As for the "what come next" question, redistribution is obviously the right answer here. As it has been since long before Chat GPT. The right response to capitalism and a winner takes all economy that produces winners and losers is to set a reasonable floor on the life of the "losers." Not complete redistribution, we still want to reward winners, but the floor should be high enough so that everyone in the "loser" category can still lead a productive life with access to health care, public transport, food, etc.
To be clear, that doesn't mean that folks get all this for doing nothing. UBI with no strings is a bad idea for all sorts of reasons. But there will always be tons of useful work to do, whether it is keeping public spaces clean, caring for the elderly and young, building infrastructure, etc. To the extent that the private sector isn't able to offer these opportunities, the state can.
How long will people have to wait to see if your idea of redistribution happens, even if in stages? Discussion of AI impact is making me more impatient to see how this aspect of social justice works out for women and minorities....like me
I think the equity prioritization needs further justification or refinement. I've worked at multiple tech startups in the Series A/B stage where I was between the 50th and 100th employee, and they all gave me ~10k options of equity when I started. If they become the next Google that will be worth a lot, but most likely it'll be worth between $0 and a few tens of thousands of dollars. That's not life changing if your salary is six figures. Isn't working two jobs and making $300-500k is much more valuable on average? Perhaps you should specify that when you say equity is the top priority you mean exclusively founder or very early employee level equity?
Fair. I do mean enough equity to matter, but I do often see quite a lot of value in equity (if with very high risk attached) given away to employees even relatively late in the game.
Maybe at some point software engineers become a hybrid between low level managers and testers. Managers because we spend a lot of time instructing and prompting GPT-N and testers because the output is not fully reliable and we need to create good tests. This would honestly make our jobs a lot more productive but a lot less fun :(
Good post! (I wanted to say something nice to counterbalance disagreeing with you on Twitter)
I've really been feeling the pressure to somehow get equity in future growth from AI but short of founding a startup (not sure it's a good match for my talents or that the talents I do have are up to it), do you think there are investments one can make?
Note that having multiple jobs is usually against your contract. And at least outside the US, where companies file your taxes for you, it's going to be very hard to hide the fact you're working for a different company without the taxman recieving two separate tax filings for you.
Here in Germany, the idea of "having several 'full time" jobs" will *immediately* run into laws limiting the total number of hours worked. If your contracts stipulate that two jobs are for 40 hours a week (or even just 30...), taking the second will *technically* be illegal, as the official limit is 50 hours a week (on average over some medium term).
Yes, OF COURSE there are many people working more than 50 hours a week. For knowledge worker/white collar type jobs and some professions, everyone mostly sort of politely ignored this fact. It seems only really enforced for blue collar jobs. (I *think* I'm using these terms right?)
Still, this seems like the kind of thing Germany would *very quickly* put a stop to. Sadly, my expectation is that Germany's reaction will lean more towards the "bullshitization" solution, as the culture leans more than average towards the idea that <comedy German accent> everywon must have EXACTLY won dschob and wörk ze CORRECT number of hours!
I do not expect that the productivity increase from AI will be uniform. Most 1x workers will remain 1x workers. A few will become 100x workers. They will be joined by some of the former 10x workers. If I were a manager, I would hire the 100x workers, pay them huge salaries, and deal with jealous 1x workers by giving them the choice between accepting inequality gracefully or getting fired
Hmm I've only worked in small (manufacturing) companies, so I have limited experience. But I don't see how that works long term. Whomever you are calling 100x workers, that is something that if motivated enough some of the 1x worker can learn how to do. So the company next door takes those 1x workers, pays them a bit more (than you were), and eats your lunch. At least that's my world view, I could be wrong.
Or in other words, the huge salaries attract more workers.
It depends on where high productivity comes from. Is it almost entirely innate talent or can anyone be motivated to be highly productive? If it's the latter, then you are right. Wages are driven to equality.
But if innate talent differences are large, then highly talented workers can earn superior incomes. Zvi says that they have to become business owners to do so. I am suggesting an alternative mechanism.
Yeah I've known some geniuses, (many not paid enough) but this looks like scribes to the printing press. And if you were a scribe, and found out about the printing press, you could make a killing, till everyone had also heard about the printing press. I'm good with slowing down for a while. But I need to add that I hate writing, and so chat-gpt may be good for other people like me, and bad for people like you and Zvi and EY, who are good at writing. An AI who can help those like me at the writing game... will be good for those like me who are not so good. (and it's in your short term interest to work against wordsmith AI.) Long term we need to reward good ideas, less than how good you are at saying them.
I’m a psychiatrist in the UK: genuinely curious if we’ll see any near term productivity gains from AI. Would obviously be ideal for summarising medical records, writing letters, automatic note taking, as a diagnostic aid (or psychiatrist replacement!)... with the caveat that we can’t send medical notes outside the UK and will have to rely on whatever NHS LLM we can run on 10,000 ZX Spectrums.
You can already transcribe English on an average laptop in near real-time with free software that doesn't use a network connection (e.g. whisper.cpp), and decent language models now run on that same laptop (e.g. ChatGPT4all-J with llama.cpp). These will be packaged in many different ways for non-techies as low cost apps this year. Making the glue for the missing 20% is likely to take the rest of the decade, and some parts won't get done at all, but if you just want the boring basics then this is soon cheap and can be run on a personal laptop or a phone.
Nice. I wonder how many jobs are only doing one (or a few) thing(s) that can be done faster now. In most of my previous tech type jobs I can see AI helping with some fraction of my work (maybe 20%), but speeding that up would have just left me more time to do all the other tasks I had to do. So maybe I'd be 15% more productive, but that is not a 5x or even 2x improvement. (As an aside, thanks for the previous link to Eliezer on the Lunar Society podcast. I'm only through 1/2, but good stuff.)
Read Vice article; So (at the moment) Chat-GPT is mostly helping with wordsmithing and coding tasks? (It's a little strange to me that wordsmithing and coding are adjacent. I like coding, except for getting all the syntax right, and I hate writing.) It seems like in the long run chat-gpt is going to make those tasks cheaper and easier for everyone. (I can use chat-gpt to craft my first draft comments on substack, and that will be a win for everyone!) Which means getting double or triple hits on income for those first users is a transient, and should go away in time. I'm much more interested in the steady state... (or at least long term), response of the system.
So is there a better search function too? I'd love some search function better than google or duck duck go.
There is an arbitrage here in the short term where it is possible multiple jobs outstrips expected returns on equity.
Especially given these productivity tools should rapidly increase the amount of startups possible by increasing the number of people capable of doing so, the default advice in the 1-3 year employer adjustment term might be to take multiple jobs at 2-4x high floor salary than take equity in some low probability of success startup.
I am optimistic about this issue (as distinguished from x-risk, which does worry me somewhat) because there are few precedents for productivity-enhancing technological innovation with really adverse impacts.
The counterexample I sometimes hear cited is the impact of blackberries (and then smartphones) on professional services, but real wages in most of those white-collar professions are up substantially from pre-blackberry levels. As one example, Cravath paid a first-year associate $83k base in 1993 ($173k in 2023 dollars) versus $215k today.
Is quality of life worse? Being always reachable is stressful, but it also means you can go home rather than stay late in the office most nights. I'd say the impact of technology on white-collar QoL is ambiguous.
Another point to expand on your post is the large % of jobs likely to be affected by technology that measure work units in billable hours. Will be interesting to see the beneficiaries of those productivity gains.
I actually have a lot of questions regarding the 10x programmers not getting paid 10x issue.
1. Why aren't there businesses specializing at identifying and employing such workers?
2. If they're not compensated accordingly, why are they still 10x programmers instead of 1x programmers and 9x shitposters?
3. Aren't there extremely strong norms/laws around discussing salaries? It was in every one of my labor contracts, and I haven't observed it being broken even once.
Re #2: In the US you have the absolute legal right to share your own salary information. There have long been social norms against doing so - it could be regarded as tacky/bragging (if your salary is higher than people expect) or could lower your social status (if lower than people expect) - but I've shared salary info with friends and had them share back on occasion (especially when negotiating for a new position). There also exist sites like glassdoor which formalize this, requiring you to share some of your own info in order to see what others have shared.
I haven't played with chat-gpt, but it seems like a reasonable transient response, from what I've seen/read.
Good piece.
As for the "what come next" question, redistribution is obviously the right answer here. As it has been since long before Chat GPT. The right response to capitalism and a winner takes all economy that produces winners and losers is to set a reasonable floor on the life of the "losers." Not complete redistribution, we still want to reward winners, but the floor should be high enough so that everyone in the "loser" category can still lead a productive life with access to health care, public transport, food, etc.
To be clear, that doesn't mean that folks get all this for doing nothing. UBI with no strings is a bad idea for all sorts of reasons. But there will always be tons of useful work to do, whether it is keeping public spaces clean, caring for the elderly and young, building infrastructure, etc. To the extent that the private sector isn't able to offer these opportunities, the state can.
How long will people have to wait to see if your idea of redistribution happens, even if in stages? Discussion of AI impact is making me more impatient to see how this aspect of social justice works out for women and minorities....like me
I think the equity prioritization needs further justification or refinement. I've worked at multiple tech startups in the Series A/B stage where I was between the 50th and 100th employee, and they all gave me ~10k options of equity when I started. If they become the next Google that will be worth a lot, but most likely it'll be worth between $0 and a few tens of thousands of dollars. That's not life changing if your salary is six figures. Isn't working two jobs and making $300-500k is much more valuable on average? Perhaps you should specify that when you say equity is the top priority you mean exclusively founder or very early employee level equity?
Fair. I do mean enough equity to matter, but I do often see quite a lot of value in equity (if with very high risk attached) given away to employees even relatively late in the game.
Maybe at some point software engineers become a hybrid between low level managers and testers. Managers because we spend a lot of time instructing and prompting GPT-N and testers because the output is not fully reliable and we need to create good tests. This would honestly make our jobs a lot more productive but a lot less fun :(
Good post! (I wanted to say something nice to counterbalance disagreeing with you on Twitter)
I've really been feeling the pressure to somehow get equity in future growth from AI but short of founding a startup (not sure it's a good match for my talents or that the talents I do have are up to it), do you think there are investments one can make?
Note that having multiple jobs is usually against your contract. And at least outside the US, where companies file your taxes for you, it's going to be very hard to hide the fact you're working for a different company without the taxman recieving two separate tax filings for you.
Yeah don't try to do that in Japan
Here in Germany, the idea of "having several 'full time" jobs" will *immediately* run into laws limiting the total number of hours worked. If your contracts stipulate that two jobs are for 40 hours a week (or even just 30...), taking the second will *technically* be illegal, as the official limit is 50 hours a week (on average over some medium term).
Yes, OF COURSE there are many people working more than 50 hours a week. For knowledge worker/white collar type jobs and some professions, everyone mostly sort of politely ignored this fact. It seems only really enforced for blue collar jobs. (I *think* I'm using these terms right?)
Still, this seems like the kind of thing Germany would *very quickly* put a stop to. Sadly, my expectation is that Germany's reaction will lean more towards the "bullshitization" solution, as the culture leans more than average towards the idea that <comedy German accent> everywon must have EXACTLY won dschob and wörk ze CORRECT number of hours!
I do not expect that the productivity increase from AI will be uniform. Most 1x workers will remain 1x workers. A few will become 100x workers. They will be joined by some of the former 10x workers. If I were a manager, I would hire the 100x workers, pay them huge salaries, and deal with jealous 1x workers by giving them the choice between accepting inequality gracefully or getting fired
Hmm I've only worked in small (manufacturing) companies, so I have limited experience. But I don't see how that works long term. Whomever you are calling 100x workers, that is something that if motivated enough some of the 1x worker can learn how to do. So the company next door takes those 1x workers, pays them a bit more (than you were), and eats your lunch. At least that's my world view, I could be wrong.
Or in other words, the huge salaries attract more workers.
It depends on where high productivity comes from. Is it almost entirely innate talent or can anyone be motivated to be highly productive? If it's the latter, then you are right. Wages are driven to equality.
But if innate talent differences are large, then highly talented workers can earn superior incomes. Zvi says that they have to become business owners to do so. I am suggesting an alternative mechanism.
Yeah I've known some geniuses, (many not paid enough) but this looks like scribes to the printing press. And if you were a scribe, and found out about the printing press, you could make a killing, till everyone had also heard about the printing press. I'm good with slowing down for a while. But I need to add that I hate writing, and so chat-gpt may be good for other people like me, and bad for people like you and Zvi and EY, who are good at writing. An AI who can help those like me at the writing game... will be good for those like me who are not so good. (and it's in your short term interest to work against wordsmith AI.) Long term we need to reward good ideas, less than how good you are at saying them.
I’m a psychiatrist in the UK: genuinely curious if we’ll see any near term productivity gains from AI. Would obviously be ideal for summarising medical records, writing letters, automatic note taking, as a diagnostic aid (or psychiatrist replacement!)... with the caveat that we can’t send medical notes outside the UK and will have to rely on whatever NHS LLM we can run on 10,000 ZX Spectrums.
You can already transcribe English on an average laptop in near real-time with free software that doesn't use a network connection (e.g. whisper.cpp), and decent language models now run on that same laptop (e.g. ChatGPT4all-J with llama.cpp). These will be packaged in many different ways for non-techies as low cost apps this year. Making the glue for the missing 20% is likely to take the rest of the decade, and some parts won't get done at all, but if you just want the boring basics then this is soon cheap and can be run on a personal laptop or a phone.
Nice. I wonder how many jobs are only doing one (or a few) thing(s) that can be done faster now. In most of my previous tech type jobs I can see AI helping with some fraction of my work (maybe 20%), but speeding that up would have just left me more time to do all the other tasks I had to do. So maybe I'd be 15% more productive, but that is not a 5x or even 2x improvement. (As an aside, thanks for the previous link to Eliezer on the Lunar Society podcast. I'm only through 1/2, but good stuff.)
Read Vice article; So (at the moment) Chat-GPT is mostly helping with wordsmithing and coding tasks? (It's a little strange to me that wordsmithing and coding are adjacent. I like coding, except for getting all the syntax right, and I hate writing.) It seems like in the long run chat-gpt is going to make those tasks cheaper and easier for everyone. (I can use chat-gpt to craft my first draft comments on substack, and that will be a win for everyone!) Which means getting double or triple hits on income for those first users is a transient, and should go away in time. I'm much more interested in the steady state... (or at least long term), response of the system.
So is there a better search function too? I'd love some search function better than google or duck duck go.
There is an arbitrage here in the short term where it is possible multiple jobs outstrips expected returns on equity.
Especially given these productivity tools should rapidly increase the amount of startups possible by increasing the number of people capable of doing so, the default advice in the 1-3 year employer adjustment term might be to take multiple jobs at 2-4x high floor salary than take equity in some low probability of success startup.
I am optimistic about this issue (as distinguished from x-risk, which does worry me somewhat) because there are few precedents for productivity-enhancing technological innovation with really adverse impacts.
The counterexample I sometimes hear cited is the impact of blackberries (and then smartphones) on professional services, but real wages in most of those white-collar professions are up substantially from pre-blackberry levels. As one example, Cravath paid a first-year associate $83k base in 1993 ($173k in 2023 dollars) versus $215k today.
Is quality of life worse? Being always reachable is stressful, but it also means you can go home rather than stay late in the office most nights. I'd say the impact of technology on white-collar QoL is ambiguous.
Another point to expand on your post is the large % of jobs likely to be affected by technology that measure work units in billable hours. Will be interesting to see the beneficiaries of those productivity gains.
I actually have a lot of questions regarding the 10x programmers not getting paid 10x issue.
1. Why aren't there businesses specializing at identifying and employing such workers?
2. If they're not compensated accordingly, why are they still 10x programmers instead of 1x programmers and 9x shitposters?
3. Aren't there extremely strong norms/laws around discussing salaries? It was in every one of my labor contracts, and I haven't observed it being broken even once.
Re #2: In the US you have the absolute legal right to share your own salary information. There have long been social norms against doing so - it could be regarded as tacky/bragging (if your salary is higher than people expect) or could lower your social status (if lower than people expect) - but I've shared salary info with friends and had them share back on occasion (especially when negotiating for a new position). There also exist sites like glassdoor which formalize this, requiring you to share some of your own info in order to see what others have shared.
https://www.nlrb.gov/about-nlrb/rights-we-protect/your-rights/your-rights-to-discuss-wages