The US has about 150 million workers. How long did it take them to learn to do what they do? I'm interested in this because of the analogous question for AI. Running an AI on a job and training it to do the job are different costs. In a recent post I found that when an AI can do a task, running it is far cheaper than paying a human, while training it looks about as expensive as the human's wages over the time they took to learn. So it is at least conceivable that we end up in a world where we can afford to run AI on every job but cannot afford to train it on every job. To know whether that is a real worry, you need to know how much learning there is to do.
Start with formal education. Every US worker spends over a decade in primary and secondary school learning basic skills like English and arithmetic, and many spend further years in college, professional programs, apprenticeships, and so forth. There are only so many of these programs, and in Appendix A I count them all and estimate how long each takes. It would take about 7 million hours, or 3,500 full-time years, for one person to complete every one of them.
This is clearly a substantial underestimate, because a great deal of what people know about their jobs is learned on the job. There are two measures of how much. The BLS Occupational Requirements Survey asks employers the minimum amount of training a worker needs after being hired, excluding orientation, and finds a mean of 28 days. O*NET asks the people actually doing each job how much on-the-job training a new employee needs to perform it as they do, and finds a mean of about 8 months. Given there are 150 million workers in the US, if each of them learned their job separately, that would be hours on the first measure and on the second.
These estimates assume that every worker's job is unique, which isn't true. Many jobs are literally identical, with several people doing the same work at the same site. The BLS's establishment data lets us count how many distinct positions there actually are, meaning an occupation at a particular site, and as I show in Appendix B the answer is about 60 million. Learning each of those once, at 28 days or 8 months apiece, is to hours.
This is still an overestimate. Even when two positions are distinct, they are often largely the same job: someone moving between two Walgreens has to learn the store layout, but already knows the work. It is hard to estimate how large this effect is, since the surveys above measure new hires, and no real worker arrives already knowing a thousand other Walgreens. There is a further effect of the same kind: someone who had somehow already learned half of all work would find much of the rest familiar, so the later jobs would take less to learn than the earlier ones. Neither effect can be sized from the data here; I'd guess the time required for one person to learn every job in America is somewhere between and hours.
So far I have counted formal education and on-the-job training, but in many professions people keep improving over decades of experience. I haven't tried to estimate how much this adds, but my sense is that it is not much larger than formal education, for two reasons. First, most workers are not in professions where experience adds much beyond formal training and the first few years on the job. Second, professional experience is substantially fungible. A software engineer with twenty years of experience is valuable for the twenty years, but the specifics of those years don't matter much; most of the value is in general skills that most software engineers with twenty years have also acquired.
What does this mean? We can think in terms of training costs: how expensive would it be to train AI models with the knowledge to perform every job? In that post I estimated that current systems learn skills at around FLOP per second of human learning time. At that rate, to hours of learning is to FLOP. This is about a hundred times more compute than is used in the current largest training runs. At current hardware prices, this would cost between $10 and $100 billion, a one-off cost of well under one percent of a single year of US wages.
Compute costs are not the only issue: one also needs the right data and algorithms. But my guess is that data acquisition is overrated as a problem. to hours of learning sounds like a lot, but it's a similar magnitude to all the video on YouTube, or a day of the world's installed camera capacity. Collecting it would also not be that difficult; at worst, companies could stick cameras and microphones on their workers.
Current training methods are much less sample efficient than human learning. But models already know far more than any person does, and per FLOP they do not seem obviously worse than the brain at learning facts. To the extent models are less efficient at learning skills and tacit knowledge, humans still provide a lower bound on possible efficiency.
Of course, none of this accounts for continued learning on the job. Most jobs change over time, and an AI that had learned every job would still have to keep learning. I don't think this adds much to the estimates. But it may not fit the current paradigm, in which a model is trained once and then acts: if what changes can't be handled in context, continual learning could be more of a qualitative barrier to widespread deployment.
Appendix A: learning all of formal education
The US has a limited number of high school courses, college majors, professional degrees, medical specialties, doctoral fields, and apprenticeships. If one person did all of them, how long would it take?
For each level I count the credentials, multiply by the hours each takes, and add up. Hours are the full hours a student puts in, including study outside class. A year of study is 1,200 hours, from the time-use data in Babcock and Marks (2011); a year of residency or fellowship is 3,500 hours, from the IOM (2009) survey of duty hours; and a year of research or apprenticeship is a 2,000-hour work year.
All of formal education, then, comes to about hours, or about 4,800 FTE years, which is roughly 100 working lifetimes. This includes some overlapping, for example the general-education year inside college majors or shared coursework in related doctorates. Removing the easily identifiable overlaps cuts the total by 30% to around hours, or 3,500 FTE years.
Appendix B: the number of distinct positions
The Quarterly Census of Employment and Wages counts all 12 million establishments covered by unemployment insurance, where each establishment is a single physical location such as a store, a plant, or an office. The Occupational Employment and Wage Statistics survey reports, for each industry, what share of establishments employ each occupation. Summing over occupations gives the average number of distinct occupations per establishment. Multiplying by the number of establishments in the industry and adding up over industries gives the number of distinct occupation-at-site positions in the country:
Industry
Occupations per establishment
Law offices
2.8
Restaurants
5.9
Junior colleges
30
General hospitals
37
All industries, weighted by establishments
5.3
An average 5.3 occupations per establishment across 12 million establishments gives a total of 63 million distinct positions.
I'm unsure what you mean, I'm already assuming that all learning is shared across instances. There's a separate question of transfer learning between different tasks, which I expect probably somewhat reduces the totals, but the effect size is harder to estimate.
It looks like your math is assuming that different coursework, majors, residencies, etc. are NOT shared in any way. The AI would need to take all of them independently. This is incredibly untrue. Even in humans, a double-major doesn't double the time taken, and for an AI there's probably only a few % of uniqueness across a lot of fields.
These estimates assume that every worker's job is unique, which isn't true. Many jobs are literally identical, with several people doing the same work at the same site.
Another effect to consider: the identical jobs will get automated first, and the resulting economic growth may lead to the creation of higher-skilled jobs (see: Lump of labour fallacy) that require even more training / specialization and are thus harder to automate than the jobs they replaced. This might systematically drive up the time to full labour automation and throw off our predictions.
The US has about 150 million workers. How long did it take them to learn to do what they do? I'm interested in this because of the analogous question for AI. Running an AI on a job and training it to do the job are different costs. In a recent post I found that when an AI can do a task, running it is far cheaper than paying a human, while training it looks about as expensive as the human's wages over the time they took to learn. So it is at least conceivable that we end up in a world where we can afford to run AI on every job but cannot afford to train it on every job. To know whether that is a real worry, you need to know how much learning there is to do.
Start with formal education. Every US worker spends over a decade in primary and secondary school learning basic skills like English and arithmetic, and many spend further years in college, professional programs, apprenticeships, and so forth. There are only so many of these programs, and in Appendix A I count them all and estimate how long each takes. It would take about 7 million hours, or 3,500 full-time years, for one person to complete every one of them.
This is clearly a substantial underestimate, because a great deal of what people know about their jobs is learned on the job. There are two measures of how much. The BLS Occupational Requirements Survey asks employers the minimum amount of training a worker needs after being hired, excluding orientation, and finds a mean of 28 days. O*NET asks the people actually doing each job how much on-the-job training a new employee needs to perform it as they do, and finds a mean of about 8 months. Given there are 150 million workers in the US, if each of them learned their job separately, that would be
These estimates assume that every worker's job is unique, which isn't true. Many jobs are literally identical, with several people doing the same work at the same site. The BLS's establishment data lets us count how many distinct positions there actually are, meaning an occupation at a particular site, and as I show in Appendix B the answer is about 60 million. Learning each of those once, at 28 days or 8 months apiece, is
This is still an overestimate. Even when two positions are distinct, they are often largely the same job: someone moving between two Walgreens has to learn the store layout, but already knows the work. It is hard to estimate how large this effect is, since the surveys above measure new hires, and no real worker arrives already knowing a thousand other Walgreens. There is a further effect of the same kind: someone who had somehow already learned half of all work would find much of the rest familiar, so the later jobs would take less to learn than the earlier ones. Neither effect can be sized from the data here; I'd guess the time required for one person to learn every job in America is somewhere between
So far I have counted formal education and on-the-job training, but in many professions people keep improving over decades of experience. I haven't tried to estimate how much this adds, but my sense is that it is not much larger than formal education, for two reasons. First, most workers are not in professions where experience adds much beyond formal training and the first few years on the job. Second, professional experience is substantially fungible. A software engineer with twenty years of experience is valuable for the twenty years, but the specifics of those years don't matter much; most of the value is in general skills that most software engineers with twenty years have also acquired.
What does this mean? We can think in terms of training costs: how expensive would it be to train AI models with the knowledge to perform every job? In that post I estimated that current systems learn skills at around
Compute costs are not the only issue: one also needs the right data and algorithms. But my guess is that data acquisition is overrated as a problem.
Current training methods are much less sample efficient than human learning. But models already know far more than any person does, and per FLOP they do not seem obviously worse than the brain at learning facts. To the extent models are less efficient at learning skills and tacit knowledge, humans still provide a lower bound on possible efficiency.
Of course, none of this accounts for continued learning on the job. Most jobs change over time, and an AI that had learned every job would still have to keep learning. I don't think this adds much to the estimates. But it may not fit the current paradigm, in which a model is trained once and then acts: if what changes can't be handled in context, continual learning could be more of a qualitative barrier to widespread deployment.
Appendix A: learning all of formal education
The US has a limited number of high school courses, college majors, professional degrees, medical specialties, doctoral fields, and apprenticeships. If one person did all of them, how long would it take?
For each level I count the credentials, multiply by the hours each takes, and add up. Hours are the full hours a student puts in, including study outside class. A year of study is 1,200 hours, from the time-use data in Babcock and Marks (2011); a year of residency or fellowship is 3,500 hours, from the IOM (2009) survey of duty hours; and a year of research or apprenticeship is a 2,000-hour work year.
All of formal education, then, comes to about
Appendix B: the number of distinct positions
The Quarterly Census of Employment and Wages counts all 12 million establishments covered by unemployment insurance, where each establishment is a single physical location such as a store, a plant, or an office. The Occupational Employment and Wage Statistics survey reports, for each industry, what share of establishments employ each occupation. Summing over occupations gives the average number of distinct occupations per establishment. Multiplying by the number of establishments in the industry and adding up over industries gives the number of distinct occupation-at-site positions in the country:
An average 5.3 occupations per establishment across 12 million establishments gives a total of 63 million distinct positions.