So they studied change in "headcount" (full-time, part-time?), but what are these heads doing? A more interesting metric to study would be changes in wages for a normalized position (e.g. 40h/week, 4w PTO/y). Then you could disentangle the different effects of actually hiring a highly paid worker or replacing a highly paid worker with two new assistants + an AI subscription for less total compensation. AI is just the latest stage of a long trend in automation. If you check the first chart in the recent article of the NY FED [1] you'll see that the labor share has been on a decline for decades, and the trend hints at an accelerating decline. In light of this larger trend, I wouldn't expect that the wages of the "increased headcount" according to ramp's article is growing or holding ground.
What is the value of saying this when ChatGPT was released in 2022? There’s no way there’s enough data to make any meaningful extrapolation about anything.
Bingo. Bullshit work grows to accommodate productivity gains.
Is it easier to implement a design now? Then let’s redesign every quarter.
Is it easier to refactor the whole codebase? Then let’s rewrite in whatever new hotness.
This is not different from web frameworks 10-20 years ago. Is it easier to make a website using Rails? If so, why did I see an explosion in devs-per-project after I started using Rails? Because of the BS.
We can’t really say what this means. Maybe heavy AI use caused companies to perform better and grow. Or maybe the type of company that went big into AI was the type of company that was getting investments and growing because of it?
Skeptical of the study approach, but it is an entirely possible scenario if we don’t hit AGI. Many of the non-tech companies are only now starting to invest into AI. For many of them, AI might unlock new business approaches. Eg a company switches to real time pricing. Great, but suddenly you need people swapping price tags. Or you launch a project to replace paper price tags with electronic tags. Those need to be regularly maintained and charged, etc. Oops, our legacy ERP system can’t deal with that. Let’s upgrade that
Seems like a questionable correlation. The companies with money to burn on AI are those those that are growing and doing well so ofc their growing headcount too
One of the study authors, Ara, tweeted that they control for that by comparing early adopters against firms who haven't yet, and built like-for-like control groups with similar pre-adoption growth trajectories. (This is a rough quote).
Companies over-hired in 2021 assuming that their COVID metrics explosion would persist, but they didn't. We saw low attrition for years, and a low-hire/low-fire employment environment. Then the time came to pay the piper.
The tech everything's-hypergrowth era is, for now, over. Most of the low-hanging fruit that we've collected for the last 20 years of bull market (putting tech into every business) is gone, and companies hired in advance of having the next business to overlay on their current ones. For many that didn't materialize.
I am confused. Wasn't the initial claim that AI kills all jobs?
I feel the claim right now is not really correct. One needs to do a thorough analysis of the whole job market across different countries, say, over 5 years. Or at the least 3 years but very complete and unbiased either way.
The claim is and always has been that automation increases net jobs. There's never been an automation revolution that worked otherwise.
The counter claim that LLMs were special for some reason has never been supported. And a lot of its proponents have other axes to grind, like UBI, Socialism, Communism etc. It would suit their worldview for this time to be different, and so thats the message they push.
I'm not entirely sure that's the claim or conclusion. Like they said the companies increasing headcount are also being pumped full of $$$. That's a non-permanent state of affairs because it's heavily fueled by speculation about possible future scenarios which may or may not come to pass.
I think the bigger conclusion is that LLMs are, for now, having a minimal effect on the labor market. And I think this makes sense. In spite of individual claims of 10x productivity boost or whatever, their effect on the bottom line of companies seems to remain quite unclear, at best. On the contrary the token-maxing catastrophe seems to have resulted in companies becoming far more price conscious towards LLM usage.
That’s not supported by the evidence. Current rates of employment are well below historic levels.
In preindustrial societies nearly everyone was working to support themselves and their families, these days a huge percentage of the population is in education, retirement, prison, disability, etc.
The only way jobs have kept up with automation is when you ignore population growth, but more people naturally increases the required workforce. You inherently need more police, food, etc when you have more people.
>Current rates of employment are well below historic levels.
Ignoring for a second that net jobs and employment rates aren't really the same thing.
>In preindustrial societies nearly everyone was working to support themselves and their families, these days a huge percentage of the population is in education which may be investing in their future but isn’t directly producing anything. Retirement as a percentage of the population similarly exploded etc.
The first thing anyone does in employment statistics is remove non participants. Bringing them back in is weird. If you don't need or want a job its kind of a non sequitur to be lumped in with the employment seeking population. AI doomers aren't suggesting that its going to gainfully retire the population.
>preindustrial societies
Pre industrial societies can be loosely grouped into "People farming to make 3-5 times the food they need" and "city dwellers". Now that a single person can farm for 100s of people, we do have hundreds more city jobs going. We dont have a huge number of out of work farmers sitting around doing nothing. Likewise, Banks employ more people after introducing ATM's than before. Likewise cloud didnt leave IT people lining up at the dole office, but it just moved them from cleaning up on prem messes to cleaning up cloud messes and onprem messes.
> Banks employ more people after introducing ATM's than before.
Not as a percentage of the population. In 1967 the US population in 1967 was 200 million today it’s 350 million. Globally the growth was even faster.
Your farming productivity numbers are also very exaggerated. The actual percentage of the population was very high because of food spillage, and inconsistent productivity due to weather etc. They often needed ~90% of the adult population to be farmers or people starved, and famine was still common. Right before industrialization that number started to drop which is largely what caused industrialization. You got a positive feedback loop of increased productivity when a vastly larger percentage of the population could do something other than produce food.
Not in your post, if you did you could quote an actual rebuttal that larger populations don’t inherently need more food, increase the need for banking services, etc etc.
You also found it odd I didn’t “remove non participants” but preindustrial societies didn’t necessarily have formal jobs for people. You’ll note I didn’t exclude modern stay at home parents as non workers because labor is the metric not if you have formal employment.
The big 4 Australian banks (Commonwealth Bank, ANZ, National Australia Bank and Westpac) all employ pretty much the same number of people now as they did 25 years ago, but their workforce composition has changed. More software developers, fewer branch staff.
I really hate the (very australian) propensity to take an argument about X, and make it about X as a proportion of Y, pretending they are exactly the same statistic.
LNP: Taxes less.
Labor: Taxes less (as a proportion of GDP)
(Seeing as GDP includes government spending, it seems like you can make the rate of tax go down by increasing government spending by more than taxes increase)
If a claim is about X, introducing another variable doesn't invalidate the claim about X.
> The counter claim that LLMs were special for some reason has never been supported. And a lot of its proponents have other axes to grind, like UBI, Socialism, Communism etc.
No, they are mostly right wing capitalists. LLM is not a left wing project. UBI is something they talked about, but super hard to believe it is anything but a way to shut up critics.
> The claim is and always has been that automation increases net jobs.
That was definitely not the common claim last 4 years.
> There's never been an automation revolution that worked otherwise.
That is not what historical record shows. In the short term, larger unemployment are to be expected. Over long term, it evens out.
>No, they are mostly right wing capitalists. LLM is not a left wing project. UBI is something they talked about, but super hard to believe it is anything but a way to shut up critics.
The proponents of the claim not the proponents of LLMs.
These studies miss a huge point, which is: working with automation outside of programming is tedious, frustrating, and soul crushing.
I see it already with some non technical clients. I have notice a terrible trend of people just offloading everything to agents or prompting even the simplest interaction with other humans. Let me tell you, it’s fucking dire.
Inside can be as well. Please manually review these 5000 lines of slop I generated while watching YouTube and verify the unreadable readme due to all the bloat and icons is correct with the code.
If that is not soul crushing I would not know what is.
"Look at all the traditional signals of thought, quality, or attention that it inserted, which being totally counterfeit have actually become strictly counterproductive--at least to anybody who has to actually use it!"
Yesterday a fellow agency who’s sharing a client with me proceeded to send me a roughly 1000-thousand words LLM-generated message on slack explaining a strange behavior on a newsletter flow that did the following:
- explained back to me the original request I had filed with details
- assumed the request was about something it was not
- didn’t reply to my question
- had no actionable element whatsoever
- repeated all of the above twice with different words.
This is a senior person running a marketing agency.
The absurdity here is that I have a high cost per hour that falls back onto our shared client. Who I will now have to brief about how I spent about 2 hours due to this comms back and forth and deciphering LLM-novels on a problem that could have been solved in a 30min slot if the agency only took 5 minutes to read, understand and use their brain.
The added time was for me to basically go back to my comms, understand if I had communicated poorly (I asked a non-tech third party: they got it in 3 minutes), and reassess everything twice.
Result: two hours to do something that amounted to changing a link behavior.
So they studied change in "headcount" (full-time, part-time?), but what are these heads doing? A more interesting metric to study would be changes in wages for a normalized position (e.g. 40h/week, 4w PTO/y). Then you could disentangle the different effects of actually hiring a highly paid worker or replacing a highly paid worker with two new assistants + an AI subscription for less total compensation. AI is just the latest stage of a long trend in automation. If you check the first chart in the recent article of the NY FED [1] you'll see that the labor share has been on a decline for decades, and the trend hints at an accelerating decline. In light of this larger trend, I wouldn't expect that the wages of the "increased headcount" according to ramp's article is growing or holding ground.
[1]: https://libertystreeteconomics.newyorkfed.org/2026/06/the-po...
What is the value of saying this when ChatGPT was released in 2022? There’s no way there’s enough data to make any meaningful extrapolation about anything.
To get onto Hacker News and other social media sites for marketing exposure?
The current narrative is that AI is displacing jobs.
Similar to the current post there’s just not enough data to be conclusive.
Current narrative where?
I don't see this narrative being more prominent than the "AI won't be displacing jobs because things worked out in the past" counternarrative.
For example, only recently has The Economist published articles that go more towards that narrative (AFAIK) [1].
> there’s just not enough data to be conclusive.
With this, I can agree. We can still extrapolate based on what we currently know.
[1] https://www.economist.com/leaders/2026/05/14/prepare-for-an-...
[dead]
Because people are wasting time and tokens doing irrelevant to the bottom line things.
Cool you spent 2 days and $200 building a react UI for your spreadsheet.
Bingo. Bullshit work grows to accommodate productivity gains.
Is it easier to implement a design now? Then let’s redesign every quarter.
Is it easier to refactor the whole codebase? Then let’s rewrite in whatever new hotness.
This is not different from web frameworks 10-20 years ago. Is it easier to make a website using Rails? If so, why did I see an explosion in devs-per-project after I started using Rails? Because of the BS.
We can’t really say what this means. Maybe heavy AI use caused companies to perform better and grow. Or maybe the type of company that went big into AI was the type of company that was getting investments and growing because of it?
Skeptical of the study approach, but it is an entirely possible scenario if we don’t hit AGI. Many of the non-tech companies are only now starting to invest into AI. For many of them, AI might unlock new business approaches. Eg a company switches to real time pricing. Great, but suddenly you need people swapping price tags. Or you launch a project to replace paper price tags with electronic tags. Those need to be regularly maintained and charged, etc. Oops, our legacy ERP system can’t deal with that. Let’s upgrade that
[flagged]
That’s because Aldi /did/ the upgrade…?
Seems like a questionable correlation. The companies with money to burn on AI are those those that are growing and doing well so ofc their growing headcount too
One of the study authors, Ara, tweeted that they control for that by comparing early adopters against firms who haven't yet, and built like-for-like control groups with similar pre-adoption growth trajectories. (This is a rough quote).
Fair enough
That's interesting and goes directly against what everyone complains about in the employment sector.
I'm sure there is a big caveat in there somewhere.
Companies over-hired in 2021 assuming that their COVID metrics explosion would persist, but they didn't. We saw low attrition for years, and a low-hire/low-fire employment environment. Then the time came to pay the piper.
The tech everything's-hypergrowth era is, for now, over. Most of the low-hanging fruit that we've collected for the last 20 years of bull market (putting tech into every business) is gone, and companies hired in advance of having the next business to overlay on their current ones. For many that didn't materialize.
"Oh yeah, we had to hire a bunch of guys to clean up the slop"
>AI adoption and the associated gains are unevenly distributed.
is that more or less than regular headcount growth?
It's not entirely clear to me, but I believe this is 10% over the control. The 10% is therefore relative to "regular headcount growth".
So, more skynet leads to more real jobs?
I am confused. Wasn't the initial claim that AI kills all jobs?
I feel the claim right now is not really correct. One needs to do a thorough analysis of the whole job market across different countries, say, over 5 years. Or at the least 3 years but very complete and unbiased either way.
The claim is and always has been that automation increases net jobs. There's never been an automation revolution that worked otherwise.
The counter claim that LLMs were special for some reason has never been supported. And a lot of its proponents have other axes to grind, like UBI, Socialism, Communism etc. It would suit their worldview for this time to be different, and so thats the message they push.
I'm not entirely sure that's the claim or conclusion. Like they said the companies increasing headcount are also being pumped full of $$$. That's a non-permanent state of affairs because it's heavily fueled by speculation about possible future scenarios which may or may not come to pass.
I think the bigger conclusion is that LLMs are, for now, having a minimal effect on the labor market. And I think this makes sense. In spite of individual claims of 10x productivity boost or whatever, their effect on the bottom line of companies seems to remain quite unclear, at best. On the contrary the token-maxing catastrophe seems to have resulted in companies becoming far more price conscious towards LLM usage.
I't has been very good as suppressing wages.
That’s not supported by the evidence. Current rates of employment are well below historic levels.
In preindustrial societies nearly everyone was working to support themselves and their families, these days a huge percentage of the population is in education, retirement, prison, disability, etc.
The only way jobs have kept up with automation is when you ignore population growth, but more people naturally increases the required workforce. You inherently need more police, food, etc when you have more people.
> That’s not supported by the evidence. Current rates of employment are well below historic levels.
What evidence?
Wasn’t employment in engineering down before all this could’ve gotten enough widespread traction to actually displace jobs?
Programming/engineering may be of supreme concern to you, but they made a much wider claim.
> There's never been an automation revolution that worked otherwise.
Requires all automation to have zero effect on net jobs long term. Meanwhile the percentage of people working is down so that’s clearly not true.
>Current rates of employment are well below historic levels.
Ignoring for a second that net jobs and employment rates aren't really the same thing.
>In preindustrial societies nearly everyone was working to support themselves and their families, these days a huge percentage of the population is in education which may be investing in their future but isn’t directly producing anything. Retirement as a percentage of the population similarly exploded etc.
The first thing anyone does in employment statistics is remove non participants. Bringing them back in is weird. If you don't need or want a job its kind of a non sequitur to be lumped in with the employment seeking population. AI doomers aren't suggesting that its going to gainfully retire the population.
>preindustrial societies
Pre industrial societies can be loosely grouped into "People farming to make 3-5 times the food they need" and "city dwellers". Now that a single person can farm for 100s of people, we do have hundreds more city jobs going. We dont have a huge number of out of work farmers sitting around doing nothing. Likewise, Banks employ more people after introducing ATM's than before. Likewise cloud didnt leave IT people lining up at the dole office, but it just moved them from cleaning up on prem messes to cleaning up cloud messes and onprem messes.
> Banks employ more people after introducing ATM's than before.
Not as a percentage of the population. In 1967 the US population in 1967 was 200 million today it’s 350 million. Globally the growth was even faster.
Your farming productivity numbers are also very exaggerated. The actual percentage of the population was very high because of food spillage, and inconsistent productivity due to weather etc. They often needed ~90% of the adult population to be farmers or people starved, and famine was still common. Right before industrialization that number started to drop which is largely what caused industrialization. You got a positive feedback loop of increased productivity when a vastly larger percentage of the population could do something other than produce food.
>Not as a percentage of the population. In 1967 the US population in 1967 was 200 million today it’s 350 million. Globally the growth was even faster.
Didnt I address that explicitly?
Not in your post, if you did you could quote an actual rebuttal that larger populations don’t inherently need more food, increase the need for banking services, etc etc.
You also found it odd I didn’t “remove non participants” but preindustrial societies didn’t necessarily have formal jobs for people. You’ll note I didn’t exclude modern stay at home parents as non workers because labor is the metric not if you have formal employment.
> Likewise, Banks employ more people after introducing ATM's than before.
that's a common trope, and its both true and false
true: more bank employees after ATM's
false: less bank employees after smartphone banking
The big 4 Australian banks (Commonwealth Bank, ANZ, National Australia Bank and Westpac) all employ pretty much the same number of people now as they did 25 years ago, but their workforce composition has changed. More software developers, fewer branch staff.
~20% outsorced workforce in places like India, Vietnam and Philipines nowadays
I’m only counting the Australian FTE. The numbers are up if you count the outsourcing.
The Australian population size isn’t pretty much the same as it was 25 years ago.
28 million in 2026 vs 18.8 million in 2001 is a ~50% jump…
So no the numbers are down significantly even with outsourcing.
I really hate the (very australian) propensity to take an argument about X, and make it about X as a proportion of Y, pretending they are exactly the same statistic.
LNP: Taxes less.
Labor: Taxes less (as a proportion of GDP)
(Seeing as GDP includes government spending, it seems like you can make the rate of tax go down by increasing government spending by more than taxes increase)
If a claim is about X, introducing another variable doesn't invalidate the claim about X.
An outsourced job is still a job.
Its just for the government to rebalance the economy if those jobs are desirable to reshore.
> The counter claim that LLMs were special for some reason has never been supported. And a lot of its proponents have other axes to grind, like UBI, Socialism, Communism etc.
No, they are mostly right wing capitalists. LLM is not a left wing project. UBI is something they talked about, but super hard to believe it is anything but a way to shut up critics.
> The claim is and always has been that automation increases net jobs.
That was definitely not the common claim last 4 years.
> There's never been an automation revolution that worked otherwise.
That is not what historical record shows. In the short term, larger unemployment are to be expected. Over long term, it evens out.
> That is not what historical record shows
It's not the first industrial revolution, we've been in a good dozen of them now and absolutely none have resulted in large unemployment.
The only large unemployment we've experienced until now were linked to economic crisis which had nothing to do with industrial revolutions
>No, they are mostly right wing capitalists. LLM is not a left wing project. UBI is something they talked about, but super hard to believe it is anything but a way to shut up critics.
The proponents of the claim not the proponents of LLMs.
there never has been a nuclear war, so why do some people worry about it, I dont understand, probably they have some axes to grind
Too soon.
These studies miss a huge point, which is: working with automation outside of programming is tedious, frustrating, and soul crushing. I see it already with some non technical clients. I have notice a terrible trend of people just offloading everything to agents or prompting even the simplest interaction with other humans. Let me tell you, it’s fucking dire.
Inside can be as well. Please manually review these 5000 lines of slop I generated while watching YouTube and verify the unreadable readme due to all the bloat and icons is correct with the code.
If that is not soul crushing I would not know what is.
> bloat
"Look at all the traditional signals of thought, quality, or attention that it inserted, which being totally counterfeit have actually become strictly counterproductive--at least to anybody who has to actually use it!"
Yesterday a fellow agency who’s sharing a client with me proceeded to send me a roughly 1000-thousand words LLM-generated message on slack explaining a strange behavior on a newsletter flow that did the following:
- explained back to me the original request I had filed with details
- assumed the request was about something it was not
- didn’t reply to my question
- had no actionable element whatsoever
- repeated all of the above twice with different words.
This is a senior person running a marketing agency. The absurdity here is that I have a high cost per hour that falls back onto our shared client. Who I will now have to brief about how I spent about 2 hours due to this comms back and forth and deciphering LLM-novels on a problem that could have been solved in a 30min slot if the agency only took 5 minutes to read, understand and use their brain. The added time was for me to basically go back to my comms, understand if I had communicated poorly (I asked a non-tech third party: they got it in 3 minutes), and reassess everything twice. Result: two hours to do something that amounted to changing a link behavior.
[flagged]
[flagged]
[dead]
The human teams that were there doing CMS translations, or doing image assets, are no more...