The effects of AI, like everything else, depend on both supply and demand. A prelude, and then a parable. As I have been reading and discussing large language models, I find I’ve learned as much about us humans as I have about the AI that replicates some of what we do. Introspecting, am I really that much more than a large language model?
"Introspecting, am I really that much more than a large language model?" I heard Marc Andreesen make the same point sometime in the past couple of years. It has been one of the stories that stuck with *me*.
A related thought: We really only accumulate wisdom through experience. You try to communicate wisdom to kids, but it will never stick-- at best they will remember it after having had more experience. But a new AI can be endowed with the neural net type connections of a predecessor simply by copying them to the new AI (an exaggeration, but not too far off)-- effectively giving them wisdom without the need for costly and time-consuming experience. So consider the possibility that those who think AIs will need to be tempered by human wisdom, have it backwards: AIs may become a greater repository of wisdom than humans.
This is funny, but it sidesteps the real elephant in the room: China. They've declared their ambition to dominate in data and A.I., leveraging not only their data but ours, thanks to lax privacy protections—something the Patriot Act certainly didn't help. Since 2017, China has been engaging in a trade war with the U.S., dismissing free market principles. Looking for a Japan 80s-like scenario to save us is not in the cards. At the same time, they are shoring up their defenses, eyeing Taiwan, and waiting for a time of their choosing to act during some distractions in the U.S. Maybe another massive cyberattack; they may even throw in a 911 outage like the one we just experienced. Probably sprinkle in a little more elite capture, and I imagine this happens before a competent leader threatens, but are we willing to sacrifice L.A. for Tiwain?
I took your request for an AI assessment of the above blog post to Adobe Acrobat Reader. Here are AI robot's summary and 5-point characterisation of the blog post, and its response to your observation on AI's learning lapse rate.
Adobe Reader AI Summary:
“The document discusses the effects of AI on the writing and commentary business, exploring the potential impact on supply, demand, and quality.”
A. Provide a list of the 5 most important points.
1. “The effects of AI depend on both supply and demand. The document emphasizes that AI can replicate certain tasks and generate new stories, but it may struggle to replicate human creativity and the ability to train itself on new stories in a productive direction.
2. “The writing of blog posts, op-eds, and interviews may be greatly influenced by large language models (LLMs) like AI. LLMs can increase the quantity of content produced and potentially lower the cost of production, but the demand for commentary may not be flat, and the price reduction may primarily benefit the producers rather than the consumers.
3. “The demand for commentary is a quality question as well. Most commentary is considered to be of low quality, and AI editing could potentially improve the quality of commentary by raising it from a C- to a B+ level. This improvement could be beneficial in various fields, such as offering skilled companions in rural Africa.
4. “While AI can assist in generating and editing content, there will still be a need for human involvement. Humans will be required to edit AI-generated material, recommend and certify AI-produced content, and figure out what prompts will interest readers. The introduction of technology like ATMs increased bank employment by enabling the opening of more branches and offering financial services.
5. “The document suggests that AI may have a positive impact on birthrates. With declining birthrates, the emergence of artificial people (presumably referring to AI) could potentially provide a solution by shifting the focus from quantity to quality, allowing for more parental and educational investment in each child. This could lead to environmental benefits from fewer people and economic benefits from increased investment.”
D. Does AI tend to get worse on old data as it trains on new data?
“The document briefly mentions that as AI trains on new data, it tends to get worse on old data. However, it does not provide further details or explanations regarding this statement. Therefore, it is difficult to determine the specific reasons behind this phenomenon or the extent to which it applies to different AI models.
“In general, the performance of AI models can be influenced by various factors, including the quality and diversity of the training data, the algorithms used for training, and the specific task or domain being addressed. Overfitting is a common challenge in machine learning, where a model becomes too specialized in the training data and performs poorly on new, unseen data. This could potentially be one reason why AI may struggle with old data as it continues to train on new data.
“However, it is important to note that advancements in AI research and techniques are constantly being made, and newer models may be designed to mitigate or overcome such limitations. Additionally, the performance of AI models can be improved through techniques like transfer learning, where knowledge gained from training on one task or dataset is applied to another related task or dataset.
“Without further information, it is difficult to provide a definitive answer. The statement in the document suggests that AI may face challenges when dealing with old data as it continues to train on new data, but the specific details and implications would depend on the specific AI model and training process being used.”
Adobe Acrobat Reader's response to three other suggested questions about the blog post were also generated, but the responses tended to be simple elaborations around the points made in the blog post, and, as such, have been withheld in the interest of saving space.
I provide the AI robot's responses without commenting on either the blog post itself or the subject matter, as interesting as both are, in order that you can judge for yourself whether Adobe's AI robot does your musings justice or not.
Years ago, I read Isaac Asimov's "I Robot." If this old hermit remembers correctly there are "robots" running the world's economy in one of the last short stories. Perhaps Mr. Asimov isn't as sophisticated as most but he did a nice job projecting a possible future. Just the rambling thoughts of an old man.
In a recent "The Grumpy Economist" blog post, you disputed the news media's understanding of definition of 'inflation' (or, specifically, 'the rate of inflation'). "Econ 101" and "Econ 102" lessons were reprised and used to point out the economists' understanding of elementary principles. Your statement that for economists the rate of inflation is solely attributable to the rate of change in the General Price Level within an economy and not to the change in relative price levels of commodities and consumer goods/services. Bearing this in mind, the question was put to Mr. David Beckworth's AI large language model to seek confirmation or refutation of your assertion that for economists the rate of inflation is solely attributable to the rate of change in the General Price Level.
The question posed and answer provided are set out below.
Quaere: "Is the rate of inflation determined by a change in the General Price Level, as suggested by John Cochrane, or relative price level changes in commodities and retail prices as determined by Index Number Theory?"
"The rate of inflation is determined by a combination of factors, including both changes in the general price level as well as relative price level changes. John Cochrane's fiscal theory of the price level suggests that inflation is primarily driven by changes in the value of government liabilities, which can be influenced by fiscal policy. This speaks to the general price level.
"However, as highlighted by Armen Alchian, Ben Klein, and Josh Hendrickson, index number theory emphasizes that inflation should be measured by changes in the cost of a basket of goods that provides constant utility, not just current consumption flows. This incorporates relative price changes, as certain goods may become more or less expensive over time.
"George Selgin further elaborates that "good deflation" driven by productivity gains should be allowed to affect the price level, rather than being offset by monetary policy. This suggests that relative price changes stemming from supply-side factors can be an important determinant of the overall inflation rate.
"Chris Conlon's research also finds that the correlation between rising markups and rising prices is quite weak, indicating that relative price changes, not just general price level changes, are crucial for understanding inflation dynamics.
"Jonathon Hazell's work on the regional Phillips curve finds that accounting for inflation expectations is key, as this helps isolate the role of slack versus relative price changes in driving inflation.
"So in summary, the rate of inflation is determined by a combination of general price level changes, as emphasized by the fiscal theory, as well as relative price changes across different goods and services, as highlighted by index number theory and the productivity norm. Policymakers need to carefully disentangle these factors when assessing the underlying inflationary pressures in the economy."
"The following episodes were considered in answering the question:
[1] Josh Hendrickson on the Treasury Standard and Global Dollar Dominance
[2] George Selgin on the Productivity Norm, Deflation, and Monetary History
[3] Chris Conlon on the Post-COVID Inflation Surge and the Greedflation Narrative
[4] Jonathon Hazell on Phillips Curves, Wage Rigidity, and How to Measure R-Star
[5] John Cochrane on Finance, the Fiscal Theory of the Price Level, and Blogging"
Are you sure we need more parental investment in children? Statistics (admittedly of the self-reported kind) show a dramatic increase in time devoted to parenting across advanced countries (with the apparent and striking exception of France) over the past forty years or so. This is also the period that saw a plateauing if not reversal of the Flynn effect, whereby average IQ tends to increase across successive cohorts. There is also evidence, admittedly limited to the extreme case of South Korea, of the cost of investing in children’s education driving fertility rates to historic lows.
At the same time I guess we would agree that in a world of pervasive demographic decline, we can only hope that AI will accelerate the rate of labour-augmenting technical progress. Perhaps your conclusion should be read in the sense that AI will increase the productivity of teaching and learning.
I asked Gemini (Google's plebian AI) to summarize your paragraph about AI summaries as a service: "The paragraph discusses the potential of AI assistants to filter and summarize information overload. It mentions the idea of AI summarizing lengthy content into concise points, highlighting the subjective nature of what constitutes "good" content."
This is an interesting piece but first thanks for sharing. AI indeed mirrors our patterns and narratives, but human creativity and intuition still offer something unique and irreplaceable. The interplay between AI and human storytelling, influenced by supply and demand dynamics, promises a fascinating evolution rather than simple displacement.
As an old ski salesman I wonder how about this comment A last bullish AI thought. Birthrates are plummeting. Perhaps artificial people are coming along just in time. If we view what’s happening as a further shift from numbers of children to quality of children, with more parental and educational investment in each one, we might get the environmental benefits of fewer people with the economic benefits of more people.¨
Do we really get better quality of children with fewer of them? By what measure? Fewer skiers are indicative of poorer skiers normally. The further from snow, the fewer the skiers and the poorer quality. No competition!
If parents have more education, but less skiing experience, how can the qaulity of their kids skiing be better? Isn´t competition enhanced by more participants as opposed to fewer? All the education in the world won´t make better skiers.
Back in 2018, Agrawal et al did an analysis of the economics of ML used for prediction in the book, "Prediction Machines." The authors argued that prediction would become much cheaper, resulting in two main consequences:
1. We would use a lot more prediction, and in areas where it had not been used before.
2. It would raise the value of related complements (data, judgement, action) and diminish the value of substitutes (human prediction).
It's fascinating to compare that dynamic with how cheaper "language prediction" is playing out with language models now!
There is a Milton Friedman AI bot! https://friedman.ai/
A lot of stories to think about in the post. Thanks!
"Introspecting, am I really that much more than a large language model?" I heard Marc Andreesen make the same point sometime in the past couple of years. It has been one of the stories that stuck with *me*.
A related thought: We really only accumulate wisdom through experience. You try to communicate wisdom to kids, but it will never stick-- at best they will remember it after having had more experience. But a new AI can be endowed with the neural net type connections of a predecessor simply by copying them to the new AI (an exaggeration, but not too far off)-- effectively giving them wisdom without the need for costly and time-consuming experience. So consider the possibility that those who think AIs will need to be tempered by human wisdom, have it backwards: AIs may become a greater repository of wisdom than humans.
This is funny, but it sidesteps the real elephant in the room: China. They've declared their ambition to dominate in data and A.I., leveraging not only their data but ours, thanks to lax privacy protections—something the Patriot Act certainly didn't help. Since 2017, China has been engaging in a trade war with the U.S., dismissing free market principles. Looking for a Japan 80s-like scenario to save us is not in the cards. At the same time, they are shoring up their defenses, eyeing Taiwan, and waiting for a time of their choosing to act during some distractions in the U.S. Maybe another massive cyberattack; they may even throw in a 911 outage like the one we just experienced. Probably sprinkle in a little more elite capture, and I imagine this happens before a competent leader threatens, but are we willing to sacrifice L.A. for Tiwain?
I took your request for an AI assessment of the above blog post to Adobe Acrobat Reader. Here are AI robot's summary and 5-point characterisation of the blog post, and its response to your observation on AI's learning lapse rate.
Adobe Reader AI Summary:
“The document discusses the effects of AI on the writing and commentary business, exploring the potential impact on supply, demand, and quality.”
A. Provide a list of the 5 most important points.
1. “The effects of AI depend on both supply and demand. The document emphasizes that AI can replicate certain tasks and generate new stories, but it may struggle to replicate human creativity and the ability to train itself on new stories in a productive direction.
2. “The writing of blog posts, op-eds, and interviews may be greatly influenced by large language models (LLMs) like AI. LLMs can increase the quantity of content produced and potentially lower the cost of production, but the demand for commentary may not be flat, and the price reduction may primarily benefit the producers rather than the consumers.
3. “The demand for commentary is a quality question as well. Most commentary is considered to be of low quality, and AI editing could potentially improve the quality of commentary by raising it from a C- to a B+ level. This improvement could be beneficial in various fields, such as offering skilled companions in rural Africa.
4. “While AI can assist in generating and editing content, there will still be a need for human involvement. Humans will be required to edit AI-generated material, recommend and certify AI-produced content, and figure out what prompts will interest readers. The introduction of technology like ATMs increased bank employment by enabling the opening of more branches and offering financial services.
5. “The document suggests that AI may have a positive impact on birthrates. With declining birthrates, the emergence of artificial people (presumably referring to AI) could potentially provide a solution by shifting the focus from quantity to quality, allowing for more parental and educational investment in each child. This could lead to environmental benefits from fewer people and economic benefits from increased investment.”
D. Does AI tend to get worse on old data as it trains on new data?
“The document briefly mentions that as AI trains on new data, it tends to get worse on old data. However, it does not provide further details or explanations regarding this statement. Therefore, it is difficult to determine the specific reasons behind this phenomenon or the extent to which it applies to different AI models.
“In general, the performance of AI models can be influenced by various factors, including the quality and diversity of the training data, the algorithms used for training, and the specific task or domain being addressed. Overfitting is a common challenge in machine learning, where a model becomes too specialized in the training data and performs poorly on new, unseen data. This could potentially be one reason why AI may struggle with old data as it continues to train on new data.
“However, it is important to note that advancements in AI research and techniques are constantly being made, and newer models may be designed to mitigate or overcome such limitations. Additionally, the performance of AI models can be improved through techniques like transfer learning, where knowledge gained from training on one task or dataset is applied to another related task or dataset.
“Without further information, it is difficult to provide a definitive answer. The statement in the document suggests that AI may face challenges when dealing with old data as it continues to train on new data, but the specific details and implications would depend on the specific AI model and training process being used.”
Adobe Acrobat Reader's response to three other suggested questions about the blog post were also generated, but the responses tended to be simple elaborations around the points made in the blog post, and, as such, have been withheld in the interest of saving space.
I provide the AI robot's responses without commenting on either the blog post itself or the subject matter, as interesting as both are, in order that you can judge for yourself whether Adobe's AI robot does your musings justice or not.
Years ago, I read Isaac Asimov's "I Robot." If this old hermit remembers correctly there are "robots" running the world's economy in one of the last short stories. Perhaps Mr. Asimov isn't as sophisticated as most but he did a nice job projecting a possible future. Just the rambling thoughts of an old man.
In a recent "The Grumpy Economist" blog post, you disputed the news media's understanding of definition of 'inflation' (or, specifically, 'the rate of inflation'). "Econ 101" and "Econ 102" lessons were reprised and used to point out the economists' understanding of elementary principles. Your statement that for economists the rate of inflation is solely attributable to the rate of change in the General Price Level within an economy and not to the change in relative price levels of commodities and consumer goods/services. Bearing this in mind, the question was put to Mr. David Beckworth's AI large language model to seek confirmation or refutation of your assertion that for economists the rate of inflation is solely attributable to the rate of change in the General Price Level.
The question posed and answer provided are set out below.
Quaere: "Is the rate of inflation determined by a change in the General Price Level, as suggested by John Cochrane, or relative price level changes in commodities and retail prices as determined by Index Number Theory?"
Response -- by "Macro Musebot - MACRO Musings with David Beckworth" [https://www.ai.macromusings.org] -- :
"The rate of inflation is determined by a combination of factors, including both changes in the general price level as well as relative price level changes. John Cochrane's fiscal theory of the price level suggests that inflation is primarily driven by changes in the value of government liabilities, which can be influenced by fiscal policy. This speaks to the general price level.
"However, as highlighted by Armen Alchian, Ben Klein, and Josh Hendrickson, index number theory emphasizes that inflation should be measured by changes in the cost of a basket of goods that provides constant utility, not just current consumption flows. This incorporates relative price changes, as certain goods may become more or less expensive over time.
"George Selgin further elaborates that "good deflation" driven by productivity gains should be allowed to affect the price level, rather than being offset by monetary policy. This suggests that relative price changes stemming from supply-side factors can be an important determinant of the overall inflation rate.
"Chris Conlon's research also finds that the correlation between rising markups and rising prices is quite weak, indicating that relative price changes, not just general price level changes, are crucial for understanding inflation dynamics.
"Jonathon Hazell's work on the regional Phillips curve finds that accounting for inflation expectations is key, as this helps isolate the role of slack versus relative price changes in driving inflation.
"So in summary, the rate of inflation is determined by a combination of general price level changes, as emphasized by the fiscal theory, as well as relative price changes across different goods and services, as highlighted by index number theory and the productivity norm. Policymakers need to carefully disentangle these factors when assessing the underlying inflationary pressures in the economy."
"The following episodes were considered in answering the question:
[1] Josh Hendrickson on the Treasury Standard and Global Dollar Dominance
[2] George Selgin on the Productivity Norm, Deflation, and Monetary History
[3] Chris Conlon on the Post-COVID Inflation Surge and the Greedflation Narrative
[4] Jonathon Hazell on Phillips Curves, Wage Rigidity, and How to Measure R-Star
[5] John Cochrane on Finance, the Fiscal Theory of the Price Level, and Blogging"
Hey, my AI reduced your blog to just one bullet point "The law of supply and demand still applies", or more succinctly "TANSTAFL" :)
Are you sure we need more parental investment in children? Statistics (admittedly of the self-reported kind) show a dramatic increase in time devoted to parenting across advanced countries (with the apparent and striking exception of France) over the past forty years or so. This is also the period that saw a plateauing if not reversal of the Flynn effect, whereby average IQ tends to increase across successive cohorts. There is also evidence, admittedly limited to the extreme case of South Korea, of the cost of investing in children’s education driving fertility rates to historic lows.
At the same time I guess we would agree that in a world of pervasive demographic decline, we can only hope that AI will accelerate the rate of labour-augmenting technical progress. Perhaps your conclusion should be read in the sense that AI will increase the productivity of teaching and learning.
I asked Gemini (Google's plebian AI) to summarize your paragraph about AI summaries as a service: "The paragraph discusses the potential of AI assistants to filter and summarize information overload. It mentions the idea of AI summarizing lengthy content into concise points, highlighting the subjective nature of what constitutes "good" content."
This is an interesting piece but first thanks for sharing. AI indeed mirrors our patterns and narratives, but human creativity and intuition still offer something unique and irreplaceable. The interplay between AI and human storytelling, influenced by supply and demand dynamics, promises a fascinating evolution rather than simple displacement.
As an old ski salesman I wonder how about this comment A last bullish AI thought. Birthrates are plummeting. Perhaps artificial people are coming along just in time. If we view what’s happening as a further shift from numbers of children to quality of children, with more parental and educational investment in each one, we might get the environmental benefits of fewer people with the economic benefits of more people.¨
Do we really get better quality of children with fewer of them? By what measure? Fewer skiers are indicative of poorer skiers normally. The further from snow, the fewer the skiers and the poorer quality. No competition!
If parents have more education, but less skiing experience, how can the qaulity of their kids skiing be better? Isn´t competition enhanced by more participants as opposed to fewer? All the education in the world won´t make better skiers.
Back in 2018, Agrawal et al did an analysis of the economics of ML used for prediction in the book, "Prediction Machines." The authors argued that prediction would become much cheaper, resulting in two main consequences:
1. We would use a lot more prediction, and in areas where it had not been used before.
2. It would raise the value of related complements (data, judgement, action) and diminish the value of substitutes (human prediction).
It's fascinating to compare that dynamic with how cheaper "language prediction" is playing out with language models now!