I suspect there will be an AI bust before the benefits of AI (machine learning) show up in national productivity data. Computers and the Internet went through similar shakeouts. I don’t think today’s much hyped LLMs will ever produce anything but mediocre prose and six fingered people. OTOH, when companies invest the resources to use machine learning for drug discovery, power grid management, or process control there are large opportunities for improved productivity, but this will only become apparent when the LLM hype fades.
Your statement is incorrect. The "chatbot" from the company I referenced above operates container vessels delivering consumer goods. This real-world application demonstrates clear progress beyond simple conversational AI.
Bill, LLMs already have proven applications in drug discovery and biological research. DeepMind's protein exploration model demonstrates this.
But their impact isn't limited to those fields. Effective use requires skill in prompt engineering - a competency you seem to lack. Your inability to utilize a tool doesn't negate its value. An untrained civilian can't fly an F-16, but that doesn't make the jet useless.
This is an issue of terminology. The examples you give show incredible potential the of AI/machine learning. They were trained on the sum of available data in their fields and made important scientific advances. An important part of the process is the the data they were trained on was vetted before it was used. But as far as my understanding of the term, those weren’t Large Language Models (LLMs). The current crop of LLMs, like Chat GTP, were trained by scraping the internet for as much written and visual information as possible, with little regard to the quality, or context of the of the information. As a result, LLMs can make up facts or write scientific articles with references that don’t exist (my sister, an medical editor, has seen an example of this). The LLM models are being advertised as the next big thing, but I think they are vastly overhyped. There will be a shakeout. Meanwhile the people working on drug discovery or novel materials with AI are where the true potential lies.
There will certainly be an AI bust in terms of a lot of the startups going out of business, as always happens. Mediocre prose is underrated, however. The economy largely runs on mediocre prose and other similar outputs and inputs, and automating all of that stuff will free people up to do slightly more sophisticated work. LLMs have their shortcomings, but I bet they will be a critical universal language connector to other types of models like the ones you mention.
Interesting and important topic. Generally agree with the conclusions... productivity up, competitive advantages exist, and creative destruction is very important. However, the data collection and metrics leave a lot of questions. For example, how is outsourcing accounted for in this flow. This was the predominant economic reality in the period where it is claimed that productivity was flat.
It's a great question and the data is a lot more complicated than the one I showed about output/labor hours. There are different measures of total factor productivity that try to account for this change in the structure of production. I wouldn't be so sure outsourcing would mean we are overstating. If anything, that should have generate a boom, since you have fewer US labor hours to produced the same output.
On outsourcing, perhaps.. it all depends on exactly where it is recorded. A lot of the outsourcing is inside multi-national with complex international corporate structures. Also, another factor is illegal immigration..which also spiked during this period. As Milton Freedman says, its the unsaid solution for a first-world economy for productivity. Overall, I am skeptical on the claim of productivity stagnation....having lived through this period. If I had time, I would dig a lot more into the exact measurement mechanisms. Anyway, the top level conclusions make sense.
My biggest worry is that the populists will hinder substantially American productivity dynamism with poor policy choices. I am most worried that Khan and Kanter will try their best to break up the most efficiently scaled companies to appease the anti-corproate mafia.
“Productivity” is not a perfect measure of “Total Factor Productivity”
Another factor is, hopefully, that entrepreneurs now expect the Fed to keep us closer to real-income maximizing inflation than in the past. Pre-FAIT, the threat of a Phillips Curve-inspired, Fed-induced recession “to control inflation” always loomed. Maybe this is a bit of what the author means by “momentum.”
That the economy “remains” competitive should not explain a _change_ in productivity.
I agree about AI.
There is, however, a negative factor. Fiscal deficits are higher than in the past and this represents a larger shift than in the past of resources from investment to consumption. [This statement depends on believing that most expenditures do not have NPV >0 and that taxation reduces consumption less than borrowing reduces other investment, But I think those re reasonable assumptions.]
Not only do deficits mean lower investment, but they lead the Fed to have higher than oterise interest rates which attract capital inflows whihc overvalue the dollar which shifts production toward non-traded goods where, arguably, [AI could change this] it is more difficult to raise productivity.
All those reasons - American innovation, creative destruction, AI, etc are good ones, but they also (maybe not AI) existed through the lower productivity 2010s.
We’ve seen a huge increase in corporate profits in tech (and more corporate profits being booked in the US due to tax law changes) as well as depressed manufacturing employment in older industries (hangover from Covid overconsumption and overproduction- see also, Germany and China)- both of these factors boost productivity. The latter boost, in particular, could be temporary. I am not willing to extrapolate from a couple years of data.
Moreover, most of the employment gains have come in the lowest productivity sectors more aligned with government spending - healthcare, social services, government employment and leisure/hospitality (this driven by consumer spending moreso than govt).
I would be more optimistic if someone told me the writer was politically neutral rather than a partisan (I have no idea). Almost all of the “productivity boom” narratives I’ve seen so far have come from Biden and IRA cheerleaders - they want it to be true (despite actual manufacturing in the US and globally being in near recession for the last two years- see PMIs and employment). When a writer wants something to be true they can usually find a way to assemble stats to help their case.
Being a professional forecaster in my part time, I rely more heavily on base rate and it takes me awhile to start drinking any Kool-Aid (perhaps too long), so I am still in the 1 pct productivity growth camp long-term.
Of course, there is a demographic case for increased productivity and investment (with consumer spending aided by profligate governments focused on bread and circuses and AI helping out with demographics ), though it hasn’t really played out in Japan or Europe, which are ahead of the US in demographic decline.
This article was written before the BLS 818K DOWNWARD revision of jobs. The takeaway is that we cannot trust official statistics anywhere. The statistics being used in many cases are cooked. That is why consumer sentiment is at such variance with Noah and DeLong's happy talk. Facts on the ground say otherwise. I know. I am on the ground and hear what people are saying and experiencing. So, the DEGREE and IMPACT of AI improvement in productivity has to be taken with skepticism. How much productivity is because people are working longer hours to maintain their positions. I know quite a bunch of them. You can't trust hours worked metric any more than employment statistics.
When agencies lie to themselves and others with reports to support a narrative that the ruling elites desire, we end up in a situation where no believes them anymore. It started with the 2001 financial crisis and have gotten worse. Remember Bernake's "green shoots of recovery" talk? Shades of Soviet statistics.
This propensity to lie is very dangerous because demagogues can take over the public debate and say things people believe because it fits their circumstances and perceptions.
With that the AI productivity improvements are coming but the magnitude I believe will be dampened.
I suspect there will be an AI bust before the benefits of AI (machine learning) show up in national productivity data. Computers and the Internet went through similar shakeouts. I don’t think today’s much hyped LLMs will ever produce anything but mediocre prose and six fingered people. OTOH, when companies invest the resources to use machine learning for drug discovery, power grid management, or process control there are large opportunities for improved productivity, but this will only become apparent when the LLM hype fades.
I agree. "To be clear, the progress isn’t about chatbots."
https://www.grey-wing.com/
Your statement is incorrect. The "chatbot" from the company I referenced above operates container vessels delivering consumer goods. This real-world application demonstrates clear progress beyond simple conversational AI.
This happened 40 years ago with the Video Game Crash. After it bottomed out, Nintendo moved into the space and is still holding strong.
Bill, LLMs already have proven applications in drug discovery and biological research. DeepMind's protein exploration model demonstrates this.
But their impact isn't limited to those fields. Effective use requires skill in prompt engineering - a competency you seem to lack. Your inability to utilize a tool doesn't negate its value. An untrained civilian can't fly an F-16, but that doesn't make the jet useless.
This is an issue of terminology. The examples you give show incredible potential the of AI/machine learning. They were trained on the sum of available data in their fields and made important scientific advances. An important part of the process is the the data they were trained on was vetted before it was used. But as far as my understanding of the term, those weren’t Large Language Models (LLMs). The current crop of LLMs, like Chat GTP, were trained by scraping the internet for as much written and visual information as possible, with little regard to the quality, or context of the of the information. As a result, LLMs can make up facts or write scientific articles with references that don’t exist (my sister, an medical editor, has seen an example of this). The LLM models are being advertised as the next big thing, but I think they are vastly overhyped. There will be a shakeout. Meanwhile the people working on drug discovery or novel materials with AI are where the true potential lies.
There will certainly be an AI bust in terms of a lot of the startups going out of business, as always happens. Mediocre prose is underrated, however. The economy largely runs on mediocre prose and other similar outputs and inputs, and automating all of that stuff will free people up to do slightly more sophisticated work. LLMs have their shortcomings, but I bet they will be a critical universal language connector to other types of models like the ones you mention.
Interesting and important topic. Generally agree with the conclusions... productivity up, competitive advantages exist, and creative destruction is very important. However, the data collection and metrics leave a lot of questions. For example, how is outsourcing accounted for in this flow. This was the predominant economic reality in the period where it is claimed that productivity was flat.
It's a great question and the data is a lot more complicated than the one I showed about output/labor hours. There are different measures of total factor productivity that try to account for this change in the structure of production. I wouldn't be so sure outsourcing would mean we are overstating. If anything, that should have generate a boom, since you have fewer US labor hours to produced the same output.
On outsourcing, perhaps.. it all depends on exactly where it is recorded. A lot of the outsourcing is inside multi-national with complex international corporate structures. Also, another factor is illegal immigration..which also spiked during this period. As Milton Freedman says, its the unsaid solution for a first-world economy for productivity. Overall, I am skeptical on the claim of productivity stagnation....having lived through this period. If I had time, I would dig a lot more into the exact measurement mechanisms. Anyway, the top level conclusions make sense.
My biggest worry is that the populists will hinder substantially American productivity dynamism with poor policy choices. I am most worried that Khan and Kanter will try their best to break up the most efficiently scaled companies to appease the anti-corproate mafia.
Excellent read.
“Productivity” is not a perfect measure of “Total Factor Productivity”
Another factor is, hopefully, that entrepreneurs now expect the Fed to keep us closer to real-income maximizing inflation than in the past. Pre-FAIT, the threat of a Phillips Curve-inspired, Fed-induced recession “to control inflation” always loomed. Maybe this is a bit of what the author means by “momentum.”
That the economy “remains” competitive should not explain a _change_ in productivity.
I agree about AI.
There is, however, a negative factor. Fiscal deficits are higher than in the past and this represents a larger shift than in the past of resources from investment to consumption. [This statement depends on believing that most expenditures do not have NPV >0 and that taxation reduces consumption less than borrowing reduces other investment, But I think those re reasonable assumptions.]
https://thomaslhutcheson.substack.com/p/fiscal-policy-and-everything-else
Not only do deficits mean lower investment, but they lead the Fed to have higher than oterise interest rates which attract capital inflows whihc overvalue the dollar which shifts production toward non-traded goods where, arguably, [AI could change this] it is more difficult to raise productivity.
Good piece, thanks. Hope he is right.
All those reasons - American innovation, creative destruction, AI, etc are good ones, but they also (maybe not AI) existed through the lower productivity 2010s.
We’ve seen a huge increase in corporate profits in tech (and more corporate profits being booked in the US due to tax law changes) as well as depressed manufacturing employment in older industries (hangover from Covid overconsumption and overproduction- see also, Germany and China)- both of these factors boost productivity. The latter boost, in particular, could be temporary. I am not willing to extrapolate from a couple years of data.
Moreover, most of the employment gains have come in the lowest productivity sectors more aligned with government spending - healthcare, social services, government employment and leisure/hospitality (this driven by consumer spending moreso than govt).
I would be more optimistic if someone told me the writer was politically neutral rather than a partisan (I have no idea). Almost all of the “productivity boom” narratives I’ve seen so far have come from Biden and IRA cheerleaders - they want it to be true (despite actual manufacturing in the US and globally being in near recession for the last two years- see PMIs and employment). When a writer wants something to be true they can usually find a way to assemble stats to help their case.
Being a professional forecaster in my part time, I rely more heavily on base rate and it takes me awhile to start drinking any Kool-Aid (perhaps too long), so I am still in the 1 pct productivity growth camp long-term.
Of course, there is a demographic case for increased productivity and investment (with consumer spending aided by profligate governments focused on bread and circuses and AI helping out with demographics ), though it hasn’t really played out in Japan or Europe, which are ahead of the US in demographic decline.
Hope he’s right, but we’ll see.
“Deputy Assistant Attorney General for Economics at the DOJ”
Does that strike anyone else as bizarre? Is there one for Physics, Chemistry, Psychology…?
I like Brian Albrecht's "A Data-Driven Case for Productivity Optimism" Noah's substack.
Brian's "optimism comes from three different sources: recent momentum, America's enduring competitive advantage, and technological change."
I agree with Brian's "third reason for optimism: we are on the verge of an AI-driven boom."
Brian concludes, "there's ample reason for optimism about America's economic future."
With optimism as the key to innovation, resilience, and strategy, our role is to foster ever more infectiously optimistic leaders, too.
#optimism #optimist #optimistic
This article was written before the BLS 818K DOWNWARD revision of jobs. The takeaway is that we cannot trust official statistics anywhere. The statistics being used in many cases are cooked. That is why consumer sentiment is at such variance with Noah and DeLong's happy talk. Facts on the ground say otherwise. I know. I am on the ground and hear what people are saying and experiencing. So, the DEGREE and IMPACT of AI improvement in productivity has to be taken with skepticism. How much productivity is because people are working longer hours to maintain their positions. I know quite a bunch of them. You can't trust hours worked metric any more than employment statistics.
When agencies lie to themselves and others with reports to support a narrative that the ruling elites desire, we end up in a situation where no believes them anymore. It started with the 2001 financial crisis and have gotten worse. Remember Bernake's "green shoots of recovery" talk? Shades of Soviet statistics.
This propensity to lie is very dangerous because demagogues can take over the public debate and say things people believe because it fits their circumstances and perceptions.
With that the AI productivity improvements are coming but the magnitude I believe will be dampened.
I’m Larry and I’m a paranoid nut.
https://engineering.stanford.edu/magazine/robot-learns-clean-your-space-just-way-you-it