Jason Furman on how to properly worry about the U.S. national debt; Paul Krugman on the end of the claim that jobless Americans are “zero marginal product workers”; very briefly noted; Jan Hatzius of Goldman Sachs says: soft landing achieved; PerplexityAI is a lying liar, in many senses; services-led economic development appears a forlorn hope; be sure to count all of the costs of geoengineering when you think about it!; right-wing global-warming denier hyena-hack Holman Jenkins is truly deranged in his takes on Elon Musk; Paul Krugman says: soft landing achieved; & Lucidity on "AI" investments considered harmful; the great Agrarian-Age vine-&-figtree shortage; reviewing Blinder's "Monetary & Fiscal History"; Joe Wiesenthal praises my "Slouching Towards Utopia"; Isaac Asimov on dominionist ethnonationalism; & weekly briefly noted for 2024-06-15 Sa...
Strictly speaking, one should NEVER worry about "debt." Worry about the suboptimal spending and taxing policies that produced it. If [which I do not believe for a minute] the spending taxing policies were optimal, the debt is nothing to worry about.
"Tasting menus" are exactly why I hate Michelin-starred restaurants. If I have to jump through hoops just to score a reservation and pay a fortune for the privilege then by the time I get there I want the experience to be about me, not the fucking chef. I'll order what I feel like eating à la carte, thank you very much, not what the chef feels like feeding me off a glorified prix fixe.
People who worry about downwardly sticky wages ought to worry instead about downwardly sticky lease payments. Evicting a client who can pay the marginal cost of keeping a property occupied makes zero sense. Lessors and their financiers need a better business model. [And cities need to have land use restriction and building codes that are flexible enough to allow property owners to respond to changes in demand. Their business model has to change, too.]
Re: Gary Marcus: "Partial Regurgitation & how LLMs really work: ‘What the LLM does is more akin to what some high school students do when they plagiarize: change a few words here or there, while still sticking close to the original."
I think LLMs are doing a little more than that, but at this time, he is absolutely correct that they act like fresh students. I am less worsed about the attempts to avoid plagiarism and more worried that they don't understand context. LLMs also seem to have "high self esteem" (anthropomorphism) and will insist they are correct when they are "hallucinating" (i.e. BSing and lying.) To be fair, adults do the same - from parrotting ideologies, and their memes, to failing to understand context like humor, irony, etc. Personally, I don't want verbal sugar but rather correct information and task completion. Otherwise it becomes "why don't I just do this myself". I have already turned off Google's AI insert at the top of the searches because it cannot guarantee reliability and I don't want to be suckered into relying on it when I want to save time rather than working through the page results. I can be quite sure Google will not offer me a search option to remove adverts from the top of the results page whatever I might ask of an AI.
But with all this attention on LLMs, it should be noted that AI encompasses a lot more than this approach. LLMs may have grabbed the attention, but this was unfortunate as AI, even simple ML approaches offer a lot of value for relatively little computational effort. IMO, LLMs are best used as interfaces, but should then used other algorithms to extract the needed results which can then be "summarized" if desired.
The lack of real AI applications is so reminiscent of the early days of personal computing. Manufacturers were producing various configurations for computers, but they had no software to run on them.. Apple got it first big success with Visicalc which created the demand for spreadsheets allowing managers to do basic, but repetitive calculations without begging the IT team handling the corporate mainframes.
These LLMs do have uses, but whether they will translate to consumer use, idk. I keep getting invites to free courses to use LLMs and even coding applications that integrate LLMs with data, but this is for developers, not the majority of consumers. I do want a good speech-to-speech universal translator, but I don't think it needs an LLM implementation on my phone. I certainly don't want a "clippy" trying to do things rather than letting me learn how (although I find the Android interface on my phone extremely annoying and think that maybe voice commands might simplify the use of that muddled interface design).
Bootom line is that there needs to be the classic "killer app" that can only be achieved by AI to make it worth paying for LLMs on the computer of choice. Incremental improvements hardly seem worth buying more powerful hardware or for tech giants to build more server farms.
[I would note that just a few years ago there was a idea to reduce compute resources by stripping down software bloat, especial on web pages, but this seems to have died down. Certainly my old iPad v2 fails to load an ever larger number of web pages and cannot support newer versions of iTunes apps. I dread the day that Apple devices will require AI on every device. :-(
Apple has a plan for using MAMLMs—not LLMs, save perhaps as a natural language front-end to a semantic index—on-device to actually do useful large-scale classification things... Brad
As you know, that "Apple Intelligence" will only be offered on newer devices that supprt it. Time will tell how useful this will be. [BTW, have you found a compelling use for their Vison Pro AR/VR headset?] While Apple had/has a reputation of building useful hardware and software that can fulfill Job's "bicycle for the mind", in recent years, since Jony Ives departed, the IOS interface has become degraded by "featuritis". Whether this mentality will result in "AI" being unnecessarily added to applications or truly useful, remains to be seem. For example, is seach on my stored photos with feature [X] all that useful? If I had a very large database of images that might save time, but I don't. If I had a large collection of images, they would likely have already been stored in useful directories to ease access. A key test will be if the "AI" search could extract photos by "Find images of [X] place from my holiday snaps taken in location [Y]. " IOW, context sentivity that might well defy classification. [Microsoft doing the same would result in their usual pratfalls. Their rapid about face on "Recall" was faster than Liz Truss's exit as UK PM ;-) ]
Chinese stagnation: Still dangerous if the US also stagnates with high deficits, timid merit-based immigration, NIMBY-ridden cities, NEPA-ridden infrastructure and zero energy projects, and high costs of reducing CO2 emissions.
American Political Economy: All a big misunderstanding. The Fed was just not cranking out enough inflation (and promising to keep doing it) to maximize real income growth. I do not expect Republicans to understand this. Crisis and stagnation fire up their base. But for Democrats, it was inexcusable.
Strictly speaking, one should NEVER worry about "debt." Worry about the suboptimal spending and taxing policies that produced it. If [which I do not believe for a minute] the spending taxing policies were optimal, the debt is nothing to worry about.
https://thomaslhutcheson.substack.com/p/debtpocalypse
"Tasting menus" are exactly why I hate Michelin-starred restaurants. If I have to jump through hoops just to score a reservation and pay a fortune for the privilege then by the time I get there I want the experience to be about me, not the fucking chef. I'll order what I feel like eating à la carte, thank you very much, not what the chef feels like feeding me off a glorified prix fixe.
People who worry about downwardly sticky wages ought to worry instead about downwardly sticky lease payments. Evicting a client who can pay the marginal cost of keeping a property occupied makes zero sense. Lessors and their financiers need a better business model. [And cities need to have land use restriction and building codes that are flexible enough to allow property owners to respond to changes in demand. Their business model has to change, too.]
Re: Gary Marcus: "Partial Regurgitation & how LLMs really work: ‘What the LLM does is more akin to what some high school students do when they plagiarize: change a few words here or there, while still sticking close to the original."
I think LLMs are doing a little more than that, but at this time, he is absolutely correct that they act like fresh students. I am less worsed about the attempts to avoid plagiarism and more worried that they don't understand context. LLMs also seem to have "high self esteem" (anthropomorphism) and will insist they are correct when they are "hallucinating" (i.e. BSing and lying.) To be fair, adults do the same - from parrotting ideologies, and their memes, to failing to understand context like humor, irony, etc. Personally, I don't want verbal sugar but rather correct information and task completion. Otherwise it becomes "why don't I just do this myself". I have already turned off Google's AI insert at the top of the searches because it cannot guarantee reliability and I don't want to be suckered into relying on it when I want to save time rather than working through the page results. I can be quite sure Google will not offer me a search option to remove adverts from the top of the results page whatever I might ask of an AI.
But with all this attention on LLMs, it should be noted that AI encompasses a lot more than this approach. LLMs may have grabbed the attention, but this was unfortunate as AI, even simple ML approaches offer a lot of value for relatively little computational effort. IMO, LLMs are best used as interfaces, but should then used other algorithms to extract the needed results which can then be "summarized" if desired.
Yes: MAMLMs—large-scale classification and regression analyses on enormous databases of huge dimensionality...
Economics of current MAMLM AI.
The lack of real AI applications is so reminiscent of the early days of personal computing. Manufacturers were producing various configurations for computers, but they had no software to run on them.. Apple got it first big success with Visicalc which created the demand for spreadsheets allowing managers to do basic, but repetitive calculations without begging the IT team handling the corporate mainframes.
These LLMs do have uses, but whether they will translate to consumer use, idk. I keep getting invites to free courses to use LLMs and even coding applications that integrate LLMs with data, but this is for developers, not the majority of consumers. I do want a good speech-to-speech universal translator, but I don't think it needs an LLM implementation on my phone. I certainly don't want a "clippy" trying to do things rather than letting me learn how (although I find the Android interface on my phone extremely annoying and think that maybe voice commands might simplify the use of that muddled interface design).
Bootom line is that there needs to be the classic "killer app" that can only be achieved by AI to make it worth paying for LLMs on the computer of choice. Incremental improvements hardly seem worth buying more powerful hardware or for tech giants to build more server farms.
[I would note that just a few years ago there was a idea to reduce compute resources by stripping down software bloat, especial on web pages, but this seems to have died down. Certainly my old iPad v2 fails to load an ever larger number of web pages and cannot support newer versions of iTunes apps. I dread the day that Apple devices will require AI on every device. :-(
Apple has a plan for using MAMLMs—not LLMs, save perhaps as a natural language front-end to a semantic index—on-device to actually do useful large-scale classification things... Brad
As you know, that "Apple Intelligence" will only be offered on newer devices that supprt it. Time will tell how useful this will be. [BTW, have you found a compelling use for their Vison Pro AR/VR headset?] While Apple had/has a reputation of building useful hardware and software that can fulfill Job's "bicycle for the mind", in recent years, since Jony Ives departed, the IOS interface has become degraded by "featuritis". Whether this mentality will result in "AI" being unnecessarily added to applications or truly useful, remains to be seem. For example, is seach on my stored photos with feature [X] all that useful? If I had a very large database of images that might save time, but I don't. If I had a large collection of images, they would likely have already been stored in useful directories to ease access. A key test will be if the "AI" search could extract photos by "Find images of [X] place from my holiday snaps taken in location [Y]. " IOW, context sentivity that might well defy classification. [Microsoft doing the same would result in their usual pratfalls. Their rapid about face on "Recall" was faster than Liz Truss's exit as UK PM ;-) ]
On Ludicity. That was awesome! Thanks for posting.
Chinese stagnation: Still dangerous if the US also stagnates with high deficits, timid merit-based immigration, NIMBY-ridden cities, NEPA-ridden infrastructure and zero energy projects, and high costs of reducing CO2 emissions.
American Political Economy: All a big misunderstanding. The Fed was just not cranking out enough inflation (and promising to keep doing it) to maximize real income growth. I do not expect Republicans to understand this. Crisis and stagnation fire up their base. But for Democrats, it was inexcusable.