While courts continue to often find briefs written by AI, it is interesting to note that papers in top law journals have thus far not fallen to the trend.
"If you believe [that the AI can do everything a human can do, but better], then it’s almost tautological to say that there will be no jobs left for humans."
Ok I finally got around to reading this fully. Very interesting take and a very good point.
I think it depends on a few things.
First, that there are things that the AI can do that produce ~1 OOM more value than things humans already do
Second, that such things cannot get 'saturated' so to speak.
Let's say you use the AI bots for finance. An AI that is 10x better than everyone else would make tons of money, and using AI for finance would immediately be the most valuable thing to do. Until the markets settle to their actual value every time (because the agents get it perfect) and then suddenly any incremental spend on AI in finance no longer matters. So then you have the AI go and do the next most profitable thing, and then the next most, and so on. It's possible that you run out of computer before you run out thing, but unclear where that cashes out/how many things get left for the meat sacks
Oh, I'm also skeptical whether comparative advantage will be any real help. There's a certain overhead involved in delegating a task to another person.
You might delegate tasks to someone who is 2x slower than you are, just because you have more valuable things to do. But you wouldn't delegate a task to someone who takes 100x longer, because in the time it takes to explain what you want you could just do it yourself.
Sure, I think it's reasonable to be skeptical! I don't think Smith's article means that there's no way for AGI to take all of our jobs (in fact, he lists several!)—I just think that it means that we shouldn't accept the premise "If AGI, then necessarily unemployment". The premise "If AGI, then possibly unemployment, given certain conditions" is certainly valid!
I guess it depends on how fast you think AI can learn to take over a new job.
If humans can learn new skills faster than the AI, they can stay one step ahead, continually reinventing themselves for whatever new jobs are available before the AI snatches it.
If the AI can learn faster than humans (which I think it can in most cases because the training can be parallelized), that doesn't seem like a promising strategy.
This is why I've been saying for a while -- that AI summons us to be truly human. It really is all that's left. And forcing / empowering us to confront the question of what it means to be human, truly human, is scary and difficult and by definition a spiritual question. AI is not a normal technology, it is a metaphysical one.
The answer is verification. The jobs that new technologies can't automate away are ones humans keep for themselves by regulatory capture. If my fleet of self driving cars require by law a 40 cars to one human ratio as safety backups, then that's that.
You underestimate the political and social savvy of humankind.
The benefit of a human being less competent than a program is that they are less competent. The benefit is then bodies you can trust more, even if you are a program.
The world is full of tacit knowledge , domain expertise, knowledge that can’t easily be trained on or extracted, and the inertia of institutions and interpersonal relationships . If the cost of extracting incredible amounts of high quality data on a task isn’t economically feasible it’s not gonna happen. What would need to happen is orders of magnitude better sample efficiency in training and that’s something I haven’t seen yet
the only setting where what you're saying is true is where the cost of extracting the data is < the total cost of all of the people who are doing that thing, but i can't think of many jobs like that
It gets a lot more granular than that though. A manager at Wendy’s knows about the interpersonal relationships between the employees, the specific way to fix this frier when it breaks down, the nuances around collecting garbage for that location, how to deal with their specific suppliers. And every job has thousands of threads of tacit knowledge that only lives in the employee’s head
It still can't count fingers, or EVEN follow the rules when playing chess, or solve river crossing problems (without some developer sitting down and writing a river crossing solver and putting it in as a big if statement), or any of the actually hard problems Gary Marcus put in his list. They are good at problems where you can easily verify the solution like Math (lean) or Writing code (tests)... But the millionth example of progress in those categories are not steps towards AGI. Sorry.
(it can also definitely do most of those things, idk which version of the tooling Gary is using or when you yourself last interacted with one on the frontier models)
I use frontier models for coding almost every day. Definitely didn't see any movement on anything other than coding. If anything the writing quality is going down.
i strongly agree that robotics is hard, but also it is hard not to feel like the objections raised here are ~equivalent to people being like 'ai doesnt understand image composition it cant put a horse on top of a cowboy' -- the bitter lesson has come for many things that are harder than dexterity and latency
Looking at the history of AI, we often learn in retrospect that our intuitions about what’s easy versus hard were wrong. Of course, my intuition about this could be wrong, too! This can’t be settled by comparing vibes. It deserves a real investigation.
While courts continue to often find briefs written by AI, it is interesting to note that papers in top law journals have thus far not fallen to the trend.
https://reason.com/volokh/2026/09/07/how-often-are-articles-in-top-law-reviews-written-in-part-by-ai/
"If you believe [that the AI can do everything a human can do, but better], then it’s almost tautological to say that there will be no jobs left for humans."
I believed this avidly all the way from college to when I read this one piece by Noah Smith that completely changed my mind: https://www.noahpinion.blog/p/plentiful-high-paying-jobs-in-the
Ok I finally got around to reading this fully. Very interesting take and a very good point.
I think it depends on a few things.
First, that there are things that the AI can do that produce ~1 OOM more value than things humans already do
Second, that such things cannot get 'saturated' so to speak.
Let's say you use the AI bots for finance. An AI that is 10x better than everyone else would make tons of money, and using AI for finance would immediately be the most valuable thing to do. Until the markets settle to their actual value every time (because the agents get it perfect) and then suddenly any incremental spend on AI in finance no longer matters. So then you have the AI go and do the next most profitable thing, and then the next most, and so on. It's possible that you run out of computer before you run out thing, but unclear where that cashes out/how many things get left for the meat sacks
Oh, I'm also skeptical whether comparative advantage will be any real help. There's a certain overhead involved in delegating a task to another person.
You might delegate tasks to someone who is 2x slower than you are, just because you have more valuable things to do. But you wouldn't delegate a task to someone who takes 100x longer, because in the time it takes to explain what you want you could just do it yourself.
Sure, I think it's reasonable to be skeptical! I don't think Smith's article means that there's no way for AGI to take all of our jobs (in fact, he lists several!)—I just think that it means that we shouldn't accept the premise "If AGI, then necessarily unemployment". The premise "If AGI, then possibly unemployment, given certain conditions" is certainly valid!
I guess it depends on how fast you think AI can learn to take over a new job.
If humans can learn new skills faster than the AI, they can stay one step ahead, continually reinventing themselves for whatever new jobs are available before the AI snatches it.
If the AI can learn faster than humans (which I think it can in most cases because the training can be parallelized), that doesn't seem like a promising strategy.
This is why I've been saying for a while -- that AI summons us to be truly human. It really is all that's left. And forcing / empowering us to confront the question of what it means to be human, truly human, is scary and difficult and by definition a spiritual question. AI is not a normal technology, it is a metaphysical one.
My husband’s running joke since 2018 is that the goal of all data scientists is to be the last data scientist on a heap of data scientist corpses.
The answer is verification. The jobs that new technologies can't automate away are ones humans keep for themselves by regulatory capture. If my fleet of self driving cars require by law a 40 cars to one human ratio as safety backups, then that's that.
You underestimate the political and social savvy of humankind.
The benefit of a human being less competent than a program is that they are less competent. The benefit is then bodies you can trust more, even if you are a program.
If it can do everything a person can, better, then we're not talking about a tool — we're talking about a replacement, and the induction holds.
Funny how "let's ignore the hollowed-out towns" gets filed under a parenthesis, like grief owed a footnote.
I'll take the argument seriously the day someone names the job that survives the machine and isn't just supervising it.
The world is full of tacit knowledge , domain expertise, knowledge that can’t easily be trained on or extracted, and the inertia of institutions and interpersonal relationships . If the cost of extracting incredible amounts of high quality data on a task isn’t economically feasible it’s not gonna happen. What would need to happen is orders of magnitude better sample efficiency in training and that’s something I haven’t seen yet
the only setting where what you're saying is true is where the cost of extracting the data is < the total cost of all of the people who are doing that thing, but i can't think of many jobs like that
It gets a lot more granular than that though. A manager at Wendy’s knows about the interpersonal relationships between the employees, the specific way to fix this frier when it breaks down, the nuances around collecting garbage for that location, how to deal with their specific suppliers. And every job has thousands of threads of tacit knowledge that only lives in the employee’s head
I suspect this is one of those "it happens slowly and then all at once" things, though it's already happening pretty fast so IDK
It still can't count fingers, or EVEN follow the rules when playing chess, or solve river crossing problems (without some developer sitting down and writing a river crossing solver and putting it in as a big if statement), or any of the actually hard problems Gary Marcus put in his list. They are good at problems where you can easily verify the solution like Math (lean) or Writing code (tests)... But the millionth example of progress in those categories are not steps towards AGI. Sorry.
(it can also definitely do most of those things, idk which version of the tooling Gary is using or when you yourself last interacted with one on the frontier models)
I use frontier models for coding almost every day. Definitely didn't see any movement on anything other than coding. If anything the writing quality is going down.
As I said at the top, the only debate that matters is if you agree that AI will have certain capabilities. Sounds like you disagree.
> the world’s capital is funneling into robotics and major advances are happening constantly.
This is the crux of the argument. It would be worth devoting a blog post to it. For another perspective, see:
https://secondthoughts.ai/p/14-reasons-robotics-is-hard
i strongly agree that robotics is hard, but also it is hard not to feel like the objections raised here are ~equivalent to people being like 'ai doesnt understand image composition it cant put a horse on top of a cowboy' -- the bitter lesson has come for many things that are harder than dexterity and latency
Looking at the history of AI, we often learn in retrospect that our intuitions about what’s easy versus hard were wrong. Of course, my intuition about this could be wrong, too! This can’t be settled by comparing vibes. It deserves a real investigation.
*there