“They were careless people, - they smashed up things and creatures and then retreated back into their money or their vast carelessness or whatever it was that kept them together, and let other people clean up the mess they had made.” - F. Scott Fitzgerald, The Great Gatsby
I marvel at the idea of partnering with the government and transforming company support into a “patriotic duty.” “What’s good for OpenAI is good for the country.”
The government taking a stake in what is really a very speculative venture will make the whole thing too big to fail, setting up a huge, budget breaking bailout.
Ed Zitron (Google him) makes a strong case that when the LLM bubble bursts, there's nothing left to bail out. The Fed could buy all the data centers, but they will be worth near zero if LLMs don't work. The US auto industry and banks had future value to bail out.
Actual value might not matter to Trump as much as keeping the bubble intact (enough) with taxpayer money so he can continue his own internet/crypto related scams.
Great essay, Mike. It pulls together so many of the dynamics shaping the AI phenomenon.
One thing occurred to me: the thing we’ve gotten better at since the tulip bulb bubble burst in 1637 is building ever-more-complex truthy narratives to justify larger and larger bubbles. The housing bubble burst when everyone started realizing all the swaps and securitized mortgages, the most sophisticated narrative to justify a burst bubble to date, were essentially worthless. But this new AI narrative plays both sides against the middle and creates circular firing/capital squads where hardware manufacturers pay for their own product. And it’s all based on what most likely is an impossible future state of AGI. I’m getting the exact same feeling I had in 2006 when I got an inkling of the size of the CDO and credit default swap markets. I don’t quite grasp the scale or the vapor it’s all built on, but it’s going to be a spectacular implosion when it happens.
A sprawling essay. I'm not sure I agree with every micro-conclusion, but it's impressive and amazingly on-point nonetheless. For my money, this line (a rephrasing of the famous Chomsky/Herman "Propaganda Model" of Media Literacy) is worth the price of admission:
"Institutional influence rarely requires the vulgarity of a script."
Not sure that current AI is up to the level of coding you assume. I've had my own trivial problems with chatGPT assuring me it can code something: it couldn't. This was IMHO trivial stuff. My wife tried to get Copilot to draw an 8 pointed star. The prompt was "2 identical squares, one on top of the other. Rotate the top square through 45 degrees. You should have an 8 pointed star. Draw it." Copilot drew a 6 pointed star. When the error was pointed out it drew a 5 pointed star. Impressive stuff. She was teaching software requirements at the time, and the activity was a "must do" principally because Microsoft paid the company to talk up Copilot. It wasn't recorded what the students thought of it. Let's just say they were not fans. Meanwhile Microsoft are tiptoeing back from Copilot. Which is really chatGPT anyway.
What I hear from those working with AI code is that the code is a mess and takes more time to fix than it would have to just get humans to write the code.
I think it has uses, but I'm sceptical about the claims. This is not helped by several of the poster boys (and they do all seem to be, ahem, boys) for AI being known liars, disingenuous, and mad. I can put names to the categories if you like.
The fact Uber used all their tokens for a year in 3-4 months followed by CTO Praveen Neppalli Naga going viral after telling The Information this, adding that higher token usage didn't translate into a proportional increase in useful consumer features and identifying the ROI from AI was very difficult does nothing to stem my fears that this is a ponzi scheme (possibly unintended) and will end in tears. Meanwhile of course the Chinese have created a very capable AI with much less expenditure, and where it's easier to identify the ROI. Rumour has it many American businesses in need of some AI have been buying into Deepseek simply because the token price is comparatively low. I'm glad I'm out of the software world.
There was a recent article in The Atlantic about AI Data Centers causing Silicon Valley to meet the constraints of the real world vs the resource light hyper scaling of the digital sphere. It’s a race to build enough to turn a profit before the whole thing collapses. This makes the public backlash an even bigger threat than people realize
This week, I've found AI to be an extraordinary tool for providing structure, identifying weaknesses, and broadening perspectives. But the moment when a genuinely new idea emerges—the kind that reorganizes one's understanding of a problem and opens a path that did not previously exist—still seems to arise in the human mind. AI accompanies that process; it does not originate it.
perversly it's funny they are making the case for sovereign wealth fund and also giving the government (local to federal) and people (could be unions) a seat at the table.
Isn’t the golden goose in all this the general public, aka “consumers”? If too many people lose their jobs - and it seems the relatively comfortably off middle class (and its disposable income) is the sector being hardest hit atm, then who will be left to actually buy the stuff that’s being so evermore efficiently made?
Agreed 100%, but there is no evidence that LLMs are actually impacting the middle class (or really anyone). That's part of the illogical Big Lie they are telling. My middle/working class friends are being crushed by inflation, not LLMs.
Not sure that current AI is up to the level of coding you assume. I've had my own trivial problems with chatGPT assuring me it can code something: it couldn't. This was IMHO trivial stuff. My wife tried to get Copilot to draw an 8 pointed star. The prompt was "2 identical squares, one on top of the other. Rotate the top square through 45 degrees. You should have an 8 pointed star. Draw it." Copilot drew a 6 pointed star. When the error was pointed out it drew a 5 pointed star. Impressive stuff. She was teaching software requirements at the time, and the activity was a "must do" principally because Microsoft paid the company to talk up Copilot. It wasn't recorded what the students thought of it. Let's just say they were not fans. Meanwhile Microsoft are tiptoeing back from Copilot. Which is really chatGPT anyway.
What I hear from those working with AI code is that the code is a mess and takes more time to fix than it would have to just get humans to write the code.
I think it has uses, but I'm sceptical about the claims. This is not helped by several of the poster boys (and they do all seem to be, ahem, boys) for AI being known liars, disingenuous, and mad. I can put names to the categories if you like.
The fact Uber used all their tokens for a year in 3-4 months followed by CTO Praveen Neppalli Naga going viral after telling The Information this, adding that higher token usage didn't translate into a proportional increase in useful consumer features and identifying the ROI from AI was very difficult does nothing to stem my fears that this is a ponzi scheme (possibly unintended) and will end in tears. Meanwhile of course the Chinese have created a very capable AI with much less expenditure, and where it's easier to identify the ROI. Rumour has it many American businesses in need of some AI have been buying into Deepseek simply because the token price is comparatively low. I'm glad I'm out of the software world.
“They were careless people, - they smashed up things and creatures and then retreated back into their money or their vast carelessness or whatever it was that kept them together, and let other people clean up the mess they had made.” - F. Scott Fitzgerald, The Great Gatsby
This makes Charles Ponzi appear to be an underachiever.
If only he'd had access to such a boondoggle....
I marvel at the idea of partnering with the government and transforming company support into a “patriotic duty.” “What’s good for OpenAI is good for the country.”
The government taking a stake in what is really a very speculative venture will make the whole thing too big to fail, setting up a huge, budget breaking bailout.
Ed Zitron (Google him) makes a strong case that when the LLM bubble bursts, there's nothing left to bail out. The Fed could buy all the data centers, but they will be worth near zero if LLMs don't work. The US auto industry and banks had future value to bail out.
Actual value might not matter to Trump as much as keeping the bubble intact (enough) with taxpayer money so he can continue his own internet/crypto related scams.
I agree on Trump's motivations, but LLMs have nothing to do with crypto scams.
Nice companion article to what Gary Marcus, Ed Zitron et al have been saying.
Great essay, Mike. It pulls together so many of the dynamics shaping the AI phenomenon.
One thing occurred to me: the thing we’ve gotten better at since the tulip bulb bubble burst in 1637 is building ever-more-complex truthy narratives to justify larger and larger bubbles. The housing bubble burst when everyone started realizing all the swaps and securitized mortgages, the most sophisticated narrative to justify a burst bubble to date, were essentially worthless. But this new AI narrative plays both sides against the middle and creates circular firing/capital squads where hardware manufacturers pay for their own product. And it’s all based on what most likely is an impossible future state of AGI. I’m getting the exact same feeling I had in 2006 when I got an inkling of the size of the CDO and credit default swap markets. I don’t quite grasp the scale or the vapor it’s all built on, but it’s going to be a spectacular implosion when it happens.
A sprawling essay. I'm not sure I agree with every micro-conclusion, but it's impressive and amazingly on-point nonetheless. For my money, this line (a rephrasing of the famous Chomsky/Herman "Propaganda Model" of Media Literacy) is worth the price of admission:
"Institutional influence rarely requires the vulgarity of a script."
Not sure that current AI is up to the level of coding you assume. I've had my own trivial problems with chatGPT assuring me it can code something: it couldn't. This was IMHO trivial stuff. My wife tried to get Copilot to draw an 8 pointed star. The prompt was "2 identical squares, one on top of the other. Rotate the top square through 45 degrees. You should have an 8 pointed star. Draw it." Copilot drew a 6 pointed star. When the error was pointed out it drew a 5 pointed star. Impressive stuff. She was teaching software requirements at the time, and the activity was a "must do" principally because Microsoft paid the company to talk up Copilot. It wasn't recorded what the students thought of it. Let's just say they were not fans. Meanwhile Microsoft are tiptoeing back from Copilot. Which is really chatGPT anyway.
What I hear from those working with AI code is that the code is a mess and takes more time to fix than it would have to just get humans to write the code.
I think it has uses, but I'm sceptical about the claims. This is not helped by several of the poster boys (and they do all seem to be, ahem, boys) for AI being known liars, disingenuous, and mad. I can put names to the categories if you like.
The fact Uber used all their tokens for a year in 3-4 months followed by CTO Praveen Neppalli Naga going viral after telling The Information this, adding that higher token usage didn't translate into a proportional increase in useful consumer features and identifying the ROI from AI was very difficult does nothing to stem my fears that this is a ponzi scheme (possibly unintended) and will end in tears. Meanwhile of course the Chinese have created a very capable AI with much less expenditure, and where it's easier to identify the ROI. Rumour has it many American businesses in need of some AI have been buying into Deepseek simply because the token price is comparatively low. I'm glad I'm out of the software world.
There was a recent article in The Atlantic about AI Data Centers causing Silicon Valley to meet the constraints of the real world vs the resource light hyper scaling of the digital sphere. It’s a race to build enough to turn a profit before the whole thing collapses. This makes the public backlash an even bigger threat than people realize
You describe a potentially world-ecompassing ourobouros that is eating itself.
It’s also a bet, they placed, on how long they think this authoritarian government will last…
Not long enough for AI to turn a profit is my guess.
This week, I've found AI to be an extraordinary tool for providing structure, identifying weaknesses, and broadening perspectives. But the moment when a genuinely new idea emerges—the kind that reorganizes one's understanding of a problem and opens a path that did not previously exist—still seems to arise in the human mind. AI accompanies that process; it does not originate it.
perversly it's funny they are making the case for sovereign wealth fund and also giving the government (local to federal) and people (could be unions) a seat at the table.
Isn’t the golden goose in all this the general public, aka “consumers”? If too many people lose their jobs - and it seems the relatively comfortably off middle class (and its disposable income) is the sector being hardest hit atm, then who will be left to actually buy the stuff that’s being so evermore efficiently made?
Agreed 100%, but there is no evidence that LLMs are actually impacting the middle class (or really anyone). That's part of the illogical Big Lie they are telling. My middle/working class friends are being crushed by inflation, not LLMs.
Sam Altman is the Black Swan.
The tell is the circular "investments." It's a tilt-a-whirl about to go off its track.
Not sure that current AI is up to the level of coding you assume. I've had my own trivial problems with chatGPT assuring me it can code something: it couldn't. This was IMHO trivial stuff. My wife tried to get Copilot to draw an 8 pointed star. The prompt was "2 identical squares, one on top of the other. Rotate the top square through 45 degrees. You should have an 8 pointed star. Draw it." Copilot drew a 6 pointed star. When the error was pointed out it drew a 5 pointed star. Impressive stuff. She was teaching software requirements at the time, and the activity was a "must do" principally because Microsoft paid the company to talk up Copilot. It wasn't recorded what the students thought of it. Let's just say they were not fans. Meanwhile Microsoft are tiptoeing back from Copilot. Which is really chatGPT anyway.
What I hear from those working with AI code is that the code is a mess and takes more time to fix than it would have to just get humans to write the code.
I think it has uses, but I'm sceptical about the claims. This is not helped by several of the poster boys (and they do all seem to be, ahem, boys) for AI being known liars, disingenuous, and mad. I can put names to the categories if you like.
The fact Uber used all their tokens for a year in 3-4 months followed by CTO Praveen Neppalli Naga going viral after telling The Information this, adding that higher token usage didn't translate into a proportional increase in useful consumer features and identifying the ROI from AI was very difficult does nothing to stem my fears that this is a ponzi scheme (possibly unintended) and will end in tears. Meanwhile of course the Chinese have created a very capable AI with much less expenditure, and where it's easier to identify the ROI. Rumour has it many American businesses in need of some AI have been buying into Deepseek simply because the token price is comparatively low. I'm glad I'm out of the software world.