On the one hand, I’m joking and really don’t want anyone to go through a war.
On the other hand, you have to admit is that postwar periods seem to be the one time citizens/voters/societies are willing to consider meaningful reforms/social changes.
Let me add that I think most of the big issues AI will actually raise aren't even on people's radar. For instance, huge parts of our legal framework only work because it is impractical to enforce them broadly. AI suddenly makes it possible to apply IP, defamation, product regulations etc etc to every random Facebook post, used furniture sale or even to conversations in a bar picked up by a microphone in the same way they are applied to massive corporations and huge YouTube accounts.
I don't think that is something we are even slightly prepared for yet. And this is just what some people have actually thought of, likely there will be things that are going to totally surprise us and we will look back at our attitudes today and wonder why we so foolishly were obsessed with stuff about AI safety or how AI affected education and weren't at all worried about these other things.
Until we have had the time to actually see what happens I don't think it's a good time to start trying to guess how to regulate the technology. Regulations have a way of being very sticky so even if long term I believe we will need them I'm very skeptical there is much to be gained by guessing at what we need now.
Long post, but TLDR is that I dont think there is going to be a mass unemployment issue from AI in next 2-3 years. Regardless of how good AI models get. I think over next 3-5 years, more and more areas are automated by AI, and hiring slows rather than mass layoffs. We'll see new graduate unemployment rise fastest first, and then when we have a recession/downturn after 3-5 years, we'll see a large spike in unemployment (and many of those become structurally unemployable).
As someone who heads a somewhat large operations function (100s of people) at a company, where I am investing heavily in AI to automate as much as possible, I have the following observations:
1) AI is evolving fast in terms of capability (models today significantly over perform 6-12 months ago)
2) Low hanging fruit for AI so far is like 10% of my teams workload (e.g., more customer support enquiries handled effectively by chatbot, straightforward creation of content) - and was quite to realise the productivity gains there
3) Most other areas are blocked not by AI "intelligence" but by its ability to interact with other systems needed to complete a task. So for example "log into a portal and access x,y,z or do x,y,z", OR blocked by access to specific data / information it needs to automate the task effectively.
4) None of those blockers are insurmountable by any means, but they take time, effort and focus to overcome, 1 by 1, which is what is going to slow down the pace of AI replacing people in jobs.
5) We are not adopting like 1 or 2 AI vendors (OpenAI, Anthropic etc.) who can do it all, but dozens, new companies who are applying gen AI to specific narrow use cases. Each of these companies need to be found by people in roles like mine and then they need sell me on how they can automate a part of my problem with a good ROI - again this takes time even if im focusing a bit chunk of my priorities on this.
6) I work in a later stage tech startup, so this is not a major factor for me, but my counterparts in large companies have so many policies restricting how they can utilise AI (liability, privacy, security concerns), that it takes a long time and a lot of effort for them to overcome these to implement solutions.
7) Narrow use cases, which have less of the above barriers, and with a lot of human costs involved will be be the initial focus and will result in fastest adoption by companies (coding, marketing content). My suspicion is that hiring will freeze, and attrition will gradually whittle away the team size in these areas rather than mass layoffs (both for rationale reasons: to stop spooking the key talent you will still need in these areas; and human reasons: most managers dont like letting people in their teams go"). That is until there is an economic shock (recession), when all of a sudden mass layoffs happen for these teams everywhere at once.
High scenario: AI becomes as capable as human workers. This calls for policy to ensure we get fully automated luxury communism instead of a dystopia in which the elite have all the robots and leave the rest of us with nothing.
Middle scenario: AI never becomes good at anything but pattern recognition, and LLMs never stop hallucinating. In this case only a few legal adjustments will be needed, because AI will be just another technology that enhances productivity for some tasks and not others, keeping the broad shape of our economic world as-is.
Bad scenario: AI tries to kill us all. This is not likely in my estimation but it is not impossible. Given how catastrophic this would be, it's worth thinking carefully about it and taking steps to prevent it.
We will see the most dramatic change due to AI/automation/robots after the next financial crisis. When the "everything" bubble pops, followed by mass unemployment, we usually see a huge uptick in employment as the economy recovers. Perhaps, this time will be different. Businesses can deploy AI/automation/robots as a lower cost way to scale up their business rather than hiring and training people. What will happen to the millions who are left behind? Will they all become carpenters, plumbers and electricians? What will society look like when only 20 to 30 percent of the work age adults are working and earning a paycheck? As a world society we should be working on a vision of the world we want to live in but I don't see that happening today.
If AI is merely a normal productivity-enhancing technology, it will still magnify inequality in the way that all productivity-enhancing technologies do: some people will learn to use it effectively in fields where it's relevant, other people's work will be the same as it ever was, and the disparity between those cohorts will grow.
It's deeply depressing that we currently have people in charge who are incapable of thinking through these types of issues. Vance is probably the leading intellectual elected republican - which is really saying something as Vance really isn't much of a critical thinker, which Matt astutely points out.
>What’s a viable tax regime for a world in which AI progress sharply diminishes the economic value of labor?
I think this hinges largely on the exact path by which AI progress affects the relative costs of all the things the economy does and will do.
Right now we have the benefit of being able to assume that, at some level, everything has its price determined by the cost of labor. Even land mostly acquires value because of the things labor can do to improve it or extract from it. There are some fundamentally constrained supply problems, but how those affect costs in the modern world depends largely on how much we can spend on labor to invent substitutes and workarounds to ameliorate them.
If and when AI breaks that dependency by becoming better and cheaper than humans at all manual and mental skills, relative price fluctuations could get wild. The cost to design a better car or plane, or to sue a major corporation, or to see a therapist, could fall to near zero, without (in the near term) affecting the price of a hamburger. What that means for investment returns, for spending patterns, for the cost of other goods and services, for where people choose to live, I have no idea. The world only has room for so many celebrities who are the best in fields a lot of people care about, and the caring only matters if those other people have enough spending power to keep the remaining human-centered industries going strong.
And what if I end up in 2035 owning a household robot that costs less than a car but can serve as a full-time live-in nanny, cook, butler, maid, chauffeur, nurse, doctor, gardener, landscaper, handyman, masseuse, personal trainer, and personal shopper? Well, that seems like it's going to have a lot of implications for the rest of everything else!
Note that both boosters and doomers assume that the LLMs will actually be "intelligent", but there is a 3rd case where their limitations become very apparent at some point.
Corporations are going whole hog on AI, but what happens when they notice diminishing productivity because the AI isn't as useful as claimed?
I am interested in the conversation about how AI is affecting our children’s prospects, but what I would find fascinating would be an analysis of how AI can be used to improve public education or healthcare.
I am a school nurse, and while AI could simplify some of my charting and data management, it is hard to see how it would improve the really hard parts of my job, like coaching autistic kindergartners through toilet learning, or reassuring the parents of children with diabetes that their child is safe on my campus, or endless meetings with unhoused parents to try to convince them that school attendance is something they should prioritize for their children in a life that is already overburdened. I would LOVE to have a computer program that wipes poopy bottoms or can teach a class of bouncy kindergartners how to wash their hands.
There are a lot of dire predictions about how AI is coming for the jobs of the college-educated but no conversation about how to adapt AI to relieve some of the stress healthcare workers and educators experience daily in caring for the bodies, minds, and spirits of humans.
One of the policies that would help would be encouraging a four hour work week (with reduced working hours), perhaps accompanied with transitional tax incentives that encourage employers to use this as a means to share the benefits of improved productivity.
Not to nitpick, but the start of the decline in bank tellers seems to coincide with the global financial crisis, which led to the closing and consolidation of many banks. Fewer banks = fewer tellers. And I'd argue that it was online and mobile banking, not the ATM, that continued to fuel the decline post financial crisis. Online banking's wide acceptance more closely correlates with the decline in tellers, rather than the ATM (which had been around for two decades without a negative effect on teller employment numbers).
Sharing some common themes with Matt, here is my policy perspective on the AI: "The Skeptical Enthusiast’s Guide to “AI at FDA” https://www.fdamatters.com/fdamatters/the-skeptical-enthusiasts-guide-to-ai-at-fda. Send any comments to me at sgrossman@hpsgroup.com.
Oh hey, nothing a good war won’t solve… :)
- - - -
On the one hand, I’m joking and really don’t want anyone to go through a war.
On the other hand, you have to admit is that postwar periods seem to be the one time citizens/voters/societies are willing to consider meaningful reforms/social changes.
I don’t know what to do about that.
I would pay for a podcast series between Matt and Nate Silver where they engage with experts on these topics.
Let me add that I think most of the big issues AI will actually raise aren't even on people's radar. For instance, huge parts of our legal framework only work because it is impractical to enforce them broadly. AI suddenly makes it possible to apply IP, defamation, product regulations etc etc to every random Facebook post, used furniture sale or even to conversations in a bar picked up by a microphone in the same way they are applied to massive corporations and huge YouTube accounts.
I don't think that is something we are even slightly prepared for yet. And this is just what some people have actually thought of, likely there will be things that are going to totally surprise us and we will look back at our attitudes today and wonder why we so foolishly were obsessed with stuff about AI safety or how AI affected education and weren't at all worried about these other things.
Until we have had the time to actually see what happens I don't think it's a good time to start trying to guess how to regulate the technology. Regulations have a way of being very sticky so even if long term I believe we will need them I'm very skeptical there is much to be gained by guessing at what we need now.
Long post, but TLDR is that I dont think there is going to be a mass unemployment issue from AI in next 2-3 years. Regardless of how good AI models get. I think over next 3-5 years, more and more areas are automated by AI, and hiring slows rather than mass layoffs. We'll see new graduate unemployment rise fastest first, and then when we have a recession/downturn after 3-5 years, we'll see a large spike in unemployment (and many of those become structurally unemployable).
As someone who heads a somewhat large operations function (100s of people) at a company, where I am investing heavily in AI to automate as much as possible, I have the following observations:
1) AI is evolving fast in terms of capability (models today significantly over perform 6-12 months ago)
2) Low hanging fruit for AI so far is like 10% of my teams workload (e.g., more customer support enquiries handled effectively by chatbot, straightforward creation of content) - and was quite to realise the productivity gains there
3) Most other areas are blocked not by AI "intelligence" but by its ability to interact with other systems needed to complete a task. So for example "log into a portal and access x,y,z or do x,y,z", OR blocked by access to specific data / information it needs to automate the task effectively.
4) None of those blockers are insurmountable by any means, but they take time, effort and focus to overcome, 1 by 1, which is what is going to slow down the pace of AI replacing people in jobs.
5) We are not adopting like 1 or 2 AI vendors (OpenAI, Anthropic etc.) who can do it all, but dozens, new companies who are applying gen AI to specific narrow use cases. Each of these companies need to be found by people in roles like mine and then they need sell me on how they can automate a part of my problem with a good ROI - again this takes time even if im focusing a bit chunk of my priorities on this.
6) I work in a later stage tech startup, so this is not a major factor for me, but my counterparts in large companies have so many policies restricting how they can utilise AI (liability, privacy, security concerns), that it takes a long time and a lot of effort for them to overcome these to implement solutions.
7) Narrow use cases, which have less of the above barriers, and with a lot of human costs involved will be be the initial focus and will result in fastest adoption by companies (coding, marketing content). My suspicion is that hiring will freeze, and attrition will gradually whittle away the team size in these areas rather than mass layoffs (both for rationale reasons: to stop spooking the key talent you will still need in these areas; and human reasons: most managers dont like letting people in their teams go"). That is until there is an economic shock (recession), when all of a sudden mass layoffs happen for these teams everywhere at once.
I see three possible futures, regarding AI:
High scenario: AI becomes as capable as human workers. This calls for policy to ensure we get fully automated luxury communism instead of a dystopia in which the elite have all the robots and leave the rest of us with nothing.
Middle scenario: AI never becomes good at anything but pattern recognition, and LLMs never stop hallucinating. In this case only a few legal adjustments will be needed, because AI will be just another technology that enhances productivity for some tasks and not others, keeping the broad shape of our economic world as-is.
Bad scenario: AI tries to kill us all. This is not likely in my estimation but it is not impossible. Given how catastrophic this would be, it's worth thinking carefully about it and taking steps to prevent it.
We will see the most dramatic change due to AI/automation/robots after the next financial crisis. When the "everything" bubble pops, followed by mass unemployment, we usually see a huge uptick in employment as the economy recovers. Perhaps, this time will be different. Businesses can deploy AI/automation/robots as a lower cost way to scale up their business rather than hiring and training people. What will happen to the millions who are left behind? Will they all become carpenters, plumbers and electricians? What will society look like when only 20 to 30 percent of the work age adults are working and earning a paycheck? As a world society we should be working on a vision of the world we want to live in but I don't see that happening today.
If AI is merely a normal productivity-enhancing technology, it will still magnify inequality in the way that all productivity-enhancing technologies do: some people will learn to use it effectively in fields where it's relevant, other people's work will be the same as it ever was, and the disparity between those cohorts will grow.
It's deeply depressing that we currently have people in charge who are incapable of thinking through these types of issues. Vance is probably the leading intellectual elected republican - which is really saying something as Vance really isn't much of a critical thinker, which Matt astutely points out.
>What’s a viable tax regime for a world in which AI progress sharply diminishes the economic value of labor?
I think this hinges largely on the exact path by which AI progress affects the relative costs of all the things the economy does and will do.
Right now we have the benefit of being able to assume that, at some level, everything has its price determined by the cost of labor. Even land mostly acquires value because of the things labor can do to improve it or extract from it. There are some fundamentally constrained supply problems, but how those affect costs in the modern world depends largely on how much we can spend on labor to invent substitutes and workarounds to ameliorate them.
If and when AI breaks that dependency by becoming better and cheaper than humans at all manual and mental skills, relative price fluctuations could get wild. The cost to design a better car or plane, or to sue a major corporation, or to see a therapist, could fall to near zero, without (in the near term) affecting the price of a hamburger. What that means for investment returns, for spending patterns, for the cost of other goods and services, for where people choose to live, I have no idea. The world only has room for so many celebrities who are the best in fields a lot of people care about, and the caring only matters if those other people have enough spending power to keep the remaining human-centered industries going strong.
And what if I end up in 2035 owning a household robot that costs less than a car but can serve as a full-time live-in nanny, cook, butler, maid, chauffeur, nurse, doctor, gardener, landscaper, handyman, masseuse, personal trainer, and personal shopper? Well, that seems like it's going to have a lot of implications for the rest of everything else!
Note that both boosters and doomers assume that the LLMs will actually be "intelligent", but there is a 3rd case where their limitations become very apparent at some point.
Corporations are going whole hog on AI, but what happens when they notice diminishing productivity because the AI isn't as useful as claimed?
I am interested in the conversation about how AI is affecting our children’s prospects, but what I would find fascinating would be an analysis of how AI can be used to improve public education or healthcare.
I am a school nurse, and while AI could simplify some of my charting and data management, it is hard to see how it would improve the really hard parts of my job, like coaching autistic kindergartners through toilet learning, or reassuring the parents of children with diabetes that their child is safe on my campus, or endless meetings with unhoused parents to try to convince them that school attendance is something they should prioritize for their children in a life that is already overburdened. I would LOVE to have a computer program that wipes poopy bottoms or can teach a class of bouncy kindergartners how to wash their hands.
There are a lot of dire predictions about how AI is coming for the jobs of the college-educated but no conversation about how to adapt AI to relieve some of the stress healthcare workers and educators experience daily in caring for the bodies, minds, and spirits of humans.
One of the policies that would help would be encouraging a four hour work week (with reduced working hours), perhaps accompanied with transitional tax incentives that encourage employers to use this as a means to share the benefits of improved productivity.
This is very good, Matt. Thanks for saying it so clearly.
Not to nitpick, but the start of the decline in bank tellers seems to coincide with the global financial crisis, which led to the closing and consolidation of many banks. Fewer banks = fewer tellers. And I'd argue that it was online and mobile banking, not the ATM, that continued to fuel the decline post financial crisis. Online banking's wide acceptance more closely correlates with the decline in tellers, rather than the ATM (which had been around for two decades without a negative effect on teller employment numbers).
Within the tech industry, the hype around AI has already subsided outside companies specifically working on it. So, like, I don’t really see it.