Strategic Wealth Accumulation Under Transformative AI Expectations
arxiv.org
arxiv.org
It's an interesting idea. But if the economy grinds to a halt because of that kind of investor behavior, it seems unlikely governments will just do nothing. E.g. what if they heavily tax ownership of AI-related assets?
Since most returns go to capital, we can expect returns on capital to increase.
Some goods and services are sold to another business, not a consumer. Right now, these "B2B" sales represent about 30% of sales volume in the economy, but there is probably no impediment to that number rising to 90%.
That's just the thing though: if we transition to a reality in which 90% or 99% of economic activity does not involve consumers, governments will adapt by taxing consumers (i.e., employees) less and investors and corporations more. It's also possible that corporations will escape the governmental oversight and control that they are currently under, which means governments of the world would shrink because they have to survive on taxes on the 1% of the economy that remains under the control of employees and consumers.
I certainly want most of the economy to serve people, but there is nothing contradictory or impossible about the possibility that 99% of it is a competition between ultra-wealthy corporations.
I think the future is less a utopia where all our needs are met and more like what happens in the third world. Where most people are desperate and impoverished save for whoever owns the underlying mineral rights and can afford to squeeze them for profit. The incentives of business all point to this outcome and there is plenty of precedent already. Zero precedent for a utopia.
What do you mean by having a large group of robots working for you? to produce products? to trade for what?
Where does an AI get chartered status, admitted to the bar, and insurance cover?
> this paper focuses specifically on the zero-sum nature of AI labor automation... When AI automates a job - whether a truck driver, lawyer, or researcher - the wages previously earned by the human worker... flow to whoever controls the AI system performing that job.
The paper examines a world people will pay an AI lawyer $500 to write a document instead of paying a human lawyer $500 to write a document. That will never happen.
I’m not sure what your expectation is, but even your claim about the assumption the paper makes is incorrect.
For one thing, the paper assumes that the amount that will be transferred from the human lawyer to the AI lawyer would be $500 + the productivity gains brought by AI, so more than 100%.
But that is irrelevant to the actual paper. You can apply whatever multiplier you want as long as the assumption that human labor will be replaced by AI labor holds true.
Because the actual nature of the future is irrelevant to the question the paper is answering.
The question the paper is answering is what impact such expectations of the future would have on today’s economy (limited to modeling the interest rate). Such a future need not arrive or even be possible as long as there is an expectation it may happen.
And future papers can model different variations on those expectations (so, for example, some may model that 20% of labor in the future will still be human, etc).
The important point as far as the paper is concerned is that the expectations of AI replacing human labor and some percentage of the wealth that was going to the human labor now accrues to the owner of the AI will lead to significant changes to current interest rates.
This is extremely useful and valuable information to model.
We seem pretty likely to be headed towards a future where AI-provided services have almost no value/pricing power, and just become super low margin businesses. Look at all of the nearly-identical 'frontier' LLMs right now, for a great example.
Do you think we will be doing in 5 or 10 years the same things we do today, but with AI? Every capability increase or cost reduction stimulates demand. AI is no different, it will stimulate both demand and competition. And since everyone has AI, and AIs are not much different between them, then the differentiating factor remain the humans. Even if we solve all our current problems with AI there is no reason to stop there, we could reduce poverty, pollution, fight global warming, conquer space. The application space is unbounded. Take electricity or internet for example and think how they expand the scope of work. Programming has been automating itself for 60 years, with each new language, library or open source project, and yet we have great jobs in the field.
No matter how much we have, we want more. Our capability of desiring progress is faster than AI capability to provide it.
The question the paper is tackling is how human behavior today (limiting this to modeling interest rates) would change if people generally expected money to transfer from labor to AI.
A future where this happens isn’t necessary for this paper to be valuable. Heck, even if people never believe this will actually happen makes this paperback less valuable.
As an analogy, let’s say someone wrote a paper on how humanity could respond to a catastrophic hit by a large meteor.
This paper would be valuable even if the earth is never hit by a meteor while humans are still here. Further, it would be additionally valuable because the insights gained here could then be applied to other, less catastrophic scenarios, such as a volcano eruption that destroys a single city.
It's an absurd assumption made by AI investors everywhere. They can't handle a world where everyone already has an AI lawyer at home that they trust, that they have because they once paid $100 for it at a kiosk in the mall or pirated it. The real future is an AI lawyer on your keychain and an extreme devaluation of the skill of knowing the law and making legal arguments.
Instead, we're going to have a weirder world where you show up to court and the court already has a list of your best legal arguments that they generated completely independent of you, and they largely match the list of arguments that your own AI advisor app gave you. They'll send you messages regarding your best next steps, and if your own device agrees, all you'll have to do is reply 'Y.'
For simple document preparation, I'm pretty sure that your phone will be able to handle it, and AI at the point of submission would be able to give you helpful suggestions if the documents were inadequate.
LLMs can almost do things of this degree of difficulty reasonably well now. Where will they be (or their successors be) in 10 years? Why do we think they will be as expensive as lawyers, who you have to send to difficult schools for a long time, feed, and flatter?
The whole AI will cause interest rates to shoot up thing seems a bit mad.
Labor automation is not zero sum. This statement alone makes me sceptical of the conclusions in the article.
With sufficiently advanced AI we might not have to do any work. That would be fantastic and extraordinarily valuable. How we allocate the value produced by the automation is a separate question. Our current system would probably not be able to allocate the value produced by such automation efficiently.
The proponents of AI systems seem to mostly misunderstand what you're paying for really. It's not writing letters.
Love this story so much I just posted it. Although it's from an era in which you'd buy CDs and books containing contracts, it's still relevant with "AI".
> “No lawyer writes a clause who is not prepared to go to court and defend it. No lawyer writes words and let’s others do the fighting for what they mean and how they must be interpreted. We find that forces the attorneys to be very, very, very careful in verbiage and drafting. It makes them very serious and very good. You cook it, you eat it. You draft it, you defend it.”
Lawyers are humans. They make the same mistakes as others humans. Quality of work is variable across skills, education, and if they had a coffee or not that day.
I would appreciate a version of this paper that is worth reading, FWIW. The paper asks an important question: shame it doesn't answer it.
there will be caste of high-tech lawyers very soon which will be able to handle many times more volume of work thanks to AI, and many other lawyers will lose their jobs.
She's got international experience and connections but moved to a small town. She was a magic circle partner years ago. Now she has a FTTP connection and has picked up a bunch of contracts that she can deliver on with AI. She underbid some big firms on these because their business model was traditional rates, and hers is her cost * x (she didn't say but >1.0 I think)
Basically she uses AI for document processing (discovery) and drafting. Then treats it as the output of associates and puts the polish on herself. She does the client meetings too obviously.
I don't think her model will last long - my guess is that there will be a transformation in the next 5 years across the big firms and then she will be out of luck (maybe not at the margin though). She won't care - she'll be on the beach before then.
Lawyers have a duty of care and of candor and of confidentiality, and a ChatGPT-ish model like that can violate all three.
-- [0] https://www.americanbar.org/content/dam/aba/administrative/p...
I think "quite possibly malpractice" does a lot of work. It's "quite possibly malpractice" to use an associate to draft an opinion. The associate could secretly take photos of the document and sell them to the opposition or the press, the associate could make stuff up, the associate could get drunk in a bar and tell everyone what they are doing. Who knows. You better have appropriate management controls, working practices and training in place! Even partners do this stuff sometimes though...
It could also be malpractice to use a networked computer to prepare documents - that could be hacked into and your duty of care could be shown to be breached.
If there's no deception and appropriate care is taken then is there an issue?
I'm not saying "it's probably malpractice" just to say it, I'm saying it because the ABA has put out guidelines saying that, and so have most state bars.
If you're putting client data into ChatGPT, for example, that's violating your duty of confidentiality. Telling someone you're doing it doesn't excuse it.
And yes, many of the other situations you mentioned would also be grounds for censure or losing your license. An associate is an attorney, you might be confusing the term with paralegal.
An associate doing those things would likely lose their license. And law offices are required to use best practices for security controls.
The funny part is that people think we will run out of things to do. Most people never hire a lawyer because they are much to expensive.
And to be transparent I'm very bearish on what we are being marketed to as "AI"; I see value in the techs flying underneath this banner and it will certainly change white collar jobs but there's endless childish and comical hubris in the space from the fans, engineers, and oligarchs jockeying to control the space and narratives.
Is your theory that the next week there will be an AI lawyer that charges only 400$, then it is a race to the bottom?
There is a proven way to avoid a race to the bottom for wages, which is what a trade union does - a union by acting as one controls a large supply of labour to keep wages high.
Replace that with a company and prices, it could very well be that a handful of companies could keep prices high by having a seller's market where everyone avoids a race to the bottom by incidentally making similar pricing calls (or flat out illegally doing it).
The core problem is lawyers already automate plenty of their work, and lawyers get involved when the normal rules have failed.
You don't write a contract just to have a contract, you write one in case something goes wrong.
Litigation is highly dependent on the specific situation and case law. They're dealing with novel facts and arguing for new interpretations, not milling out an average of other legal works.
Also, you generally only get one bite at the apple, there's no do-overs if your AI screws up. You can hold a person accountable for malpractice.
this is true - and the majority of work of lawyers is in knowing past information, and synthesising possible futures from those information. In contracts, they write up clauses to protect you from past issues that have arisen (and may be potential future issues, depending on how good/creative said lawyer is).
In civil suits, discovery is what used to take enormous amounts of time, but recent automation in discovery has helped tremendously, and vastly reduced the amount of grunt work required.
I can see AI help in both of these aspects. Now, whether the newer AI's can produce the type of creativity work that lawyers need to do post information extraction, is still up for debate. So far, it doesn't seem like it has reached the required level for which a client would trust a pure ai generated contract imho.
I suspect the day you'd trust an AI doctor to diagnose and treat you, would also be the day you'd trust an AI lawyer.
US automotive, labor, and manufacturing unions couldn't remain competitive against developing economies, and the jobs moved overseas.
In the last few years, after US film workers went on strike and renegotiated their contracts, film production companies had the genius idea to start moving productions overseas and hire local crews. Only talent gets flown in.
What stops unions from ossifying, becoming too expensive, and getting replaced on the international labor market?
Labor action, such as strikes.
You can even use the same argument line against the wheel, electricity, or farming.
I was thinking religious leaders might get a good run. Outside of say, Futurama, I'm not sure many people will want faith-leadership from a robot?
I've already heard people comparing AI hallucinations to oracles (in the greek sense)
To the extent it requires actual faith it's presently a complete joke, of course, and I expect it will remain so for a long time. But I'd say the quality bar for congregation members is due for a rise.
[1] Arguably, full generality is not required, only generality across the labour domain in question. [2] Of course, the question remains: Who will be able to reap this reward. My bet is on the fabs + chip design.
If this kind of thing happens, if interest rates are 0.5%, then people on UBI could potentially have access to land and not have horrible lives, if it's 16% as these guys propose, they will be living in 1980s Tokyo cyberpunk boxes.
This is mainstream AI safety theory -- the term is "instrumental convergence". No matter what goal an optimizing system has, it tends to optimize for its own survival: after all, if it's an optimizer for <goal>, it wants to optimize for <goal>, so destroying it (or turning it off) will reduce the likelihood of achieving <goal>.
Unless that goal happens to be incredibly fine-tuned to our very complex human desires, we're not going to be happy when it goes off to do its thing.
The few exceptions are ones where you have the thing optimize for its own destruction, but those are rather less useful.
Tho I guess even post scarcity we'd have people who care about hoarding gold-pressed latinum.
That's the literal actual textbook definition of "communism".
Lmao that I actually lived to see the day when techbros seriously discuss this.
People have been making comparisons between post scarcity economics and "utopia communism" for decades at this point. This talking point probably predates your birth.
The economy cannot be sustained on the demand of a small handful of wealthy people. At a certain point, you either get a depression or hyperinflation depending on how the powers that be react to the crisis. In either case, the wealthy will have no leverage to incentivize people to do their bidding.
If your argument is, they'll just get AI to do their bidding, you have to keep in mind that "there is no moat." Outside of the ideological sphere, there is nothing that essentially ties the wealthy to the data centers and resources required to run these machines.
Revolts could be successful or not successful, with tech advancements in suppression (large scale surveillance, weaponry, various strike drones) chances of population to strike back become smaller.
Economy could totally be built around demand and wishes of super-rich, because human's greed and desires are infinite, new emperor may decide to build giant temple, and here you have multi-trillion economy how to make it running.
in a meantime, gap in wealth is keeping increasing, and AI will be one another multiplier.
There are fewer than ten nuclear states, two of which have almost all of the nukes. All but four countries in the world are signatories to the Treaty on the Non-Proliferation of Nuclear Weapon. Only nation-states possess nuclear weapons, even the wealthiest corporations and individuals would be obliterated if they ever tried to obtain or build them.
The idea that no moat exists is folly.
“I have a theoretical degree in economics”
You’re hired!
real talk though, I wish I had just encountered an obscure paper that could lead me to refining a model for myself, but it seems like there would be so many competing papers that its the same as having none
I don't expect dad to Do Your Own AI anytime soon, he'll still pay someone to set it up and run it.
1) all work gets done by AI. Owners of AI reap the benefits for a while. There is a race to the bottom concerning costs, but also because people are not earning wages and come ang really afford the outputs of production. Thus rendering profits close to zero. If the people controlling the systems do not give the people "on the bottom" some kind allowance they will not have any chance for income. They might ask horrible and sadistic things from the bottom people but they will need to do something.
2) if people get pushed into these situations they will get riot or start civil wars. "Butlerian jihads" will be quite normal.
3) another scenario is that the society controlled by the rich will start to criminalise non-work in the early stages, that will lead to a new slave class. I find this scenario highly likely.
4) one of the options that I find very likely if "useless" people do NOT get "culled" en mass is an initial period of Revolt followed an AI controlled communist "Utopia". Where people do not need to work but "own" the means of production (AI workers). Nobody needs to work. Work is LARPing and is done by people who act like workers but don't really do anything (like some people do today) A lot of people don't do this, there are still people who see non-workers as leeching of the workers, because workers are "rewarded" by ingame mechanics (having a "better job"). Parallel societies will become normal. Just like now. Rich people will give themselves "better jobs" some people dont play the game and there are no real consequences, but not being allowed to play.
5) an amalgamation of the scenario as above, but in this scenario everybody will be forced to larp with the asset owning class. They will give people "jobs" but these jobs are bullshit. Just like many jobs right now. Jobs are just a way of creating different social classes. There is no meritocracy. Just rituals. Some people get to do certain rituals that give them more social status and wealth. This is based on oligarch whims. Once in a while a revolt, but mostly not needed.
Many other scenarios exist of course.
It helps no one to explain to them just how much the boot stomps on their face. Left wing post modernist intellectuals have been doing this since the 60s and all it did was prevent any left winger from doing anything “revolutionary”.
Don’t waste your time reading “theory”. Look at what happened to Mark Fischer.
My prep is:
1) building a company (https://getdot.ai) that I think will add significant marginal benefits over using products from AI labs / TAI, ASI.
2) investing in the chip manufacturing supply chain: from ASML, NVDA, TSMC, ... and SnP 500.
3) Staying fit and healthy, so physical labour stays possible.
The only thing I see as obvious is AI is going to generate tremendous wealth. But it's not clear who's going to capture that wealth. Broad categories:
(1) chip companies (NVDA etc)
(2) model creators (OpenAI etc)
(3) application layer (YC and Andrew Ng's investments)
(4) end users (main street, eg ChatGPT subscribers)
(5) rentiers (land and resource ownership)
The first two are driving the revolution, but competition may not allow them to make profits.
The third might be eaten by the second.
The fourth might be eaten by second, but it could also turn out that competition amongst the second, and the fourth's access to consumers and supply chains means that they net benefit.
The fifth seems to have the least volatile upside. As the cost of goods and services goes to $0 due to automation, scarce goods will inflate.
It substitutes for human labour. This will reduce the price and substantially increase the benefits of land and resource ownership.
4) Develop an obsession for the customers & their experiences around your products.
I find it quite rare to see developers interacting directly with the customer. Stepping outside the comfort zone of backend code can grow you in ways the AI will not soon overtake.
#3 can make working with the customer a lot easier too. Whether or not we like it, there are certain realities that exist around sales/marketing and how we physically present ourselves.
At best AI will be a tool I use while developing software. For now I don't even think it's very good at that.
Last famous words.
Current technology can't do your job, future tech most certainly will be able to. The question is just whether such tech will come in your lifetime.
I thought the creative field was the last thing humans could do but that was the first one to fall. Pixels and words are the cheapest item right now.
I'm not aware of any big changes in writer/artist employment either.
The only argument you can have is to be cheaper than the machine, and at some point you won't be.
Things change and people adapt. Maybe my job won't be the same in 20 years, maybe it will. But I'm pretty sure I'll still have a job.
If you want to make big decisions now based on vague predictions about the future go ahead. I don't care what you do. I'm going to do what works now, and if things change I'll make whatever decisions I need to make once I have the information I need to make them.
You call me naive, I'd say the same about you. You're out here preaching and calling people naive based on what you think the future might look like. Probably because some influencer or whatever got to you. I'm making good money doing what I do right now, and I know for a fact that will continue for years to come. I see no reason to change anything right now.
The question is which probability do you assign to getting TAI over time? From your comment it seems you say 0 percent in your career.
For me it's between 20 to 80 percent in the next ten years ( depending on the day :)
Huge asteroid? We know full well how to deflect an asteroid -- launch rockets, deploy thrusters or explosives or &c, give it a little nudge and it'll miss the Earth by a huge margin. And better yet: everyone would be aligned on the need to do so!
AI on the other hand -- I don't think this applies. We couldn't stop nuclear proliferation -- what makes you so confident that we'd stop AI development before it was too late?
One believes the following:
> AI can't do my job and I doubt it will any time soon
The other believes the opposite; that AI is improving rapidly enough that their job is in danger "soon".
From a game theory stance, is there any advantage to holding the first belief over the second?
Meanwhile the engineer who isn't bothered by this bullshit prophecy goes about their day, making lots of money and becoming less replaceable every day. Maybe they learn to use these AI tools to be more efficient, which is really the only realistic endgame of AI tools anyway. You don't just fire all the devs and have some manager do the prompting. Maybe you fire some devs and keep the best ones as prompt engineers. Maybe this isn't even a management-driven process at all, maybe the developers just start using these tools of their own volition, become more productive and everyone's happy. It's not like we're running out of development work any time soon, whenever we meet a goal they set a new one. Being able to move faster doesn't necessarily mean we need fewer developers.
Setting aside hypotheticals and game theory, it's completely unrealistic to expect that software developer suddenly won't be a job any more. If it even happens it will be a slow, gradual process. The people working as software developers today will be prime candidates for using AI tools to create software. You still need to understand what you're doing, what's possible and what isn't etc. There is absolutely no reality where some business person just tells an AI to make a banking system and it does that perfectly without any human intervention.
I try to do 3 as much as possible.
My current work explicitly forbids me from doing 1. Currently just figuring out the timing to leave.
Sure, the rich might hire some more people to aim the AI for them, but who's got a monopoly on problems? Nobody. Every freelancer, farmer, or startup's got their own problems to fix, and cheap AI access means they can. The paper's obsessed with wealth grabbing all the future benefits, but problems are everywhere, good luck cornering that market. Every one of us has their own problems and stands to get personalized benefits from AI.
In the age of AI having problems is linked to receiving its benefits. Imagine for example I feel one side of my face drooping and have speech difficulty, and I type my symptoms into a LLM, and it tells me to quickly visit the doctors. It might save my life from stroke. Who gets the largest benefit here?
Problems are distributed even if AI is not.
If this were true, AWS wouldn't have pulled in well over $100bn in 2024. Nvidia wouldn't be worth $3.3tn.
The owners and builders of infra make a ton of money.
Almost everything on HN gets those comments. Look at the top comments of almost any discussion - they will be a rejection / dismissal of the OP.
- one is expanding on the topic without expressing disagreement
- one is a eulogy
- one expresses both agreement on some points and disagreement on other points
It asks the equivalent of "what if magic were true" (human-level AI) and answers with "the magic economy would be different." No kidding.
FWIW, the author is listed as a fellow of "The Forethought Foundation" [0], which is part of the Effective Altruism crowd[1], who have some cultish doomerism views around AI [2][3]
There's a reason this stuff goes up on a non-peer reviewed paper mill.
--
[0] https://www.forethought.org/the-2022-cohort
[1] https://www.forethought.org/about-us
[2] https://reason.com/2024/07/05/the-authoritarian-side-of-effe...
[3] https://www.techdirt.com/2024/04/29/effective-altruisms-bait...
Isn't developing AGI basically the mission of OpenAI et al? What's so bad about considering what will happen if they achieve their mission?
>who have some cultish doomerism views around AI [2][3]
Check the signatories on this statement: https://www.safe.ai/work/statement-on-ai-risk
And yes, that link is exactly the behavior I'm talking about.
It makes it sound like it's a top issue for all these people. Most of them work at EA-related places, but I spot-checked (really, I did) five people who worked in CS research but not at an EA company.
I couldn't find any other public statements that they thought AI was that important. Do they know their names are on this page?
Appeal to authority is a classic cult/fraud/misinformation/conspiracist tactic. If it says "so-and-so believes this" and not "so-and-so believes this and here's why", it's suspect.
This isn't science fiction.