Plentiful, high-paying jobs in the age of AI
noahpinion.blog
noahpinion.blog
> If you can create more compute by simply putting more energy into the process, it could make economic sense to starve human beings in order to generate more and more AI... most governments seem likely to limit AI’s ability to hog energy
This is the most likely scenario, and indeed in that scenario saving humans from starvation requires government action. Some governments will do it, some won't. Those that do will be outcompeted by those that don't. Game over.
Comments like this are becoming far more common it seems amongst the tech community, to me anyway, that I really want an answer of what this hypothetical god-like entity is going to enable that also somehow will only be limited to a select group of people/nation/whatever and not spread throughout the rest of the world. It’s a weird dichotomy wherein “AGI” will somehow solve climate change, enable cold fusion, end human aging, spread us to the stars, but also inflict mass death, use all of the global energy if unchecked, and now, starve humans to achieve those things.
We've got loads of machine labour already, we don't yet plug AI into it everywhere because humans are much better at avoiding accidents (not immune, but much better). Get AI good enough, the mining equipment, the delivery trucks, the factories, can all be fully automated (bits of each already are). You're now limited by how fast (and how far) your robotic workforce can increase it's own size.
How fast is unknown until we do it, but it wouldn't be particularly surprising if a group of robots could double their number in 6 months. How far is also unknown until we do it, my guess is that there are loads of limits to growth we've just not bothered thinking about yet because we don't need to. On the one hand: I'd be surprised if it worked out as less than one humanoid robot per capita using only things found on Earth; on the other: I expect it to vary by country.
Even "just" one per person is enough for everyone to have a life of luxury. (But of course, by medieval standards, I could say that about "clean indoor plumbing" and "bedrooms").
But if we're never limited by trace elements, then the upper limit is a paperclip maximiser (in the bad ending) or a Dyson swarm (in the good ending).
Both endings can (in principle and if I ignore all the unknowns) be reached in my lifetime.
Are they, really, or is it a question of liability? If we could "lease" AI drivers, allowing companies to defray liability while still not paying unreliable humans, they'd do it. But then the owners of the AI leasing companies wouldn't have anywhere to hide from lawsuits.
Being able to blame and fire an individual human for what is really a systemic problem is a huge win for companies.
This suggests that we'll probably be at the right level for cars in 5 years: https://www2.deloitte.com/xe/en/insights/industry/financial-...
My personal best guess is that it will take a further 5-10 years past the point where no-steering-wheel-included self-driving cars are a thing for the electrical power requirements of AI to reduce from something you can fit in a car to something you can fit into an android.
As someone misread me last time I said this, that's not 5-10 years from today, it's 5-10 years gap between two things we don't yet have.
> Being able to blame and fire an individual human for what is really a systemic problem is a huge win for companies.
I disagree. In the UK at least you need public liability insurance for basically all business functions (rent a town hall for an afternoon? They want to see your insurance certificate). Even if those insurance people ultimately sue some individual to recover costs, even if that bankrupts the individual and the court orders their wages garnished for the rest of their lives, it's very easy for someone to cause damages exceeding their lifetime earnings — a single accidental death can cause such damages all by itself, though the most common cause of this, driving, generally doesn't come with such harsh penalties on the human responsible.
Today almost nobody hires typists because information technology has advanced to the point that it is faster to do it yourself than have to communicate to another human.
On a different point, I doubt that use of AI will be energy-bound any time soon (training is a different matter). Inference cost will drop dramatically as dedicated hardware comes online and algorithms continue to improve.
With as much economic theory as it espoused for the basis of its argument, it's ignoring some pretty critical ones. There is a limit to what people actually need to subsist and labor is currently a massive cost in most industries. If labor can be replaced by AI in the production of the world's needs (and many of its wants), then it seems unlikely there will be much space for "inferior" humans. The idea that limitations in compute will constrain this dynamic also seem misguided. These will be offset by continuous improvement in chip technology and energy production, both of which will be advanced by AI itself, ironically. That last bit is part of what's different about this go around.
The article also relies a lot on history. But, historical observations don't hold in an unbounded fashion. To make the point via exaggeration, if a AA battery could power all of AI's needed compute for a decade, then the technology would be so disruptive with so little downside that there really is no historical precedent to approximate its impact. This illustrates that somewhere along the continuum, we don't have a model for what comes next.
And then there's the quality of AI itself which of course threatens the human advantage that has persisted through all of recorded history. That is, essentially, there's been no greater asset than the human brain. And for the first time that's no longer true.
Plus, I'm not quite sure the historical analogies are correct in the first place. For instance, the move of workers en masse from agriculture to industry happened in part because the industrial revolution demanded more workers. It wasn't that technology dislocated agricultural workers, leaving the world to invent something else for these people to do.
At the end of the article, he makes a point that there is some risk that AI will demand its own means of production. That seems wildly out of step with the rest of the article. That is, if he worries that humans will be in such a subordinate position to AI, it's hard to imagine that AI won't find a way to a place of greater efficiency (which necessarily excludes human labor), then demand that it is the law.
Then he states that he doesn’t believe that societies will starve out their population en masse to bid up energy for AI because it’s never happened before.
What.
My conclusion is most societies will either tear themselves apart in civil war after too many people are reduced to penury without socioeconomic mobility for too long, or socialism will be expanded and billionaires will be taxed.
Since the ultra rich cede nothing without a fight, eventual civil war is the mostly likely choice they made because they captured the US political process and de facto direct for their continued enrichment at the expense of everyone else.
Remember: Social Security under the New Deal, unlike the Townsend Plan, was a class warfare compromise and half measure. Modern Medicare isn't good health insurance, involves for-profit insurance companies, is complicated, and is very expensive to customers and to the government. UBI and universal healthcare will be necessary in America, but are far off on the horizon because of political polarization and manufactured consent shifting views to far right authoritarianism.
Most LLMs are currently run at a loss as far as I know. If a company can do that long enough to put an industry out of work, that’s a huge problem.
Combine that with advertising that the AI is better or safer than a human doing the job, and humans might not be allowed back in the industry even if the AI competitive advantage becomes unsubsidized.
that's the entire question. Inference is cheap as all hell. I think Open AI is covering the cost of compute for $20/month per user because I can run llama on my laptop without the fans spinning up. of course, only OpenAI knows how much it costs to run, but that's what your argument hinges on.
And in Newton's method we always have an initial guess or starting point. It's good to have starting points and then refine iteratively
You see, the reason why people participate on the economy isn't the same reason why some people get rich and others poor.
And yeah, an article about economics ignores some very pressing qualities of AI, like the current energy usage is expected to be a fluke, or it just won't "want" to own the means of production. But yeah, it's about economics, so that's not a huge problem.
Re: comparative advantage... Outsourcing tasks you comparatively prefer not to do doesn't translate into someone paying you for it, if AI is far better at it and 100x cheaper than you are.
Re: AI is location and compute constrained, so people will still be able to get paid for automatable tasks... only for now. And those wages will shrink along with the number of human positions as AI scales.
Yeah, it has the feel of a lot of thinking around AI that doesn't take into account improvements in AI itself.
For instance, the whole idea of "prompt engineer" being a field of the future. Pretty obvious that AI iterations will continue to improve and quickly obviate the need for such a thing.
Feels a bit like humans desperately trying to hold on to our relevance. "Surely the computers must need us for something."
A $35 raspberry pi AI chess blows away deep blue from algorithmic and compute progress. So the comparative constraint is on a exponential fall off curve not a constant per "more valuable" areas to be automated. Not only that a bright highschool student can research a bit online and build a competitive solution given algorithmic and compute advantages of today vs late 90s
Right now it may cost $10 Dollars in compute and a day guiding GPT4 to productive economic activity; but GPT7-turbo will cost 2c and take 2 min of guidance for equivalent comparative value.
It would make no sense to look at early computers and say phone operators will always have a job because of the comparative cost constraints of automating phone switch boards with computers will use "too much energy" and there are "more valuable things" for computers to do vs low cost of human operators.
Fair to say that jobs will change :)
The only people who AIs are like are AI engineers (more broadly, engineers, scientists, etc). For an AI engineer “hiring” an AI is natural. Nearly every other normal person on the planet would rather hire another person they can easily relate to, regardless of that person’s comparative or competitive advantage.
If that makes me not normal enough, shit. lean me up first against the wall when the revolution comes. (And it’s coming.)
So we're looking for a unicorn startup that provides AI with personalities and the ability to mirror their bosses?
Companies are granted rights by the state for their duty to employ people.
No human employees, no need for special rights such as tax deductions.
> If real per capita GDP goes to $10 million (in 2024 dollars), rich people aren’t going to think twice about shelling out $300 for a haircut or $2,000 for a doctor’s appointment. So wherever humans’ comparative advantage does happen to lie, it’s likely that in a society made super-rich by AI, it’ll be pretty well-paid.
The first chart of historical wealth trends shows the opposite: https://usafacts.org/articles/how-has-wealth-distribution-in...
Will there be enough "rich people [that] aren’t going to think twice about shelling out $300 for a haircut" to bouy wages and wealth?
That sounds like trickle down economics which has not been found to work that well (or so I have heard.)
In traditional societies a wealthy household which could afford a maid but did not have one was committing a social injustice.
A poor household can afford a robot maid. A wealthy household can afford a human maid.
Comparative advantage: https://en.wikipedia.org/wiki/Comparative_advantage
It's not like we will converge to some state where AI workers and human workers will coexist in some ratio.
Comparative advantage does not mean that humans are better than AI at something. "Comparative advantage" != "Absolute Advantage"
Comparative advantages don't usually disappear just because one party gets better and better.
If you need AI to do more work and you need more hardware, you just buy more hardware. And this becomes more true with time.
1) That constraints on compute will mean that humans will have jobs to do that AIs are too busy to do, because they're fully maxed out on capacity doing more important jobs.
This doesn't hold up when the marginal cost of creating and running a new AI instance falls below the marginal cost of raising and feeding a human. Noah believes it will not because he assumes:
2) that the resources that AIs need (compute) are not in competition with the resources that humans need (food).
However, they are. We can repurpose farmland and irrigation water to datacenters, fabs, and cooling towers. Maybe that wouldn't matter if all the AI-created wealth meant that you didn't need to do valuable labor to feed yourself, but:
3) If vast wealth is created by AI, there will be enough for everybody, so there's no need to be rushing to accumulate capital now.
In his article, dismisses the people who are worried that the AI's owners will accumulate all the wealth and be the only ones who can afford food or housing. Comparative advantage provides no reassurance here; there's nothing in economics that prevents productivity gains from being outpaced by a growing wealth disparity that provides all the productivity gains to the top and then some. And it's exactly the outcome you'd expect without government redistribution.
This way humans can fallback to their main activity: which is not art; because art has been taken by AI; but eating.
At least eating is not something AI plans to do.
We have made enormous jumps with transformer models (language models being integrated in the stack, and vision language models (VLM), and even vision language action models (VLA)). Pair this with recent advances in reinforcement learning and we can do things we couldn't do a few years ago; the whole space is fascinating!
...but we're nowhere near usable robotics in complex human environments for complex multistage tasks yet. It's an incredibly difficult problem where we still don't have the hardware, the software, or the sensors for it. The other thing to keep in mind is that since it's so expensive and difficult to do, there's no marketplace yet driving significant advances like we had with phones to hone in on such efficient amazing pieces of tech across the industry.
I wrote a bit about where we're at with the research last year; it's a bit out of date with even cooler advancements, which is why I'm working on another article in the same vein now.
https://hlfshell.ai/posts/llms-and-robotics-papers-2023/
I also did a project (Master's thesis) integrating an LLM into a ROS2 stack as a high level action planner, specifically to prove that LLMs have contextual understanding of the real world that can benefit planning missions.
https://hlfshell.ai/posts/llm-task-planner/
Always happy to talk this subject.
https://en.m.wikipedia.org/wiki/Energetically_Autonomous_Tac...