high performance airplane components are 3D printed thanks to SLS delivering full strength metal parts. This is the good outcome for 3D printing as a practical technology.
TheAntipopulist
Formerly Ben___Garrison
2hr ago
I think a plateau is inevitable, simply because there’s a limit to how efficient you can make the computers they run on. Chips can only be made so dense before the laws of physics force a halt.
What if we get an AI so smart that it figures out a way to circumvent these particular laws of physics? I'm 50% joking with this follow-up question, and 50% serious.
The military exists in short to make enemies dead. They don’t want an AI that is going to get morally superior when told to bomb someone.
The fact that there will be a big emphasis on designing AI to be able to bomb people without question is not exactly something that increases my confidence in alignment! I think the argument you're making here is more along the lines of 'following directions will always be a critical component of AI effectiveness, so the problem with largely solve itself'. I think that's somewhat plausible for simplish AI, but it gets less plausible for an AI that's 2000x smarter than people.
Pynewacket
2hr ago
interesting, thanks for sharing.
TheAntipopulist
Formerly Ben___Garrison
2hr ago
Thank you for the kind words.
FiveHourMarathon
Wawa Nationalist
2hr ago
I'm of the opinion that most countries, most of the time, are just coasting off a good run of 20-50 years for 100-200 years at a time. You get one good run, and the momentum and institutions and values created carry a civilization for decades.
Even when you have a multi-century run like the Roman Republic or the 20th century United States where the line goes up and to the right for a long time, there are a lot of crises that we barely remember, and a lot of mediocre leaders mixed in with the studs. You see a jump from a generation of great leaders, and then a long dwindling, until another great generation jumps up.
ThisIsSin
Cainanites and Abelists
3hr ago
I believe that overtime, they experience a sort of unconscious strain that builds up with each casual sexual encounter, that worsens their mental health, separately from the regret they feel due to the aforementioned physical and social costs.
Of course. After all, who would want to buy goods from a seller that casually exposes their ware(s) to hazardous, ionizing XXX-rays?
But really, this is just "buyers being buyers", in the exact way that "sellers being sellers" is. You instinctively negotiate- used items are inherently worth a little less than new items, it's best to have a service history and low mileage, shops that take care to polish and present their merchandise command higher prices from you just based on the confidence it demonstrates in the product, etc. etc. blah blah blah.
Or is there some other reason I'm not anticipating why you'd believe this is true in the absence of evidence? You can just say "I don't want to spend the money", lol; that's just as morally neutral as sellers who say "that is not enough money".
FiveHourMarathon
Wawa Nationalist
3hr ago
But it is? I'm extremely pro-immigration because
Exactly my point!
Admitting that people emigrating is bad for the country they are from is an argument that, once admitted by the pro-immigration crowd, is fully generalizable.
If it's bad for South Africa to get rid of discontent whites, then why is it good for us to take refugees from Syria and Afghanistan and...?
faceh
3hr ago
Do you seriously imagine that economies of scale in a nation with 5x the American workforce will amount to Wile-E-Coyote running off a cliff.
Yes. Find me a single instance in history where a nation was able smoothly transition through a period of declining population as the old begin to outnumber the young.
Especially one that is utterly dependent on continued imports of agricultural products and energy and most raw materials for that workforce to do anything productive.
https://www.cfr.org/article/china-increasingly-relies-imported-food-thats-problem
https://www.eia.gov/todayinenergy/detail.php?id=62804
They are simply unprepared to weather any situation where they can't afford to purchase necessary economic inputs from other countries.
Which is what their population cliff threatens to cause.
2rafa
3hr ago
Yes, due to intracommunal ethnic dynamics in the Bantu population, specifically between the Zulus and Xhosas, not to mention the fact that most ethnic resentment is toward black immigrants from the Congo and elsewhere, and to a lesser extent toward Indians.
If you ask the average black South African they might agree with a statement like ‘white people still have an unfair share of the nation’s wealth’, but it’s a distant fourth or fifth place behind other grievances.
faceh
3hr ago
I don't see the disagreement?
The economy will have to contract, this will lead to lower standards of living, and thus there's no way China can maintain its status as a continually growing economy?
2rafa
3hr ago
It wasn’t a genius move and he came unnecessarily close to actual ruin on a couple of occasions. You can disagree, but this one of very few areas where I personally know a couple of people involved in lending for commercial real estate in NYC and NJ at that time and they both say the same thing. (And one of them definitely voted for him, the other I would guess did).
lagrangian
3hr ago
I completely agree. This is exactly what I tried to say a couple weeks ago, but better written and less inflammatory. Thanks for taking the time.
MaiqTheTrue
Renrijra Krin
3hr ago
I think a plateau is inevitable, simply because there’s a limit to how efficient you can make the computers they run on. Chips can only be made so dense before the laws of physics force a halt. This means that beyond a certain point, more intelligence means a bigger computer. Then you have the energy required to run the computers that house the AI.
A typical human has a 2lb brain and it uses about 1/4 of TDEE for the whole human, which can be estimated at 500 kcal or 2092 kilojoules or about 0.6 KWh. If we’re scaling linearly, if you have a billion human intelligences the energy requirement is about 600 million KWh. An industrial city of a million people per Quora uses 11.45 billion KWH a year. So if you have something like this you’re going to need a significant investment in building the data center, powering it, cooling it, etc. this isn’t easy, probably doable if you’re convinced it’s a sure thing and the answers are worth it.
As to the second question, im not sure that all problems can be solved, there are some things in mathematics that are considered extremely difficult if not impossible. And a lot of social problems are a matter of balancing priorities more the than really a question of intellectual ability.
As to the third question, I think it’s highly unlikely that the most likely people to successfully build a human or above level AI are people who would be least concerned with alignment. The military exists in short to make enemies dead. They don’t want an AI that is going to get morally superior when told to bomb someone. I’m suspecting the same is true of business in some cases. Health insurance companies are already using AI to evaluate claims. They don’t want one that will approve expensive treatments. And so there’s a hidden second question of whether early adopters have the same ideas about alignment that we assume they do. They probably don’t.
TheAntipopulist
Formerly Ben___Garrison
3hr ago
Definitely does, but some hype bubbles do pan out (smartphone, social media, internet/ecommerce [with a bust along the way], arguably solar power).
Sure, these are all true to some extent. Like, social media is obviously very important, but I remember some people claiming it would end all wars since people would empathize with one another too much. The most extreme claims never come true.
Also, the 3 tech examples you posted all mostly occurred during the 2000-2010 decade, whereas a lot of the flops (crypto, blockchain, NFTs, VR, etc. ) are considerably more recent. Maybe there's a recency bias or availability heuristic going on that makes people excessively discount tech-based hype now.
TheAntipopulist
Formerly Ben___Garrison
3hr ago
Good post.
Thank you!
On your question 4, while that will certainly be an interesting topic and one that many people want to discuss, it's fairly pedestrian. "How should we share the benefits of scientific advancements" is something humanity has been dealing with for centuries. It's utterly dependent on how the other 3 questions resolve. If (1) is false and we don't get further major AI advances, then nothing really needs to change from the status quo. If (1) is true but (2) is false and AI revolutionizes some areas but not others, then maybe we have jobs programs so people in affected industries can reskill. If (1), (2), and (3) are true, then something like UBI can be seriously considered and we can all live in Fully Automated Luxury Gay Space Communism. If (1) and (2) are true but (3) is false then we're all dead anyways so who cares.
This is kind of the crux of the AI 2027 project Scott shared recently. Not coincidentally, it also claims to have good answers to (1), though I didn’t really dive into their reasoning. I’m curious about your thoughts on Kokotajlo’s scenario.
I wasn't particularly convinced by any evidence they posted in regards to question 1. It was mostly handwaving at recursive self-improvement, drawing straight lines on graphs and zooming off to infinity, and stuff like that. AI 2027 was one of the primary reasons I wrote this piece, as it's probably the best-researched pieces I've seen on the topic, and there's still just almost no substance. Nothing anyone could really use to make confident claims one way or the other.
Currently reading SM Stirling's To Turn the Tide. Which is exactly the sort of ISOT story I signed up for. Not the deepest characters but still enjoyable enough. Except...
I just wish Stirling didn't crib from his own - much better - genre namer. You read enough of a small circle of althist writers like him and Eric Flint and you start to see the same tropes.
More options
Context Copy link