& I do not have one. I do have scattered observations on how it might work out for the tech industry. How it is affecting the overall business cycle—that is another equally complicated topic that...
"Programmer copilots like GitHub Copilot boost output for the already-productive if not yet for experts"
You have made this claim more than once, but I think you are too credulous. I would like to raise 3 points.
1. At least one study has found that programmers self-report significant productivity improvements in situations where measured productivity declines.
2. I use an AI programming assistant at work and I am required to report productivity gains regularly. I am not allowed to report negative gains. So at least some surveys are based on truncated data.
3. I have a friend who teaches practical programming courses at university. He tells me that new students are very bullish on AI, smugly do their assignments with AI, are always discussing and comparing various AI tools. But once they start getting jobs, they gradually stop using AI. This is long before the "expert" stage.
This is not to say that AI assistants are useless; I think they are suitable front ends for scripting languages like Python to write throw-away data munging programs, for example. But that is far short of the claims being made for them.
Not too put too fine a point on it, the US is one big grift. The President is a notorious con artist, who has nonetheless secured near-majorities of the vote three times in a row, and is busy looting the Treasury. The valuations of AI hyperscalers look restrained compared to that of Tesla, a company with declining sales valued at 100 times earnings. And the entire financial corporate class has signed on to crypto, an utterly valueless set of notional assets with a value in the trillions. Sooner or later this will stop, but who can say when? There's a lot of ruin in a country as big as the US.
> John Quiggin: 'Not too put too fine a point on it, the US is one big grift. The President is a notorious con artist, who has nonetheless secured near-majorities of the vote three times in a row, and is busy looting the Treasury. The valuations of AI hyperscalers look restrained compared to that of Tesla, a company with declining sales valued at 100 times earnings. And the entire financial corporate class has signed on to crypto, an utterly valueless set of notional assets with a value in the trillions. Sooner or later this will stop, but who can say when? There's a lot of ruin in a country as big as the US...
"The obvious parallel, of course, is in the stories we tell of the California Gold Rush: miners rarely struck it rich, but sellers of Levi’s, shovels, and whiskey did."
Hmmmm, I think a more appropriate parallel is the building out of the US railroad system in the second half of the 19th century. Richard White's wonderful book, "Railroaded: The Transcontinentals and the Making of Modern America" is required reading on this. Lots of cautionary lessons, though Stanford University was one positive result.
I believe the "promises" of AI are overblown and will have an impact on only a few employment sectors. I continue to comment on a variety of Substacks, that nobody is focusing on jobs that will never be replaced by AI. This is a far more useful thought experiment than worrying about the build out of lots of cloud storage centers that most certainly will be vacated in ten years (maybe sooner).
I, too, find the railroads along with the telegraph appropriate analogies. It is worth noting how important the two together are to their eventual impact. Likewise, the automobile and electricity. Part of my skepticism about the eventual size and scope of the AI transformation is that it does not have a partner to dance alongside it, creating opportunities.
But unlike railway tracks and even fiber-optic cable, those GPUs can be sold off in units to willing buyers as cents on teh dollar, allowing their processing pipeline power to be used for local applications in businesses and consumer home computing.
Apollo did produce some spinoffs, but idk if they were sufficient compensation for the costs of Apollo. Transistor technology used by the space program probably was important in getting that technology used that supplanted valves. Maybe it was even important in pushing forward microelectronics, idk. Will data centers dedicated to AI be useful technology that is repurposed for general internet use, or even supercomputers, or will they become obsolete and the individual servers and boards auctioned off cheaply?
[I remember when older Sun servers became available as cheap computers for the home user. OTOH, I never saw Silicon Graphics computers ever become available, nor cheap "Lisp machines" or thin Java clients.
How many home users even need a computer? Most everyone uses just phones and tablets for the vast majority of their needs. Maybe I am a special case as I build my own workstation for home use (photography, financial analysis and writing). I certainly don't need a top GPU for these uses, and it is really on gamers and video editors that do. I don't think these NVidia GPUs can be repurposed at all. Even crypto mining of Bitcoin is nearing the end as Satoshi designed it so that is another use that is disappearing. What I don't know is what happens to all the data centers that get built out and then are not needed. I'm on MSFT Azure right now for storage, webhosting and a couple of other things but my use is still under the 2 TB level.
I know gamers who use powerful desktop gaming machines. I am not a gamer, but I have 2 desktop machines. I only use a tablet for reading, and my smartphone when I am away from home.
I far prefer a desktop with large screens and a mechanical keyboard to read and type.
I thought the NVIDIA GPUs use CUDA so that they are able to accelerate any code when run via CUDA to parallelize the calculations. While overkill for using office applications, I do notice that several applications use the graphics card GPU when running, so I imagine a much more powerful GPU would be good when running code that can be parallelized rather than running loops in serial mode.
Admittedly, I am somewhat old and also old school.
"Admittedly, I am somewhat old and also old school."
Me too!! I learned Fortran programming on a mainframe when I was an undergraduate. When I was a post-doc at Cornell, we had a DEC PDP-11 that had to be booted every morning using punched paper tape and toggle switches. Most of what we did in the lab was in BASIC. I learned C on my own and now use Python to fool around with. Most of my work can easily be done in Excell.
That is wrong. Nobody's work can be done in Excel, because Excel is **not** debuggable at all.
Only spreadsheets that have been replicated by each of five different people working in five different cities with zero channels of communication between them can be trusted:
> Alan Goldhammer: '"Admittedly, I am somewhat old and also old school." Me too!! I learned Fortran programming on a mainframe when I was an undergraduate. When I was a post-doc at Cornell, we had a DEC PDP-11 that had to be booted every morning using punched paper tape and toggle switches. Most of what we did in the lab was in BASIC. I learned C on my own and now use Python to fool around with. Most of my work can easily be done in Excel...
"That is wrong. Nobody's work can be done in Excel, because Excel is **not** debuggable at all."
I will have to contact my former neighbor, Cosma Shalizi, to weigh in on this!!!!😉 I've been using Excel to do discounted cash flow analyses of stocks for a number of years now (as well as some other things). Are you saying that this is just wrong? Inquiring minds want to know.
This is super weird. Derek Thompson's newest podcast that dropped today is about the transcontinental railroad and features Richard White!!! I am prescient.
LLM summary of railroads book seems pretty good, but what are realistic alternative "better" scenarios for how that endeavor could have played out? (As in card game, not mining)
(And what are its lessons for llms today, besides "lots of people aren't going to be winners in the short run")
Thanks for putting words tom my nagging sense: "For now, the boom props up the broader economic expansion as well, offsetting the drag on the economy created by the random chaos-monkey actions of the White House."
But what if the chaos-monkeys are just getting started?
What if the chaos-monkeys leap to the perhaps not unjustified conclusion that the only thing keeping them in power is more chaos?
In that case the window of opportunity for AI to produce investment justifying returns is closing fast.
An AI investment bust brought forward by chaos-monkey policies WITH chaos-monkeys in charge is a scary thought indeed.
Re: where green shoots might emerge - considering the "attention-hacking" economy and AI-slop that you mention, I think we can view attention-hacking platform algorithms as a response to an sort of Malthusian content/attention environment where "The power of [content] is indefinitely greater than the power in the earth to produce [attention] for [content]" - thus attention needs to be "rationed" by attention hacking algorithms, producing all of the pernicious effects we're aware of. I think LLMs clearly have the capacity to become what the industrial revolution was for Malthus -- that is, they materially enhance the attentional bandwidth of humanity to absorb a given amount of digital content by "compressing" huge amounts of content into something digestible (I have in mind here the analogy you've made between LLMs and Hayekian markets + James Scott's state bureaucracies). While I still certainly don't think the profitability of this is likely to justify hundreds of billions in CapEx, I think LLMs can be a useful solution to replace attention-hacking curation algorithms as a means of alleviating the Malthusian explosion in the content/attention ratio (ironically challenging the very model which has given the likes of Meta enough cash flow to pursue this CapEx in the first place).
Yes. Constructing reliable information-attention butlers to stand between you and the infosphere in order to allow you to manage the repeated tsunamis of information overload plus keep Mark Zuckerberg and Rupert Murdoch from hacking your brain for profit—getting you to glue your eyeballs to the screen and making you miserable just to sell a few ads.
> Donovan Berry: 'Re: where green shoots might emerge - considering the "attention-hacking" economy and AI-slop that you mention, I think we can view attention-hacking platform algorithms as a response to an sort of Malthusian content/attention environment where "The power of [content] is indefinitely greater than the power in the earth to produce [attention] for [content]" - thus attention needs to be "rationed" by attention hacking algorithms, producing all of the pernicious effects we're aware of. I think LLMs clearly have the capacity to become what the industrial revolution was for Malthus -- that is, they materially enhance the attentional bandwidth of humanity to absorb a given amount of digital content by "compressing" huge amounts of content into something digestible (I have in mind here the analogy you've made between LLMs and Hayekian markets + James Scott's state bureaucracies). While I still certainly don't think the profitability of this is likely to justify hundreds of billions in CapEx, I think LLMs can be a useful solution to replace attention-hacking curation algorithms as a means of alleviating the Malthusian explosion in the content/attention ratio (ironically challenging the very model which has given the likes of Meta enough cash flow to pursue this CapEx in the first place)...
Lo these years, The Economist pointed out that a major early use case for new tech was typically (nods to elon) X rated. Is it also the case that societal autophagy ranks highly? I am not sure how we curb this.
Google: "its search is crippled because it swims in an ocean of SEO‑optimized content."
As a point of detail, its search is crippled because of an internal decision to make it worse, one which involved a substantial internal struggle which is well documented. Inefficient search is more profitable than efficient search: the desired optimum seems to be three searches for one useful result. Much of the slop is paying for its position, and much of the paid slop is fraudulent, of a type Google could easily identify algorithmically, should it choose to do so. Google has done to search something very like what Amazon has done to its own search algorithm (more visibly): namely, fully monetized it at the expense of the search itself. (The original thesis laying out the idea of Google noted out that this would be inevitable in an ad-based model.)
Forbes now has a slop generating offshoot that is worth more than the parent company, and which is parked under their main URL, for SEO. Google knows this but has no interest in addressing it. They rate everything bearing any version of the Forbes imprimatur as valuable.
In a different direction, here's a current view (and snapshot) of how things stand presently.
Back in the 1970s, my father and a bunch of his colleagues then working at the FTC tried to think of whether and how the corporate practice of not monetizing but rather aggressively liquidating your brand equity as a maker of high-quality goods and services by cutting all the corners you could and serving up slop. But creating a way to, legally, classify that process as false advertising, and hence a consumer protection violation, has not been something that any Neoliberal Order FTC has wanted to pursue:
> glc: Re: Google: "its search is crippled because it swims in an ocean of SEO‑optimized content."
> As a point of detail, its search is crippled because of an internal decision to make it worse, one which involved a substantial internal struggle which is well documented. Inefficient search is more profitable than efficient search: the desired optimum seems to be three searches for one useful result. Much of the slop is paying for its position, and much of the paid slop is fraudulent, of a type Google could easily identify algorithmically, should it choose to do so. Google has done to search something very like what Amazon has done to its own search algorithm (more visibly): namely, fully monetized it at the expense of the search itself. (The original thesis laying out the idea of Google noted out that this would be inevitable in an ad-based model.)
> Forbes now has a slop generating offshoot that is worth more than the parent company, and which is parked under their main URL, for SEO. Google knows this but has no interest in addressing it. They rate everything bearing any version of the Forbes imprimatur as valuable.
> In a different direction, here's a current view (and snapshot) of how things stand presently.
A day late to comment but how does the thinking of today’s post change when you consider what you’ve previously written about specialized small language models? Iirc you, Brad, made the point that LLMs face some limitations and that having smaller more context appropriate models running locally on your device would be necessary. That certainly paints a different picture than the gigantic data centers everywhere story.
This may well be wrong. But in many—most?—ChatBot use cases, one wants not a SubTuring simulacrum of a friend or a software pet but rather a Natural-Language Interface to some structured source of data: for example, what goes into a tagine and how to best cook it. The GPT LLM part then exists to be the natural-language interface. And behind that there is some store of structured or unstructured data that is to be searched, the best search results to be then sent to the GPT LLM, and then used in RAG—Retrieval-Augmented Generation.
Anything more in the GPT LLM than simply basic linguistic fluency is then a source of potential error rather than a plus.
If this view of mine is correct—and it is probably not—then the end point has (a) everyone running a small LLM created and curated by Apple or Google on-device on their smartphones for an NLI to structured and unstructured databases, (b) the GPT LLMs of the data centers focused on being SubTuring friends and software pets, & (c) a huge amount of excess data-center capacity being used to train larger and larger models that are not much wanted, and so looking for something else to do.
Basically, my view is thus that Apple has the right idea about what to do here, even though it has overpromised and failed to deliver what it needs to deliver:
> James: 'A day late to comment but how does the thinking of today’s post change when you consider what you’ve previously written about specialized small language models? Iirc you, Brad, made the point that LLMs face some limitations and that having smaller more context appropriate models running locally on your device would be necessary. That certainly paints a different picture than the gigantic data centers everywhere story...
"Programmer copilots like GitHub Copilot boost output for the already-productive if not yet for experts"
You have made this claim more than once, but I think you are too credulous. I would like to raise 3 points.
1. At least one study has found that programmers self-report significant productivity improvements in situations where measured productivity declines.
2. I use an AI programming assistant at work and I am required to report productivity gains regularly. I am not allowed to report negative gains. So at least some surveys are based on truncated data.
3. I have a friend who teaches practical programming courses at university. He tells me that new students are very bullish on AI, smugly do their assignments with AI, are always discussing and comparing various AI tools. But once they start getting jobs, they gradually stop using AI. This is long before the "expert" stage.
This is not to say that AI assistants are useless; I think they are suitable front ends for scripting languages like Python to write throw-away data munging programs, for example. But that is far short of the claims being made for them.
Not too put too fine a point on it, the US is one big grift. The President is a notorious con artist, who has nonetheless secured near-majorities of the vote three times in a row, and is busy looting the Treasury. The valuations of AI hyperscalers look restrained compared to that of Tesla, a company with declining sales valued at 100 times earnings. And the entire financial corporate class has signed on to crypto, an utterly valueless set of notional assets with a value in the trillions. Sooner or later this will stop, but who can say when? There's a lot of ruin in a country as big as the US.
I cannot say that any of this is wrong:
> John Quiggin: 'Not too put too fine a point on it, the US is one big grift. The President is a notorious con artist, who has nonetheless secured near-majorities of the vote three times in a row, and is busy looting the Treasury. The valuations of AI hyperscalers look restrained compared to that of Tesla, a company with declining sales valued at 100 times earnings. And the entire financial corporate class has signed on to crypto, an utterly valueless set of notional assets with a value in the trillions. Sooner or later this will stop, but who can say when? There's a lot of ruin in a country as big as the US...
"The obvious parallel, of course, is in the stories we tell of the California Gold Rush: miners rarely struck it rich, but sellers of Levi’s, shovels, and whiskey did."
Hmmmm, I think a more appropriate parallel is the building out of the US railroad system in the second half of the 19th century. Richard White's wonderful book, "Railroaded: The Transcontinentals and the Making of Modern America" is required reading on this. Lots of cautionary lessons, though Stanford University was one positive result.
I believe the "promises" of AI are overblown and will have an impact on only a few employment sectors. I continue to comment on a variety of Substacks, that nobody is focusing on jobs that will never be replaced by AI. This is a far more useful thought experiment than worrying about the build out of lots of cloud storage centers that most certainly will be vacated in ten years (maybe sooner).
I, too, find the railroads along with the telegraph appropriate analogies. It is worth noting how important the two together are to their eventual impact. Likewise, the automobile and electricity. Part of my skepticism about the eventual size and scope of the AI transformation is that it does not have a partner to dance alongside it, creating opportunities.
And fiber, around 2000.
But unlike railway tracks and even fiber-optic cable, those GPUs can be sold off in units to willing buyers as cents on teh dollar, allowing their processing pipeline power to be used for local applications in businesses and consumer home computing.
Apollo did produce some spinoffs, but idk if they were sufficient compensation for the costs of Apollo. Transistor technology used by the space program probably was important in getting that technology used that supplanted valves. Maybe it was even important in pushing forward microelectronics, idk. Will data centers dedicated to AI be useful technology that is repurposed for general internet use, or even supercomputers, or will they become obsolete and the individual servers and boards auctioned off cheaply?
[I remember when older Sun servers became available as cheap computers for the home user. OTOH, I never saw Silicon Graphics computers ever become available, nor cheap "Lisp machines" or thin Java clients.
How many home users even need a computer? Most everyone uses just phones and tablets for the vast majority of their needs. Maybe I am a special case as I build my own workstation for home use (photography, financial analysis and writing). I certainly don't need a top GPU for these uses, and it is really on gamers and video editors that do. I don't think these NVidia GPUs can be repurposed at all. Even crypto mining of Bitcoin is nearing the end as Satoshi designed it so that is another use that is disappearing. What I don't know is what happens to all the data centers that get built out and then are not needed. I'm on MSFT Azure right now for storage, webhosting and a couple of other things but my use is still under the 2 TB level.
I know gamers who use powerful desktop gaming machines. I am not a gamer, but I have 2 desktop machines. I only use a tablet for reading, and my smartphone when I am away from home.
I far prefer a desktop with large screens and a mechanical keyboard to read and type.
I thought the NVIDIA GPUs use CUDA so that they are able to accelerate any code when run via CUDA to parallelize the calculations. While overkill for using office applications, I do notice that several applications use the graphics card GPU when running, so I imagine a much more powerful GPU would be good when running code that can be parallelized rather than running loops in serial mode.
Admittedly, I am somewhat old and also old school.
"Admittedly, I am somewhat old and also old school."
Me too!! I learned Fortran programming on a mainframe when I was an undergraduate. When I was a post-doc at Cornell, we had a DEC PDP-11 that had to be booted every morning using punched paper tape and toggle switches. Most of what we did in the lab was in BASIC. I learned C on my own and now use Python to fool around with. Most of my work can easily be done in Excell.
That is wrong. Nobody's work can be done in Excel, because Excel is **not** debuggable at all.
Only spreadsheets that have been replicated by each of five different people working in five different cities with zero channels of communication between them can be trusted:
> Alan Goldhammer: '"Admittedly, I am somewhat old and also old school." Me too!! I learned Fortran programming on a mainframe when I was an undergraduate. When I was a post-doc at Cornell, we had a DEC PDP-11 that had to be booted every morning using punched paper tape and toggle switches. Most of what we did in the lab was in BASIC. I learned C on my own and now use Python to fool around with. Most of my work can easily be done in Excel...
"That is wrong. Nobody's work can be done in Excel, because Excel is **not** debuggable at all."
I will have to contact my former neighbor, Cosma Shalizi, to weigh in on this!!!!😉 I've been using Excel to do discounted cash flow analyses of stocks for a number of years now (as well as some other things). Are you saying that this is just wrong? Inquiring minds want to know.
This is super weird. Derek Thompson's newest podcast that dropped today is about the transcontinental railroad and features Richard White!!! I am prescient.
LLM summary of railroads book seems pretty good, but what are realistic alternative "better" scenarios for how that endeavor could have played out? (As in card game, not mining)
(And what are its lessons for llms today, besides "lots of people aren't going to be winners in the short run")
Thanks for putting words tom my nagging sense: "For now, the boom props up the broader economic expansion as well, offsetting the drag on the economy created by the random chaos-monkey actions of the White House."
But what if the chaos-monkeys are just getting started?
What if the chaos-monkeys leap to the perhaps not unjustified conclusion that the only thing keeping them in power is more chaos?
In that case the window of opportunity for AI to produce investment justifying returns is closing fast.
An AI investment bust brought forward by chaos-monkey policies WITH chaos-monkeys in charge is a scary thought indeed.
I wonder about the huge demands on limited water resources especially if there is overbuilding.
Re: where green shoots might emerge - considering the "attention-hacking" economy and AI-slop that you mention, I think we can view attention-hacking platform algorithms as a response to an sort of Malthusian content/attention environment where "The power of [content] is indefinitely greater than the power in the earth to produce [attention] for [content]" - thus attention needs to be "rationed" by attention hacking algorithms, producing all of the pernicious effects we're aware of. I think LLMs clearly have the capacity to become what the industrial revolution was for Malthus -- that is, they materially enhance the attentional bandwidth of humanity to absorb a given amount of digital content by "compressing" huge amounts of content into something digestible (I have in mind here the analogy you've made between LLMs and Hayekian markets + James Scott's state bureaucracies). While I still certainly don't think the profitability of this is likely to justify hundreds of billions in CapEx, I think LLMs can be a useful solution to replace attention-hacking curation algorithms as a means of alleviating the Malthusian explosion in the content/attention ratio (ironically challenging the very model which has given the likes of Meta enough cash flow to pursue this CapEx in the first place).
Yes. Constructing reliable information-attention butlers to stand between you and the infosphere in order to allow you to manage the repeated tsunamis of information overload plus keep Mark Zuckerberg and Rupert Murdoch from hacking your brain for profit—getting you to glue your eyeballs to the screen and making you miserable just to sell a few ads.
I find myself thinking here of Cora <http://cora.computer> from Every <http://every.to>:
> Donovan Berry: 'Re: where green shoots might emerge - considering the "attention-hacking" economy and AI-slop that you mention, I think we can view attention-hacking platform algorithms as a response to an sort of Malthusian content/attention environment where "The power of [content] is indefinitely greater than the power in the earth to produce [attention] for [content]" - thus attention needs to be "rationed" by attention hacking algorithms, producing all of the pernicious effects we're aware of. I think LLMs clearly have the capacity to become what the industrial revolution was for Malthus -- that is, they materially enhance the attentional bandwidth of humanity to absorb a given amount of digital content by "compressing" huge amounts of content into something digestible (I have in mind here the analogy you've made between LLMs and Hayekian markets + James Scott's state bureaucracies). While I still certainly don't think the profitability of this is likely to justify hundreds of billions in CapEx, I think LLMs can be a useful solution to replace attention-hacking curation algorithms as a means of alleviating the Malthusian explosion in the content/attention ratio (ironically challenging the very model which has given the likes of Meta enough cash flow to pursue this CapEx in the first place)...
How much of San Francisco's 21st C. character is explained by its "Gold rush supplier town" origins?
Lo these years, The Economist pointed out that a major early use case for new tech was typically (nods to elon) X rated. Is it also the case that societal autophagy ranks highly? I am not sure how we curb this.
Google: "its search is crippled because it swims in an ocean of SEO‑optimized content."
As a point of detail, its search is crippled because of an internal decision to make it worse, one which involved a substantial internal struggle which is well documented. Inefficient search is more profitable than efficient search: the desired optimum seems to be three searches for one useful result. Much of the slop is paying for its position, and much of the paid slop is fraudulent, of a type Google could easily identify algorithmically, should it choose to do so. Google has done to search something very like what Amazon has done to its own search algorithm (more visibly): namely, fully monetized it at the expense of the search itself. (The original thesis laying out the idea of Google noted out that this would be inevitable in an ad-based model.)
Forbes now has a slop generating offshoot that is worth more than the parent company, and which is parked under their main URL, for SEO. Google knows this but has no interest in addressing it. They rate everything bearing any version of the Forbes imprimatur as valuable.
In a different direction, here's a current view (and snapshot) of how things stand presently.
https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
I don't find it that sort of thing valuable other than as a part of an ongoing conversation, but there it is.
Back in the 1970s, my father and a bunch of his colleagues then working at the FTC tried to think of whether and how the corporate practice of not monetizing but rather aggressively liquidating your brand equity as a maker of high-quality goods and services by cutting all the corners you could and serving up slop. But creating a way to, legally, classify that process as false advertising, and hence a consumer protection violation, has not been something that any Neoliberal Order FTC has wanted to pursue:
> glc: Re: Google: "its search is crippled because it swims in an ocean of SEO‑optimized content."
> As a point of detail, its search is crippled because of an internal decision to make it worse, one which involved a substantial internal struggle which is well documented. Inefficient search is more profitable than efficient search: the desired optimum seems to be three searches for one useful result. Much of the slop is paying for its position, and much of the paid slop is fraudulent, of a type Google could easily identify algorithmically, should it choose to do so. Google has done to search something very like what Amazon has done to its own search algorithm (more visibly): namely, fully monetized it at the expense of the search itself. (The original thesis laying out the idea of Google noted out that this would be inevitable in an ad-based model.)
> Forbes now has a slop generating offshoot that is worth more than the parent company, and which is parked under their main URL, for SEO. Google knows this but has no interest in addressing it. They rate everything bearing any version of the Forbes imprimatur as valuable.
> In a different direction, here's a current view (and snapshot) of how things stand presently.
> <https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/>
> I don't find it that sort of thing valuable other than as a part of an ongoing conversation, but there it is.
>>consider just how gonzo the AI-economy truly is<<
Professor DeLong is out front on the AI discussion & one reason I'm a happy subscriber...
I also enjoy the comments section here - Thanks everybody
A day late to comment but how does the thinking of today’s post change when you consider what you’ve previously written about specialized small language models? Iirc you, Brad, made the point that LLMs face some limitations and that having smaller more context appropriate models running locally on your device would be necessary. That certainly paints a different picture than the gigantic data centers everywhere story.
This may well be wrong. But in many—most?—ChatBot use cases, one wants not a SubTuring simulacrum of a friend or a software pet but rather a Natural-Language Interface to some structured source of data: for example, what goes into a tagine and how to best cook it. The GPT LLM part then exists to be the natural-language interface. And behind that there is some store of structured or unstructured data that is to be searched, the best search results to be then sent to the GPT LLM, and then used in RAG—Retrieval-Augmented Generation.
Anything more in the GPT LLM than simply basic linguistic fluency is then a source of potential error rather than a plus.
If this view of mine is correct—and it is probably not—then the end point has (a) everyone running a small LLM created and curated by Apple or Google on-device on their smartphones for an NLI to structured and unstructured databases, (b) the GPT LLMs of the data centers focused on being SubTuring friends and software pets, & (c) a huge amount of excess data-center capacity being used to train larger and larger models that are not much wanted, and so looking for something else to do.
Basically, my view is thus that Apple has the right idea about what to do here, even though it has overpromised and failed to deliver what it needs to deliver:
> James: 'A day late to comment but how does the thinking of today’s post change when you consider what you’ve previously written about specialized small language models? Iirc you, Brad, made the point that LLMs face some limitations and that having smaller more context appropriate models running locally on your device would be necessary. That certainly paints a different picture than the gigantic data centers everywhere story...
TMSC, I believe.