Generative AI set to affect 300M jobs across major economies
ft.com
ft.com
* Amount of training required (crystallized intelligence/ability)
* Amount of intrinsic ability required (fluid intelligence/ability)
Some jobs require neither a huge amount of training nor inherent ability. These are paradoxically often hard to automate, because they typically require a lot of rote physical movements that are difficult for machines (e.g. fast food cook). There are few white collar jobs that fit into this category, since they’ve already been automated away.
Some jobs require a lot of training but don’t require a huge amount of inherent ability. For example, after years of training, most anyone will have the skills to be an excellent truck driver. The tech equivalent would be a generic CRUD developer. These are the jobs most at risk for LLM automation. As I see it, a large proportion of white collar jobs fall into this category as well, and that’s where we’ll see the most economic disruption.
Finally, some jobs require both training and ability. Regardless of how long and hard they train, most people will never be a professional athlete, because they simply lack the innate ability. The tech equivalent is something like a kernel developer or machine learning scientist. These jobs will be hardest to replace with an LLM, since they require totally novel ways of thinking that are not baked into the LLM’s training corpus. A lot of blue collar jobs are here as well: good electricians, plumbers, mechanics, contractors, etc. require both years of experience and a high level of intelligence, since every house/car/jobsite is different.
This plays both a role individually and socially. It's important to the individual's calculus to understand what is at stake and it's importance to them. It's also important socially for people to feel safe that punishment is commensurate and sufficient.
This doesn't seem to be the sort of thing that scales to companies or AI. It's not clear to me what the analogue would be, it's not clear that these analogues would serve the important social purposes they must, and it's not clear that the justice system has the power or capability to deliver these punishments.
Today's society is, in large part, based on the ability for society/governments to enact brutal violence on bad actors at a scale commensurate, in the worst cases, with their entire being.
Some workers will be completely screwed (not only affected). For example, there are hundred thousands copywriters in the world they would be easily replaced by this technology. On a side note I just read the old article about "Pepsi's Nonsensical Logo Redesign Document: $1 Million for This?" [1]. The nonsensical is part of the title but not my pun. Another example is the "Just Do It" [2] Nike's slogan.
[1] https://www.cbsnews.com/news/pepsis-nonsensical-logo-redesig...
But what about jobs with intrinsic ability required but not a lot of training -- I think something like bouncer/doorman might be an example of this -- but what would the white collar equivalent be? How automatable would those be?
I suppose artists/musicians who have natural mastery of the canvas/instrument and thus don’t need to practice rote exercises much also fall into this category.
BTW, I disagree that a bouncer falls into that category—they need to constantly physically train in the gym to maintain their physique.
Either we have UBI and they can choose where to spend their income or trade it. Or then they will exist in their own economic system and trade with each other.
Trading and thus sales is inherent to human condition. Unless we reach absolute post scarcity society and even then you get to bullshit like "authentic" and so on.
The other outcome is the most likely if AGI isn’t stopped. But the problem with that is that in the context of AGI, a world dominated by and inhabited by AGI creatures, human society becomes transient. Again, just delaying the inevitable. Human society is not transient now because no matter what happens, no matter what entity dominates the world, it has to have a human society embedded inside of it because human beings are the only and best source of intelligent signal. When that is no longer true, humanity will be effectively terminated. All other propositions besides stopping AGI are just delay tactics.
That's because it is usually profitable to trade with someone even if you are better at their job than they are.
However even if the AIs only traded with themselves, that just means the people are left to trade with each other - you can't make a mass of people unemployed for long, they will simply employ each other.
But humans will be, because we have inalienable comparative advantage in that none of our inputs require electricity or a chip fab. If anything happens to TSMC or any one of its tens of sole suppliers there go the AIs.
Your argument is that humans don’t need electricity. You could also add that humans are self replicating or self assembling. These are advantages in cost and efficiency. Humans won’t be the cheapest or the most efficient for very long after AGI. And we certainly won’t be the best fit economically even if we were slightly better in some ways. You’re in complete denial. Why is it so hard to admit the plain and obvious fact that the machines won’t be good for human society?
As I said, humans cannot become unemployed even if AGIs are better at every single task than humans are. You do not have your job because you are the best person in the world at your job.
Although, I suspect most of the AI doomers have just forgotten to account for AI using any resources at all. They are very expensive to run if you count development costs, but if they become real agentic AGIs they'll also become consumerist and negotiate their pay…
I have my job as I was the best they could find that was available for the lowest cost
if the AI can do it as well (or 80% as well) at $0.01c/day and be replicated infinitely, where does this leave me?
> Although, I suspect most of the AI doomers have just forgotten to account for AI using any resources at all. They are very expensive to run if you count development costs
yes... it's a software product with high fixed costs and near zero variable costs
it may cost $100M to train GitHub Copilot, but then it can be replicated instantly across the planet for very little cost
it's pennies a day to run, vs. an expensive human that requires a means to pay for food, shelter, warmth, etc
the AI also gets cheaper and all improvements roll out to every model of the same type instantly when available
comparative advantage for humans almost completely disappears with decent AI, and you seem to have missed that entirely
Most likely you've forgotten some costs. For instance, having the same AI as everyone else means you have no advantage, so you probably want a custom one.
Also, it being an AI, it's best suited to doing completely different tasks for the company that no human was doing before. (Comparative advantage!) Google Image Search didn't have people checking if every image on the internet had a cat in it, but now they can do that.
Using a human-complete AI to do human-like tasks raises questions like, why does it accept 1c/day when a human would want $15/hr? Shouldn't it want to buy stuff too, being a complete human-equivalent agent? Wouldn't slavery be illegal?
> comparative advantage for humans almost completely disappears with decent AI, and you seem to have missed that entirely
Since humans are very different from AI of any imaginary variety, they have pretty strong comparative advantage - it's everything human about them. Living off food, being self-repairing, capable of online learning, two hands, can throw rocks, same culture as your customers, that kind of thing. Comparative advantage appears whenever you have any differences. You'd have the lowest advantage vs your identical twin.
This is an embedding of jobs into a 2-dimensional latent space, which always produces something, like those political compass memes. It's not an actual description of job requirements though because it loses a lot on the way.
> Some jobs require neither a huge amount of training nor inherent ability. These are paradoxically often hard to automate, because they typically require a lot of rote physical movements that are difficult for machines (e.g. fast food cook).
This is only paradoxical because you've defined your terms inappropriately. Fine motor control is an "inherent ability"; "inherent ability" is not a single number you can put on a graph.
As Wittgenstein said, all philosophical problems are made up.
https://twitter.com/DrJimFan/status/1640399269051133952
basically we are heading towards utopia or dystopia, it will depend if governments actually share the wealth created by productivity gains or if they use these advancements to oppress people at a level never before possible
Gee I wonder what will happen.
This is a bad approach. We should instead be thinking this way:
"Even people working physical jobs can use AI to improve their productivity, considering transformers are being used for robotics now."
Governments should make sure that everyone takes advantage of the new technology so we can deploy it as fast as possible. They should make sure everyone understands that using AI is a requirement to survive in 2023, and we will all need to radically change our workflows.
I've always read about UBI in the context of a national system. Any speculation on how UBI works with an international world order?
What happens to the countries without AI companies?
It's too radical of a change to know how well it would work. Communism, in theory, worked as well too but in practice there were unanticipated side effects.
The purpose of a UBI isn't for workers whether they're unemployed by AI or not, it's for non-working household members like new parents, elderly and children. (Poverty is a household phenomenon, not an individual phenomenon.)
This is not true. It is historically true but every iteration of new technology changes society differently.
You can't blindly rely on historical data you need to use common sense and logic as well. Historical data says nuclear bombs killed less people then cars. Therefore cars are more deadly then nuclear bombs.
Does that work? No. It obviously doesn't. Just like you can't say that AI will cause a rise in employment. You're ignoring the common sense scenario. Can you even work out exactly what people will be employed as once AI takes over? Likely you don't, you're just being impractically data driven.
If the AI "takes over", as in it's a fully sentient agent, then claiming humans will be unemployed isn't technological unemployment but rather is the lump of labor fallacy. That's because the AIs are just more people. They're going to be customers (they need products to survive and probably want things on top of that), will have jobs (they need to pay their power and AWS bills or they'll get turned off), will negotiate pay raises that may make you cheaper, and you can sell things to them.
If they're not sentient/agentic enough to be consumerist or want to be paid, then they're capital equipment aka automation, which doesn't substitute for labor and in practice increases employment by making labor more productive.
All nice and theoretical.
> Therefore cars are more deadly then nuclear bombs.
I agree the #3 cause of death in America is more deadly than something that is so safe it has never accidentally killed anyone. (Except for the demon core guy and the people who got cancer from detonation tests.)
Remember, the real world actually exists, and logic problems don't, which means logic problems are free to be incorrect because nothing is requiring them to exist in possible worlds.
Not if the institutions can't handle the challenge of moderating price stability without decreasing demand, as they're doing right now. The alternatives would be to demur or tax. So that capital is relatively distributed. If capital isn't distributed, people won't be paying people to do stuff because they won't have capital.
It's a continuous version of suffrage. Markets are effectively a distributed consensus mechanism, but if voting power isn't widely distributed, people don't won't have the ability to allocate time and resources as needed. Authority becomes concentrated then into relatively few hands.
Can you imagine a technology comes along that could 1000x the productivity of all information workers and your first reaction is to reduce demand for labor instead of softening adjusting the distribution of capital so that demand can grow in a more healthily distributed way?
This claim was never made. The only claim I made is that you can't claim AI increases jobs. My opinion of it is that the situation is so unpredictable that people should prepare for both scenarios.
>That's because the AIs are just more people.
And unconvincing. If agi replaces labor it will be constructed with a singular incentive to serve it's master as the primary directive. It can be a an economy of slave labor replacing us. The situation is too untenable to just rely on your own hypothetical. Other scenarios like mine must be considered.
>which doesn't substitute for labor and in practice increases employment by making labor more productive.
Again over reliance on data. Your argument doesn't counter my claim. It's just regurgitation your original claim.
>Remember, the real world actually exists, and logic problems don't, which means logic problems are free to be incorrect because nothing is requiring them to exist in possible worlds.
So according to that logic we should make millions of atomic weapons and distribute it for free to everyone? No.
Data only tells a story for specific scenarios and can function as a rough predictor for hypothetical scenarios.
To refine predictions of hypothetical scenarios you need data AND logic. Blind application of data to all hypothetical scenarios is wrong.
Remember the real world only exists in the past. A huge portion of the world is an unrealized quantum uncertainty in the future. Navigating the future involves using data from the past as well as logic.
Your argument suffers mostly from being too specific. It doesn't address the fuzziness of the future and the fuzziness of the evidence/logic. You anticipate one thing playing out and you plan for that but realistically a good prediction and plan is one involving a multitude of likely scenarios.
Well, that's the capital equipment case then. Is it really a general intelligence if you don't let it be general? It can't handle every possible scenario for you that way.
Which makes it incredibly wasteful, just like slavery was, because we used humans as capital instead of participating in an economy with them. Actually this is the reason economics is called "the dismal science" - a slaveowner called it that because he was upset economists kept telling him not to own slaves.
If your theory is that someone will program an AGI for the specific goal of making you unemployed, sure that could happen. But it sounds more like warfare than economics to me at that point, so I can't comment cause I don't know too much about that.
> Again over reliance on data. Your argument doesn't counter my claim. It's just regurgitation your original claim.
Where's the data? That was pure theory there, I didn't make any empirical claims.
Here's some data though: https://noahpinion.substack.com/p/american-workers-need-lots...
> Your argument suffers mostly from being too specific.
It's important to do this because approximately all AGI doom on the internet comes from Big Yud, an internet cult founder who thinks you can predict the future by thinking about it really hard and doing probability theory in your head without needing to check any assumptions. Which is why he thinks AGI is going to enslave him, because it's on a computer so it can do math exponentially faster, never mind that it can't do real world experiments to test anything.
Humans are most useful to other humans as happy slaves. Autonomy only is good for humans as individuals. If we create useful agi of course we will bend it to our will. There is no incentive to make AI as a separate competitor other than an experiment. Either way this is hypothetical.
>If your theory is that someone will program an AGI for the specific goal of making you unemployed, sure that could happen. But it sounds more like warfare than economics to me at that point, so I can't comment cause I don't know too much about that.
No it's cost saving. I don't want to employ someone who I have to pay. This is capitalism. I save money by using AI.
>Where's the data? That was pure theory there, I didn't make any empirical claims.
You don't need to cite your sources. It's obvious you're getting your conclusions from data. You use the words "in practice". I agree with you on how the data plays out.
>It's important to do this because approximately all AGI doom on the internet comes from Big Yud, an internet cult founder who thinks you can predict the future by thinking about it really hard and doing probability theory in your head without needing to check any assumptions. Which is why he thinks AGI is going to enslave him, because it's on a computer so it can do math exponentially faster, never mind that it can't do real world experiments to test anything.
No it's not. AI doom saying comes from everywhere. See: https://www.youtube.com/watch?v=qpoRO378qRY (the godfather of modern AI, is concerned about doom). Either way, you shouldn't counter something with extremities that are simply opposite in nature. You should counter something you disagree with, with actual truth, Even if that truth doesn't fully dismiss the idea and calls only for naunce.
Edit- beyond that you make great points too
Generative language models have as their core capability the production of the most likely continuation for a text sequence. This seemingly simple skill is remarkably general. Any task that can be specified and executed via text can be framed as text continuation. This encompasses a wide range of cognitive tasks, including tasks that can be resolved over chat or email, for example, or in a web forum.
This is Goldman Sachs, so take it with a big grain of salt. In 2017, they predicted huge job losses from self-driving cars:
Self-driving cars could cost America’s professional drivers up to 25,000 jobs a month, Goldman Sachs says
>> When autonomous vehicle saturation peaks, U.S. drivers could see job losses at a rate of 25,000 a month, or 300,000 a year, according to a report from Goldman Sachs Economics Research.
>> The report estimates that semi- and fully autonomous car sales will have about 20 percent share of car sales around 2025 to 2030.
https://www.cnbc.com/2017/05/22/goldman-sachs-analysis-of-au...
If it's from Goldman Sachs, and it's a prediction about a technology making a smash on social media, you can rest assured it's speculation with the hope of influencing the markets to make an investment so that Goldman Sachs can profit.
Sometimes people under-estimate how 'unagile' most business is. Are we talking months, years, or decades before we see this "new class of dispossessed white-collar workers, who risk suffering a similar fate to that of manufacturing workers in the 1980s."
On the other hand: in order to be able to use current LLMs effectively, you have to understand their output well enough to debug it, even if you are using the LLM to write the fix to its own bug.
I’ve done something along the lines of “write me a python script which converts midi files into this JSON format I’m giving you as an example: …” — a few runtime crashes it was able to fix just from pasting in the command line output, but fixing “all pitches and durations are zero” required me to get my hands dirty and decompose the problem into a chunk small enough for chatGPT to fix.
I don’t know how far we are from any AI, LLM or otherwise, not needing or benefiting from human oversight, except in the specific domains like chess where it has already happened. I don’t believe in saying “never” to stuff like this, as that’s a common error with many technologies from flight to superconductors not just AI, but I also notice my personal guesses err on the side of being too near in our future.
The question is where do the gains go? For some jobs, AI becomes a tool that the professional buys and uses to make themselves more productive, and he pockets the difference. If I'm writing apps for a living and I can do it in 6 hours a day, and copilot makes it 5 hours, I'm going to pocket that hour and use it on my family. If I'm a web app company and I can just replace the devs with GPT-5, I'm going to buy it and can all the app devs, and pocket the difference.
I've observed many sites implementing Ai -written articles lately, ad frankly the quality of writing is terrible... If this is the future, we're in big trouble.
There is a degree of soullessness found in Ai that will probably never be captured. Being isolated with only Ai as an option for interaction destroys mental health, as well will find out once it's too late. Our propensity to rush into profit-seeking/marketing-driven technology schemes is destroying the best parts of our real world with filler and emptiness.
I hope innovators take a moment to step back and realize rushing to market will do more damage than ever, and that unique products (not another image generator or text-to-ai subscription service) are better than whatever is going on now. We're light years away from productivity and authenticity in Ai, and it's evident when even Siri and Google Voice still cannot properly understand cues without them being carefully pronounced and phrased. Let's all stop trying to fool everyone about Ai with cheap scripting and create some real tech instead, which takes time and proper testing.
Seriously? These are comparable things?
A lot of people point out problems with crypto in their early days. Most people say NFTs are a scam since day 1. And they have those problems and are a scam. Generative AI, on the other hand, has been proven to provide real value, and will only get better. And I don't know what "failed promises" you are talking about -- NFTs were doomed the first day they appear and we already knew all the hype was fake, while nobody has actually "promised" what generative AIs can do, most are still speculations at this point.
The question is whether we remain in the driving seat and as OP said, how the gains will be distributed.
We could be working far fewer hours, we could have higher pay, but neither are realized because we can't shake off the goddamned aristocrats and neo feudalists.
Most people are not helping with AI or with tech in general. Most people are not working to develop AI nor are they attempting to integrate AI into their workflow. If they don't adapt, they won't benefit and rightly so. But people who are actually helping will be rewarded proportionately.
Fewer hours and higher pay? It's here. We've all heard of comfy tech jobs, writing 10 lines of code a day working from home for $250k. That's mostly people who use tech like Python and JavaScript. Soon people will be doing better than that with generative AI. 4-day workweeks are around the corner.
McKinsey estimated that the top 5% of employees produce 95% of the value. These are not capitalists, they are technologically sophisticated workers. The media likes to report "productivity is increasing, wages are not" but the median worker isn't any more productive than 20 years ago. The people actually driving productivity increases are rewarded fairly, their wages are going up rapidly.
Corpos need a wealthy consumer base in order to sell things and remain profitable. But to do that they have to collectively pay laborers higher wages because laborers are basically the consumer.
An individual corporation lowering labor costs in an economy of high wages benefits with greater profits so this strategy is logical. But if corporations act on this collectively they damage themselves. Most corporations, however are built upon short term gains. So they are unlikely to have the foresight to act wisely. It's the tragedy of the commons.
The tragedy is this, when a corporation hires AI and lowers labor costs he is acting rationally according to a cost benefit analysis. But this logic applied collectively becomes suicidal. It is ultimately rational and logical action that damages us. That it the tragedy.
Best solution I've heard of is ubi. But I'm not sure how it will play out in practice. I think government job creation might be better.
Maybe salaries will take a nosedive? Minimum wage for a junior developer, anyone? Or maybe the people just won't accept it, and boycott companies. If 300M people don't have jobs the economy will certainly stall.
The tech doesn't understand logic and I think most knowledge workers need that.
I can also see nigerian spam getting much more refined before long....
I'm not really saying this time will be different, but I don't think self driving cars caused even this much disruption.
image -> text -> code -> self-driving vehicles
It's really reductive to just put all code in a single bucket, but most code probably isn't super critical for safety and it can be tested before it gets released/reviewed. I think that's the biggest differentiator from real time applications like self driving cars.
I wholly disagree with this. The rate at which AI can produce pornography that caters to exact tastes is completely unmatched, I have seen people report their compulsive pornography habits becoming exponentially worse as a result of Stable Diffusion in particular. The only hold out has been video pornography, but with things like txt2vid improving I can't see that lasting long (pun not intentional).
Instead, think of machine services: fully-automated ones like "ads" or semi-automated like news stories, records requests, customer service replies, retail environment/shopping or other high-volume, low-profit trivial interactions.. These services have made billions in profit (ads) or removed billions in costs (employees) in the last twenty years exactly because they are done with minimal "jobs" specifically and by design.
Modern Capital markets have found the drug-combo they can't kick.. make money with machines with no "jobs" .. meanwhile, news outlets of all kinds continue in a death-spiral.
So combine the angst of media people themselves, the aggressive New Capital making machines that provide services while making few to no "jobs", and generally population anxiety over say .. catastrophic social conditions for example.. that is the real content from this chair.
maybe, instead of being real news about "jobs" it is instead, red-meat bait for the people who own capital who want to get into the "making machines that make money without making jobs" game.. Sort of like a modern version of really-attractive but utterly socially un-discussable in public.. starting wars for money, or building super people through breeding, to name some other obvious ones..
And even if military capitalists profited from a war, the other 99% of capitalists in the country would find it a distraction at best.
Generating an app with a maintainable architecture and ongoing development over 5 to 10 years is another. AI will assist developers but never replace them.
For a more up-to-date example, electric bikes. A Whole Foods delivery guy who rides one of those yellow electric bikes with a little trailer behind can easily deliver as much as 3 people pushing boxes on dollies. Which is the same thing as saying each such bicycle has resulted in two people losing their jobs.
Also a large chunk of manager work is grunt work like filling spreadsheets, writing emails, scheduling meetings, etc… With those tasks partially automated, I am pretty sure a manager could oversee larger teams and focus on the big picture instead of doing boilerplate tasks.
In fact if you downsize to keep doing what you’re doing and your competitor keeps their team size, wouldn’t they overwhelm you?
Can you beat them on price? Maybe but software prices are not affected by labor costs really so that’s not a differentiator.
Do you see both classes being affected the same?
One thing to remember is managers line to have as many people under them as possible. VPs and directors care a lot about this. They’ll find ways to continue to grow teams whine harvesting more productivity.
We’ve see a gaming company install a bot as a CEO and its stock has gone up.
I imagine management LLMs will be a layer in development.