Gradual Disempowerment: How Even Incremental AI Progress Poses Existential Risks
arxiv.org
arxiv.org
(small spoiler)
In Money Monster it turns out the hedge fund manager who is blaming their state of the art AI bot for malfunctioning and rapidly selling off certain stock, tanking it because of that, did so out of a machine code error. He can’t explain what the error was or how or why it happened cause ‘he didn’t program their trading bot, some guy in Asia did.’ But as it turns out he did do it in some way.
I feel like using AI as a way to abstract blame even more when something goes wrong will be a big thing, even when secretly it was not the AI (ML) or who trained the thing’s fault.
Jim Keller said it best, about every 10 orders of magnitude in available computation the paradigm of computing shifts to a higher level of mathematical building blocks. In the beginning we had pure logic. Then came addition and subtraction, then came vectors, then came matrices, and now we're at tensors, he believes the next building block is graphs.
This is citing from memory while I'm sleep-deprived, but I think the general idea holds. Approximately every 10x increase in computation, there's a paradigm shift in what is possible.
But it's just algorithms. Always has been, and still is.
It's supposed to be applied equally to all citizens but we see cases everyday where it's not. The wealthy, the enforcers, and their allies are frequently spared from "the law"
It feels like a lot of "Existential AI Risk" types are divorced from the physical aspects of maintaining software - eg. your model needs hardware to compute, you need cell towers and fiber optic cables to transmit.
It feels like they always anthropomorphize AI as some sort of "God".
The "AI Powered States" aspect is definetly pure sci-fi. Technocratic states have been attempted, and econometrics literally the exact same mathematical models used in AI/ML (Shapely values are an Econometrics tool, Optimization Theory itself got it's start thanks to GosPlan and other attempts and modeling and forecasting economic activity, etc).
As we've seen with the Zizian cult, very smart people can fall into a fallacy trap of treating AI as some omnipotent being that needs to either be destroyed or catered to.
And that's what I'm getting it. It's basically bad science fiction being used to create some form of a pseudo-religious experience.
At least your local Mandirs, mosques, churches, synagogues, gurdawara, and other formal religions do food drives and connect a subset of your local community.
What do AI Doomers do? Nothing.
You seem very keyed up about this, so it is a good thing you are wrong. AI doomers are far more likely people to be giving to GiveWell, or actually thinking about how they are helping society. Source: Am AI doomer. Donate to GiveWell.
Do you disagree there are significant incentives to scale up the power, ubiquity, and direct physical impact of these systems?
Please answer with yes or no, and not some snarky “bad science fiction” straw man.
> I also wasn’t commenting on how a model is created…?
That comment wasn't targeted at you. I just find AI washing arguments laughable sometimes. It's very similar to the ambulance chasing and scare tactics that is a major part of Cybersecurity GTM.
Total red herring. Your argument for why AI cannot be a risk is because it depends on so much physical infrastructure. Sure, this infrastructure isn't going to come from thin air, but if there are strong incentives to put in place then it makes no difference whether it came from thin air or VC checkbooks. It is then in place, ergo your rationale for why this cannot be a risk is invalid.
A decent amount.
I started off as a NatSec adjacent staffer and have helped build Cybersecurity and Defense Tech companies, so a lot of my peers and friends are working directly on US-China, US-Russia, Israel-Iran, Saudi-Iran, Saudi-Turkey, India-Pakistan, and India-China relations.
These are all relations that could explode into cataclysmic wars (and have already sparked or exacerbated plenty of wars like the Syrian Civil War, Yemen Civil War, Libyan Civil War, Ethiopian Civil War, Russia-Ukraine War, Myanmar Civil War, Afghan War, etc). We are already going through a global trend of re-armament, with every country expanding their conventional, nuclear, and non-conventional warfighting capabilities. Just about every nuclear state has the nuclear triad or is in the process of implementing a nuclear triad. And China's nuclear rearmament race has forced India to rearm which has forced Pakistan to rearm, and is causing a bunch of regional arms races.
I think the world is more likely to end due to bog standard conflicts escalating into an actual war. Not some sort of AGI/ASI going skynet
WOPPER, not skynet. You saw War Games right?
Did Wargames age well? I tried watching a couple iconic 80s movies like Scarface but I just couldn't stand it.
"A strange game. The only winning move is not to play."
Not to mention that the race for AI technology is likely going to make geopolitics more volatile. If something happens, it doesn’t matter how we bucketed it conceptually. Reality doesn’t care about where we draw the lines.
The false dichotomies abound in many of these discussions.
The decision of whether or not to shell Chinese troops and start a war between two nuclear states came down to one brigadier [0]
Conventional War between two nuclear states is a MAJOR risk, and has already happened before in 1999 (even if Clinton's State Department didn't want to call it that).
These kinds of flashpoints are way more likely to cause suffering in the near future than some sort of AGI/ASI.
[0] - https://theprint.in/defence/nearing-breaking-point-gen-narav...
Ok. Then where does your thinking lead?
One good starting question is: what set of mitigation strategies do we pursue for a probabilistic mix of failure modes?
A pure technocratic driven state leveraging econometrics or optimization theory for the sake of it has failed multiple times. For example, the failure of the USSR's planned economy, the failure of the US's Vietnam War objectives due to a hyper-metrics driven workflow, etc.
On a separate note, I have always considered returning to grad school and making a small career of translating esoteric Econometrics models into ML models to make a brief publishing career. A Russian American friend of mine did something similar by essentially basing his CS research career on older Soviet optimization theory research that wasn't translated into English, so he could boost his publishing ability.
People are the problem here. As always. But of course AI itself need to be managed by people too so it can rose similar problem. Politics itself is the issue.
AI is not some runaway Skynet type of a thing. It's controlled by people, who will use it for good and bad.
Not to mention that AI is and will be owned by people who are already concentrating power in their own hands. They're not going to, voluntarily, relinquish that power.
Lets call this AI theomorph)
It's not like that. It is that. They're playing Pascal's Wager against an imaginary future god.
The most maddening part is that the obvious problem with that has been well identified by those circles, dubbed "Pascal's Mugging", but they're still rambling on about "extinction risk" whilst disregarding the very material ongoing issues AI causes.
They're all clowns whose opinions are to be immediately discarded.
The same effort could be expended on plenty of other problems that are unsolved.
The "Mugging" going on is that "AI safety" folks proclaim that AI might have an "extinction risk" or infinite-negative outcome.
And they proclaim that therefore, we should be devoting considerable resources (i.e. on the scale of billions) to avoiding that even if the actual chance of this scenario is minimal to astronomically small. "ChatGPT won't kill us now, but in 1000 years it might" kinda shit. For some this ends with "and therefore you need to approve my research funding application", for others (including Altman) it has mutated into "We must build AGI first because we're the only people who can do it without destroying the world".
The problem is that this is absurd. They're focussing on a niche scenario whilst ignoring horrific problems caused in the here and now.
"Skynet might happen in Y3K" is no excuse to flood the current internet with AI slop, create a sizeable economic bubble, seek to replace entire economic sectors with outsourced "Virtual" employees, and perhaps most ethically concerning of all: create horrific CSAM torment nexuses where even near-destitute gig economy workers in Kenya walk out of the job.
Yet "AI safety" folks would have you believe so.
SMBC is quite funny on the AI risks eg. https://www.smbc-comics.com/comic/signal-2
It's called absurd because it does not deserve to be humoured the effort of writing out those arguments.
> Here's some names in the field. 15/19 think the risk is significant
A list that is largely a pile of clowns and morons, many with direct financial interests in amplifying the "danger"/power of AI.
This is why the doomsday cult is not taken serious.
So I think we might be on the same side on this one.
Back in HS, I introduced a buddy of mine to ELIZA mode in emacs, and it completely broke her mind the same way LLMs did for a lot of people - like she actually conversed with ELIZA and used it to solve her anxiety during the college admissions process. Yet ELIZA used very simple heuristics to extrude human sounding text. And my friend wasn't some dummy or Luddite - she ended up going to an Ivy to study economics and medicine.
It goes to show that User Experience is all that really matters for technology.
Here's a more prosaic analogy that might be helpful. Imagine tomorrow there's a new country full of billions of extremely conscientious, skilled workers. They're willing to work for extremely low wages, and to immigrate to any country and don't even demand political representation.
Various countries start allowing them to immigrate because they are great for the economy. In fact, they're so great for economies and militaries that countries compete to integrate them as quickly and widely as possible.
At first this is great for most of the natives, especially business owners. But the floor for being employable is suddenly really high, and most people end up in a sort of soft retirement. The government, still favoring natives, introduces various make-work and affirmative action programs. But for anything important, it's clear that having a human in the loop is a net drag and they tend to cause problems.
The immigrant population grows endlessly, and while GDP is going through the roof and services are all cheaper than ever, people's savings eventually dwindle as the cost of basic resources like land gets bid up. There are always more lucrative uses for their capital by the immigrants and capital owners compared to the natives. Educating new native humans for important new skills is harder and harder as the economy becomes more sophisticated.
I don't have strong opinions about what happens from here, but the point is that this is a much worse position for the native population to be in than currently.
Does that make sense? Even if this scenario doesn't seem plausible, do you agree that I'm not talking about anything omnipotent, just more competitive?
Another way to understand AI in my view is to look at (often smaller) resource rich countries around the world (Oil, Minerals etc). Often the government is more worried about the resource rather than the people that live in the country. The government often does not bother educate them, take good care of them, give them a voice in the future of the country etc. because those citizens are not the ones that pay the bills, or are the main source of GDP output or source of political power etc.
Similarly in an AI heavy economy, unless systems are not designed right governments might start ignoring their citizens. If democracy is not robust or money has a big role in elections the majority voice of humans is likely to matter less and less going forward.
Norway is a good example of resource rich country that still looks out for its citizens. So it should be possible to be resource rich/AI rich and have a happy citizenry. I suppose balancing all moving parts would be difficult.
The way to deal with the risks of AI would be to make AI available to all -- this is my strong belief. There are more risk to AI being walled off to select nations / classes of citizens on grounds of various real / imagined risks. This could create a very privileged class of countries and people that have AI while the others citizens don't / cant. With such huge advantages, AI would wreak greater havoc on the "have-nots". (Over)regulation of AI can have worse consequences under some conditions.
For example:
> "Instead of merely(!) aligning a single, powerful AI system, we need to align one or several complex systems that are at risk of collectively drifting away from human interests. This drift can occur even while each individual AI system successfully follows the local specification of its goals"
Well yes, making systems and incentives is a hard problem. But maybe we can specify a specific instance of this, instead of "what if one day it goes rogue!"
In our society, there are already many superhuman AI systems (in the form of companies) - and somehow, they successfully contribute to our wellbeing! In fact life is amazing (even for dumb people in society, who have equal rights). And the reason is, we have categorized the ways it goes rogue (monopoly, extortion, etc) and responded adequately.
So the "extinction by industrial dehumanization" reads a lot like "extinction by cotton mills" - i mean, look on the bright side!
This objection is a reasonable one. But the point of the paper is that a lot of the ways we have of addressing these systemic problems will probably not work once we're mostly unemployed (or doing make-work). E.g. going on strike, holding a coup, or even organizing politically will become less viable. And once we're a net drag on growth, the government will have incentives to route resources away from us. Right now they have to invest in their human capital to remain competitive.
Is that any more convincing?
It's quite creepy that in his view all tools, features, access and capabilities will be accessible to an AI agent which can just do the task. This sounds fine in a narrow scope, but if it's deployed at the scale of Windows, then it suddenly becomes a lot more scary. Don't just think of the home users, but businesses and institutions will be running these systems.
The core problem is that we can't be sure what a new generation of AI models will be capable of after a few years of iteration. They might find it trivial to control our machines which can provide them unprecedented access to follow an agenda. Malware exists today to do this, but they can be spotted, isolated and analyzed. When the OS by design is welcoming these attacks there's nothing we can do probably.
But please tell me I've consumed too many sci-fi.
Now we have the tools to make it happen more often.
Each time we choose to allow an AI to "improve" what we write/create, each time we choose to allow AI to "summarize" what we read/consume, we choose to take another step along this road. Eventually, it will be a simple "optimization" to allow AI to do this on a protocol level, making all of our lives "easier" and more "efficient"
Of course, I am not sure if anyone will actually see this comment, or if this entire thread is an AI hallucination, keeping me managed and docile.
I've heard it said that corporations are in many ways the forerunners of the singularity, able to act with superhuman effectiveness and marshall resources on a world-altering scale in pursuit of goals that don't necessarily align with societal welfare. This paper paints a disturbing picture of what it might look like if those paperclip (profit) maximizing systems become fully untethered from human concerns.
I was always a little skeptical of the SkyNet model of AI risk, thinking the danger was more in giving an AI owner class unchecked power and no need to care about the wants or needs of a disempowered labor class (see Swanwick's "Radiant Doors" for an unsettling portrayal of this). But this scenario, where capitalism takes on a mind of its own and becomes autonomous and even predatory, feels even bleaker. It reminds me of Nick Land's dark visions of the posthuman technological future (which he's weirdly eager to see, for some reason).
I have a research project on Nick Land's core thesis that capitalism is AI. If you want to go ultra-deep into his theory, check it out: https://retrochronic.com/
It's fundamental to understand that capital is not just teleological - converging with AI on the event horizon of the Singularity - but teleoplexic (i.e. "capitalism takes on a mind of its own and becomes autonomous").
I know you are being facetious, but it is not obvious (to me at least) what the best approach is in this scenario where future AI capabilities and use is unclear. I don't know what the correct analogy for future AIs is because they don't exists yet. If AI have some offensive advantage then nuclear weapons might be the right analogy. The world would not be safer if everyone has their own nuclear weapon. If the harmless & helpful personal assistant analogy is correct, then we should democratize. But Biden's mandate that AI training runs of a certain size should be reported to the government (not in any way hindered) was (to me) just obviously good. Such a report would take at most the effort of one person-day (per Dario Amodei) and gives the government some overview of progress.
for example, regarding human marginalization in states, it's just rehashing basic tropes about government (tldr, technology exacerbates the creation of surveillance states)
- "If the creation and interpretation of laws becomes far more complex, it may become much harder for humans to even interact with legislation and the legal system directly"
Well duh. That's why as soon as we notice these things, we pass laws against it. AI isn't posing the "existential risk", the way we set up our systems are. There are lots of dictators, coups, surveillance states today. And yet, there are more places in which society functions decently well.
So overall, I'm more of the opinion that people will adapt and make things better for themselves. All this anthropomorphization of "the state" and "AI" obscures the basic premise, which is we created all this stuff, and we can (and have) modified the system to suit human flourishing
Well, the claim is that that's the sort of thing that will get harder once humans aren't involved in most important decisions.
> which is we created all this stuff, and we can (and have) modified the system to suit human flourishing
Why did we create North Korea? Why did we create WWI? We create horrible traps for ourselves all the time by accident. But so far there has been a floor on how bad things can get (sort of, most of the time), because eventually you run out of people to maintain the horrible system. But soon horrible systems will be more self-sustaining.
BUT:
1. we evolve quickly in the grand scheme of the universe but slow when talking about human lives.
2. what would rapidly adapting to AI mean? if it’s become reliant on it for the most basic tasks in the way that people nowadays have become so reliant on calculators they barely even bother with doing math in their head, sure. If you mean adapting alongside AI where we in some way would also become way smarter as humans? Nonsense.
AI can be a great tool, nothing more though.
They could be wrong (I’m ambivalent), but assuming bad intent is almost always wrong.
Just like lawnmowers don't think of hedgehogs.
It seems wild that someone could unironically talk about tools “disempowering” their users. Like, I get it, C disempowers programmers by shielding them from assembly language, and Cuisinarts disempower chefs, and airplanes disempower travelers by not making them walk through each territory.
But… isn’t that a pretty tortured interpretation of tool use? Doesn’t it lead to “the Stone Age was a mistake”, and right back to Douglas Adams’ “Many were increasingly of the opinion that they'd all made a big mistake in coming down from the trees in the first place”
I get that AI can be scary, and hey, it might kill us all and that would be bad, but this particular framing is just embarrassing.
The risk being posed here is that AI may not land as an incremental improvement that still requires the user to maintain some understanding of the underlying skill.
We aren’t quite there yet with the current LLMs. You still need to have a base level of knowledge to use them effectively. But if they were just a little bit better, hallucinated just a little bit less, the value of actually knowing things goes way down.
What would the incentive be to learn the underlying skill or area if the LLM can handle things just fine on its own? Why not just let the LLM figure it out and do it? And at that point, it ceases to be a tool and starts to be something you are completely dependent on. That is a risk.