Nailed it. Tyler Cowen thinks marginalist logic is fading into an AI background. But your marginalist argument suggests just the opposite. I think you're right, Lynne.
"And that shift raises the value of something else: judgment."
And human judgment can assist the AI robots in their analysis. If you know how to talk productively to them. This is yet another excellent piece of analysis, and not only for the field of economics.
This is really great - I think you exactly nailed the issues here. As usual, the academy seems stuck in the past in adapting to change, but figuring out how to apply these insights to research (and teaching!) is important.
More relevant in the age of Ai than Williamson’s NIE choice of incompleteness of contracts will be Grossman and Hart’s NIE choice to own an open weight LLM to deploy within the firm (“employee”) vs contracting out for closed weight frontier LLM (“contractor”) who will own residual decision rights. Within last week, wave of podcasts (Nardelly, Theil, All-In ) suddenly realizing better to own than contract.
One last thought on NIE approaches to questions regarding use of AI that merit further research: Elinor Ostrom on conditions favorable to formation of private cooperatives of firms in same industry with similar SoR club goods to create, share, and govern an AI native venture (AI joined club goods). e.g hospitals systems sharing Epic EHR creating GPO at point of care
Another recent post by VC Bill Gurley that can be read as Grossman and Hart NIE choice of own vs contract for LLM. Makes analogy between current institutional choice for LLM and 1970s and 1980s choice for owning open source OS software like UNIX vs licensing proprietary closed sources from likes IBM and DEC. https://p3institute.substack.com/p/from-open-source-software-to-open?r=2swuo&utm_medium=ios
Lynne, this is fascinating, especially the point that AI may change the relative value of different kinds of economic reasoning.
One place my mind goes is back before the computer-shaped equilibrium of the last several decades. Modern economics became increasingly organized around data, modeling, identification, and measurable outcomes. That brought real discipline, but it also narrowed the kinds of questions that felt professionally valuable.
Schumpeter is interesting in this context. He was close to the early institutional formation of econometrics — he chaired the 1930 meeting that founded the Econometric Society — yet his own great strength was not narrow measurement. It was structural imagination: innovation, creative destruction, business cycles, entrepreneurship, and the patterned instability of capitalism.
That makes me wonder whether AI may reopen space for a kind of economic reasoning that became harder to reward during the high-computation, high-econometrics period.
Not a return to loose speculation.
Something different.
If AI lowers the cost of search, synthesis, modeling, comparison, and evidence-gathering, then perhaps the scarce skill becomes the ability to identify the right structure, name the relevant mechanism, and judge which pattern actually matters.
In that sense, AI may not move economics away from theory. It may force economics to recover a deeper version of theory: disciplined reasoning about institutions, constraints, incentives, technological change, and human behavior when the measurable evidence is only part of the story.
Reminds me of an old joke. A child makes a mistake and asks his father how to avoid them going forward. The father says you have to have judgement. But how do you get judgment the child asks. It takes experience the father replies. And experience, how do you get that? he asks. By making mistakes the father replies.
Loved your exposition and the application of price theory. In terms of re-imagining education, AI can threaten learning by replacing thinking/writing by students, but I believe it can be used to sharpen and scale the feedback process so students can make many errors and get much more feedback than in the current models. We aren’t positioning AI this way, but it would relieve the scarcity of the instructors time, patience and attention with a cheap and (potentially) more accurate source.
For learning and research AI saves lots of time, but for testing students' knowledge and understanding I'm thinking a return to oral exams might be the answer.
I think you'll find ai is not just good at listening it is far better at it than humans. There was a time when humans were better at listening where some centaur approach to production was possible, but that hasn't been the case for at least 12 months, once we got context windows large enough to handle a few mb of PDF files at a time. Even if it were the case that there remains some scarcity humans can approach, what's your plan six months later when the ai does it better? We will live in a world where Intelligence is too cheap to meter.
I've recently found just what you said about 'good at listening' with Google AI, which had made an erroneous comment on something I had written in a published book review. When I corrected it, the robot immediately took my brief remarks to heart and we were off to a productive discussion of my review, that required very little input from me. The speed with which it completely changed its opinion (from 'day to night') amazed me.
Nailed it. Tyler Cowen thinks marginalist logic is fading into an AI background. But your marginalist argument suggests just the opposite. I think you're right, Lynne.
"And that shift raises the value of something else: judgment."
And human judgment can assist the AI robots in their analysis. If you know how to talk productively to them. This is yet another excellent piece of analysis, and not only for the field of economics.
This is really great - I think you exactly nailed the issues here. As usual, the academy seems stuck in the past in adapting to change, but figuring out how to apply these insights to research (and teaching!) is important.
Thanks!
Well said! AI may reduce tendencies toward the McNamara Fallacy, named after Robert McNamara’s blunders as Secretary of Defense during Vietnam:
“The first step is to measure whatever can be easily measured. This is okay as far as it goes.
The second step is to disregard that which cannot be measured or give it an arbitrary quantitative value. This is artificial and misleading.
The third step is to presume that what can't be measured easily really isn't very important. This is blindness.
The fourth step is to say that what can't be easily measured really doesn't exist. This is suicide.”
The most interesting and important piece on AI and economics I've read.
Thanks!
More relevant in the age of Ai than Williamson’s NIE choice of incompleteness of contracts will be Grossman and Hart’s NIE choice to own an open weight LLM to deploy within the firm (“employee”) vs contracting out for closed weight frontier LLM (“contractor”) who will own residual decision rights. Within last week, wave of podcasts (Nardelly, Theil, All-In ) suddenly realizing better to own than contract.
Agreed. After listening to last week’s All In, I have a couple of ideas exactly along these lines.
One last thought on NIE approaches to questions regarding use of AI that merit further research: Elinor Ostrom on conditions favorable to formation of private cooperatives of firms in same industry with similar SoR club goods to create, share, and govern an AI native venture (AI joined club goods). e.g hospitals systems sharing Epic EHR creating GPO at point of care
Another recent post by VC Bill Gurley that can be read as Grossman and Hart NIE choice of own vs contract for LLM. Makes analogy between current institutional choice for LLM and 1970s and 1980s choice for owning open source OS software like UNIX vs licensing proprietary closed sources from likes IBM and DEC. https://p3institute.substack.com/p/from-open-source-software-to-open?r=2swuo&utm_medium=ios
Just posted by Microsoft Satya Nardella: another example of Grossman and Hart NIE choice of own vs contract for LLM , begins with Arrow on information paradox https://x.com/satyanadella/status/2076323181154230284?s=46&t=8tbKkYbj6q6vGBbG1GkqhQ
Lynne, this is fascinating, especially the point that AI may change the relative value of different kinds of economic reasoning.
One place my mind goes is back before the computer-shaped equilibrium of the last several decades. Modern economics became increasingly organized around data, modeling, identification, and measurable outcomes. That brought real discipline, but it also narrowed the kinds of questions that felt professionally valuable.
Schumpeter is interesting in this context. He was close to the early institutional formation of econometrics — he chaired the 1930 meeting that founded the Econometric Society — yet his own great strength was not narrow measurement. It was structural imagination: innovation, creative destruction, business cycles, entrepreneurship, and the patterned instability of capitalism.
That makes me wonder whether AI may reopen space for a kind of economic reasoning that became harder to reward during the high-computation, high-econometrics period.
Not a return to loose speculation.
Something different.
If AI lowers the cost of search, synthesis, modeling, comparison, and evidence-gathering, then perhaps the scarce skill becomes the ability to identify the right structure, name the relevant mechanism, and judge which pattern actually matters.
In that sense, AI may not move economics away from theory. It may force economics to recover a deeper version of theory: disciplined reasoning about institutions, constraints, incentives, technological change, and human behavior when the measurable evidence is only part of the story.
Todd, that’s a great insight, and I think that’s very compatible with how I’m thinking about it.
Reminds me of an old joke. A child makes a mistake and asks his father how to avoid them going forward. The father says you have to have judgement. But how do you get judgment the child asks. It takes experience the father replies. And experience, how do you get that? he asks. By making mistakes the father replies.
Yep, pretty much. Which is why AI is such an existential shock for education, which we really now have to reimagine.
Loved your exposition and the application of price theory. In terms of re-imagining education, AI can threaten learning by replacing thinking/writing by students, but I believe it can be used to sharpen and scale the feedback process so students can make many errors and get much more feedback than in the current models. We aren’t positioning AI this way, but it would relieve the scarcity of the instructors time, patience and attention with a cheap and (potentially) more accurate source.
For learning and research AI saves lots of time, but for testing students' knowledge and understanding I'm thinking a return to oral exams might be the answer.
I think you'll find ai is not just good at listening it is far better at it than humans. There was a time when humans were better at listening where some centaur approach to production was possible, but that hasn't been the case for at least 12 months, once we got context windows large enough to handle a few mb of PDF files at a time. Even if it were the case that there remains some scarcity humans can approach, what's your plan six months later when the ai does it better? We will live in a world where Intelligence is too cheap to meter.
I've recently found just what you said about 'good at listening' with Google AI, which had made an erroneous comment on something I had written in a published book review. When I corrected it, the robot immediately took my brief remarks to heart and we were off to a productive discussion of my review, that required very little input from me. The speed with which it completely changed its opinion (from 'day to night') amazed me.
Brilliant, Lynne.
Really first rate.
Thanks David!
You're welcome, Lynne. And thank you.