The best way for AI to raise GDP would not necessarily improve human welfare.
In my industry AI is used for up-coding, that is finding ways to increase the number of Hierarchical Condition Category codes that a person has. It does this precisely by prompting the doctor during the visit to write up the codes. This increases revenue from MAPD (thus, increases government spending) which "increases" GDP.
A similar dynamic seems to play out in sports gambling (AI attuned to figure out how to manipulate people into betting).
Like a lot of technology I can imagine good or ill uses. But GDP only measures whether a purchase was made or not.
Good thing everybody knows what MAPD means. But if you're too young, "MAPD stands for Medicare Advantage Prescription Drug plans, which are Medicare-approved health plans from private companies that offer all-in-one coverage for Medicare Parts A and B, and often include prescription drug coverage (Part D)."
Coincidentally I've been thinking about how GDP (and related measures) is failing to capture perceived wealth. A lot of things seems to fail to be properly captured, but technological improvements seem to be particularly poorly caught.
If I can use an AI to write half a dozen emails that would normally take me 10minutes each that's an hour of my life I've got back (less time taken to read the emails before hitting send). That number totally fails to show up in GDP
Worse, if your case 7 applies and we can use AI to program our lawnmowers, roombas etc. we'll see a diminution in the amount of human provided services in those areas and thus, for those people who pay for gardeners etc. rather than do it themselves, a reduction in calculated GDP because they no longer need to pay Jose, Maria and their friends to do it for them. (Although some of those services are already missed by GDP since they are paid in cash and undeclared by the recipient). But that dimunution is in fact false because the people who used to hire gardeners and cleaners and now have a robot to do it are saving time and money thanks to the new technology. That ought to make them, and the nation, richer not poorer
I agree that time seems to be a more relevant measure for much in life. The question is whether it can be measured adequately. A poor correlate that can be measured might be better than a perfect correlate that can only be poorly estimated.
Then again, maybe time isn't a great measure of progress. How would one factor in that we spending an increasing amount of our time looking at our phone?
Do we spend the time looking at the phone because we want to? or because we have to? If we aren't paid to do so and aren't forced to do so then I think that counts as a voluntary leisure activity and therefore in economic terms a benefit of our freedom from needing to do something else to survive
I got excited when I saw the title for #2 – i.e., an LLM’s value is as an interface, not an encyclopedia. But the focus was narrowly on the ‘personality’ of the LLM as a human-to-machine interface. I think the point is much bigger than that, and more profound.
LLMs are going to be able to take data and information in whatever form they happen to exist in, and translate them into a form that is useful in another domain.
That might be instantaneous translation, or allowing one system to access a database of information that it wasn’t programmed to interact with, or taking information of a kind that isn’t easily parsed into a traditional database and making it something you can sort and query. A company that wants to upgrade an old, archaic system would be able to relatively easily re-factor information captured in that system into something new. (I saw something on twitter about a guy taking code from a very old video game, feeding it into an LLM, and having the LLM rewrite the code to make it playable on a modern computer.) It might be useful in a company to understand what information needs to be disseminated to which people, at what time, without needing people to spend time distilling and communicating.
LLMs will reduce a tremendous amount of friction all over the place, again, in allowing information to flow from one domain and context to another. That will be hard to measure, I suppose, but it will unlock incredible efficiencies.
Arnold, your perspective of Ai is compelling and sensible. As for measuring GDP, the notion that a political economy is little more than an accounting ledger is spot on. The circular flow model of the aggregate economy that matches up well with the notion of GDP seems worse than naive these days. The CFM seemed misleading to me 45 years ago in graduate school; it seems worse now.
Ai as a replacement for GUI, (which was a huge advance over command-line) and other means of interacting with a computer is absolutely HUGE. I cannot wait for the full-blown embodiment of LLM technology for device input-output. Is it just me, or has remembering user interfaces with devices become ridiculously complicated; e.g., "press this button while holding the other button down (while winking)."
Science fiction "terminator" perspectives of Ai are baseless, in my opinion. But I also think that LLM is but the leading edge of a much bigger, future Ai tent.
It seems to me like AI assistants could be a huge help to disabled people. A family member is autistic, very forgetful, and diabetic. Not an easy combination. I could imagine an AI assistant extending her lifespan significantly.
Even for people with average cognitive abilities, I'd expect diabetes management AI could be a big help. Hopefully the FDA will not be too totalitarian about restricting such software.
If nothing else, your one suggestion - "Think of a spam filter for everything, not just email. Filter your social media feed, your text messages, and your news sources." - will save me a lot of time.
"A Large Language Model’s main value is as an interface, not an encyclopedia"
That's a bad way to frame the question; it's forcing a false choice as in reality the value is a combination of both. Just like the value of human expert contractors is a combination of both. We just take the "try to understand me / break it down for me in layman's terms" interface part of human-layman-to-human-expert interaction for granted, though, note, many human experts are -despite the benefit of being born and raised human- *still very bad* at doing these interactions well.
An encyclopedia that can only be quickly used by pre-trained human experts will have value, but it's not the *huge* value that can be unlocked in a genuinely low-learning-curve mass-user scenario. Likewise, what's the point of an interface unless it allows you to specify instructions to direct one particular outcome or instance out of the huge number of possibilities in any category or medium or even large sets thereof, as might constitute an "encyclopedia" of those possibilities?
So, the big value of LLM's in in having comprehensive encyclopedic reference mapped to a "laymen's terms" specification and interaction language. Which is just another way of saying that the big value is "replacing expert service-providers at scale at negligible marginal cost."
Productivity - In response to another comment, AK mentioned gdp as a measure of progress. As best I can tell, gdp is used as the output for calculating productivity. If the number of widgets per input doubles but the price goes down, the gdp doesn't double. One could try to account for the increased number of widgets but this runs into multiple problems including products that aren't widgets and, as mentioned in AK's substack post, accounting for product improvement and new products. Are TVs today, of the same size, 10% better than the vacuum tube tech of 50 years ago or a million times better? I think one could easily make either argument. What about vs black and white TVs? Broadcast vs video tape rentals vs streaming? I would add that all of this makes it impossible to accurately measure inflation too.
Maybe none of this truly matters. Maybe what we are currently measuring tells us what we need to know. Maybe the constant racheting up of standards via what I believe to be over-estimation of inflation is the right outcome. IDK. But I still think understanding what these measures tells us, and don't tell us, is a a useful pursuit.
I worked in r&d. I got funded for projects. I'd ask for a certain amount I might get it it I might get less. If I got less I'd doy best for that funding level. But does that mean my productivity increased because I did it for less? Other times the funding would be for an ongoing project. How does one determine the productivity on that? (My brother worked on the same Navy contract for 30 years, until it was canceled. What's the productivity on that?)
To complicate matters more, I might not have enough of my time or other planned resources available for the project. Now I go to secondary resources that probably aren't as good. Maybe the result degrades, maybe it doesn't even get completed. What's the productivity.
Not to say these complications are unimportant to productivity but they almost totally miss all the way that technology has made it easier to create presentation "slides," write print, and submit documentation and all kinds of other productivity improvement. it also hides the cost of increased bureaucracy in the process.
Yes, AI does seem to be influencing how we live in the sense that it appears to offer individuals more agency in their information consumption choices and at the same time allowing individuals to increase the limits of their curiosity.
The internet, as you note, set these forces in motion before AI. I noticed several years ago that friends and relatives were asking their cell phones questions more and more frequently. A joke was “If God had meant for us to be ignorant, he wouldn’t have given us cell phones.” And now with browser and twitter AIs able to answer questions, I find myself being less and less a passive consumer of the many newsletters and substacks that clog my inbox, and more and more of an active practitioner of curiosity, in a manner perhaps similar to a “Klingbot” but also similar to the old “answer-bot.”
For example, I woke up with Romania on the brain, and rather than checking my inbox, I went straight to an AI and asked about the reaction to the news from the different regions of the world. I found the response interesting and satisfying, and followed up with a query about late-breaking developments around the world that were not receiving much attention in the US press relative to the press in other regions of the world. Again, I was glad I took the time to do so. Answer-bot sure, but also in the vein of the Klingbot in that the AI was searching for things of interest to me and saving me from sifting through a lot of sources.
Since the news media acts in many ways as a transaction cost, inter-mediating itself between information and the information consumer, perhaps AI enabled curiosity satisfaction will show up as decline in GDP. Estimates of the GDP share of media companies seem to be all over the place, ranging from around 6% to 10% of US GDP. Maybe AI success can be measured in whether that share can be shrunk below 5%?
The best way for AI to raise GDP would not necessarily improve human welfare.
In my industry AI is used for up-coding, that is finding ways to increase the number of Hierarchical Condition Category codes that a person has. It does this precisely by prompting the doctor during the visit to write up the codes. This increases revenue from MAPD (thus, increases government spending) which "increases" GDP.
A similar dynamic seems to play out in sports gambling (AI attuned to figure out how to manipulate people into betting).
Like a lot of technology I can imagine good or ill uses. But GDP only measures whether a purchase was made or not.
Good thing everybody knows what MAPD means. But if you're too young, "MAPD stands for Medicare Advantage Prescription Drug plans, which are Medicare-approved health plans from private companies that offer all-in-one coverage for Medicare Parts A and B, and often include prescription drug coverage (Part D)."
33 million people, 54% of people enrolled in Medicare, are on mapd.
Yup, and I'll bet those old people know what MAPD means.
For me, the interesting question is whether it increases productivity and whether it changes .measured productivity.
Coincidentally I've been thinking about how GDP (and related measures) is failing to capture perceived wealth. A lot of things seems to fail to be properly captured, but technological improvements seem to be particularly poorly caught.
If I can use an AI to write half a dozen emails that would normally take me 10minutes each that's an hour of my life I've got back (less time taken to read the emails before hitting send). That number totally fails to show up in GDP
Worse, if your case 7 applies and we can use AI to program our lawnmowers, roombas etc. we'll see a diminution in the amount of human provided services in those areas and thus, for those people who pay for gardeners etc. rather than do it themselves, a reduction in calculated GDP because they no longer need to pay Jose, Maria and their friends to do it for them. (Although some of those services are already missed by GDP since they are paid in cash and undeclared by the recipient). But that dimunution is in fact false because the people who used to hire gardeners and cleaners and now have a robot to do it are saving time and money thanks to the new technology. That ought to make them, and the nation, richer not poorer
Coyle talks about focusing on the use of time rather than GDP as a measure of progress
I agree that time seems to be a more relevant measure for much in life. The question is whether it can be measured adequately. A poor correlate that can be measured might be better than a perfect correlate that can only be poorly estimated.
Then again, maybe time isn't a great measure of progress. How would one factor in that we spending an increasing amount of our time looking at our phone?
Do we spend the time looking at the phone because we want to? or because we have to? If we aren't paid to do so and aren't forced to do so then I think that counts as a voluntary leisure activity and therefore in economic terms a benefit of our freedom from needing to do something else to survive
It was a comment meant more for the heart than the brain.
I got excited when I saw the title for #2 – i.e., an LLM’s value is as an interface, not an encyclopedia. But the focus was narrowly on the ‘personality’ of the LLM as a human-to-machine interface. I think the point is much bigger than that, and more profound.
LLMs are going to be able to take data and information in whatever form they happen to exist in, and translate them into a form that is useful in another domain.
That might be instantaneous translation, or allowing one system to access a database of information that it wasn’t programmed to interact with, or taking information of a kind that isn’t easily parsed into a traditional database and making it something you can sort and query. A company that wants to upgrade an old, archaic system would be able to relatively easily re-factor information captured in that system into something new. (I saw something on twitter about a guy taking code from a very old video game, feeding it into an LLM, and having the LLM rewrite the code to make it playable on a modern computer.) It might be useful in a company to understand what information needs to be disseminated to which people, at what time, without needing people to spend time distilling and communicating.
LLMs will reduce a tremendous amount of friction all over the place, again, in allowing information to flow from one domain and context to another. That will be hard to measure, I suppose, but it will unlock incredible efficiencies.
Arnold, your perspective of Ai is compelling and sensible. As for measuring GDP, the notion that a political economy is little more than an accounting ledger is spot on. The circular flow model of the aggregate economy that matches up well with the notion of GDP seems worse than naive these days. The CFM seemed misleading to me 45 years ago in graduate school; it seems worse now.
Ai as a replacement for GUI, (which was a huge advance over command-line) and other means of interacting with a computer is absolutely HUGE. I cannot wait for the full-blown embodiment of LLM technology for device input-output. Is it just me, or has remembering user interfaces with devices become ridiculously complicated; e.g., "press this button while holding the other button down (while winking)."
Science fiction "terminator" perspectives of Ai are baseless, in my opinion. But I also think that LLM is but the leading edge of a much bigger, future Ai tent.
It seems to me like AI assistants could be a huge help to disabled people. A family member is autistic, very forgetful, and diabetic. Not an easy combination. I could imagine an AI assistant extending her lifespan significantly.
Even for people with average cognitive abilities, I'd expect diabetes management AI could be a big help. Hopefully the FDA will not be too totalitarian about restricting such software.
If nothing else, your one suggestion - "Think of a spam filter for everything, not just email. Filter your social media feed, your text messages, and your news sources." - will save me a lot of time.
"A Large Language Model’s main value is as an interface, not an encyclopedia"
That's a bad way to frame the question; it's forcing a false choice as in reality the value is a combination of both. Just like the value of human expert contractors is a combination of both. We just take the "try to understand me / break it down for me in layman's terms" interface part of human-layman-to-human-expert interaction for granted, though, note, many human experts are -despite the benefit of being born and raised human- *still very bad* at doing these interactions well.
An encyclopedia that can only be quickly used by pre-trained human experts will have value, but it's not the *huge* value that can be unlocked in a genuinely low-learning-curve mass-user scenario. Likewise, what's the point of an interface unless it allows you to specify instructions to direct one particular outcome or instance out of the huge number of possibilities in any category or medium or even large sets thereof, as might constitute an "encyclopedia" of those possibilities?
So, the big value of LLM's in in having comprehensive encyclopedic reference mapped to a "laymen's terms" specification and interaction language. Which is just another way of saying that the big value is "replacing expert service-providers at scale at negligible marginal cost."
Productivity - In response to another comment, AK mentioned gdp as a measure of progress. As best I can tell, gdp is used as the output for calculating productivity. If the number of widgets per input doubles but the price goes down, the gdp doesn't double. One could try to account for the increased number of widgets but this runs into multiple problems including products that aren't widgets and, as mentioned in AK's substack post, accounting for product improvement and new products. Are TVs today, of the same size, 10% better than the vacuum tube tech of 50 years ago or a million times better? I think one could easily make either argument. What about vs black and white TVs? Broadcast vs video tape rentals vs streaming? I would add that all of this makes it impossible to accurately measure inflation too.
Maybe none of this truly matters. Maybe what we are currently measuring tells us what we need to know. Maybe the constant racheting up of standards via what I believe to be over-estimation of inflation is the right outcome. IDK. But I still think understanding what these measures tells us, and don't tell us, is a a useful pursuit.
I worked in r&d. I got funded for projects. I'd ask for a certain amount I might get it it I might get less. If I got less I'd doy best for that funding level. But does that mean my productivity increased because I did it for less? Other times the funding would be for an ongoing project. How does one determine the productivity on that? (My brother worked on the same Navy contract for 30 years, until it was canceled. What's the productivity on that?)
To complicate matters more, I might not have enough of my time or other planned resources available for the project. Now I go to secondary resources that probably aren't as good. Maybe the result degrades, maybe it doesn't even get completed. What's the productivity.
Not to say these complications are unimportant to productivity but they almost totally miss all the way that technology has made it easier to create presentation "slides," write print, and submit documentation and all kinds of other productivity improvement. it also hides the cost of increased bureaucracy in the process.
Yes, AI does seem to be influencing how we live in the sense that it appears to offer individuals more agency in their information consumption choices and at the same time allowing individuals to increase the limits of their curiosity.
The internet, as you note, set these forces in motion before AI. I noticed several years ago that friends and relatives were asking their cell phones questions more and more frequently. A joke was “If God had meant for us to be ignorant, he wouldn’t have given us cell phones.” And now with browser and twitter AIs able to answer questions, I find myself being less and less a passive consumer of the many newsletters and substacks that clog my inbox, and more and more of an active practitioner of curiosity, in a manner perhaps similar to a “Klingbot” but also similar to the old “answer-bot.”
For example, I woke up with Romania on the brain, and rather than checking my inbox, I went straight to an AI and asked about the reaction to the news from the different regions of the world. I found the response interesting and satisfying, and followed up with a query about late-breaking developments around the world that were not receiving much attention in the US press relative to the press in other regions of the world. Again, I was glad I took the time to do so. Answer-bot sure, but also in the vein of the Klingbot in that the AI was searching for things of interest to me and saving me from sifting through a lot of sources.
Since the news media acts in many ways as a transaction cost, inter-mediating itself between information and the information consumer, perhaps AI enabled curiosity satisfaction will show up as decline in GDP. Estimates of the GDP share of media companies seem to be all over the place, ranging from around 6% to 10% of US GDP. Maybe AI success can be measured in whether that share can be shrunk below 5%?