Timmy. good question. Maybe the NSA will have to download the DeepSeek kit, too...
Gordon, It looks like DeepSeek has a way of recognizing the odds of an incorrect response are rising so it stops and starts over. If that's true, it should reduce hallucinations.
Per the stock market: technically, a case can be made that it was poised on a precipice of over-valuation awaiting a leaf to fall and tip it into the abyss. DeepSeek might just be that leaf.
But the "smart money" will be desperately trying to persuade the "dumb money" to not sell but rather "buy the dip," so the "smart money" can liquidate their positions and the "dumb money" will be left holding the bag as the bubble deflates. This is the usual course of bubbles popping,
This is excellent content Charles. My bet is President Trump was unaware of Deep Seek before he initiated AI STARGATE. This link explains why President Trump was pushing AI and Stargate:
Therefore Charles, it appears Charina Chou, James Manyika, and Hartmut Neven of Google Quantum AI deceived President Donald J. Trump into a "$500 billion fire drill" to save Google and Sam Altman, Larry Ellison, Bill Gates, Jensen Huang, Mark Zuckerberg and the others pushing AI onto Trump.
I think it's wise to look at the implosion of the AI "growth" fantasy holding up the entire global stock market in conjunction with the reversal of the Treasury yield inversion, which has been a 100% accurate indicator of recession, and the rise in interest rates in Japan, which blows up the yen carry trade. To Carl's point, each of these can be viewed as an omen of things falling apart--Black Swan, Gray Swan, take your pick. An 80% decline in tech stocks would simply repeat what happened in the dot-com bubble, which is eerily similar to today's hype / euphoria / over-valuation bubble. warm regards, charles
Precisely Charles. "An 80% decline in tech stocks would simply repeat what happened in the dot-com bubble, which is eerily similar to today's hype / euphoria / over-valuation bubble."
FYI Charles: This AM on the Glenn Beck radio show, Beck made the statement: "Is this our Sputnik moment?" as he explained "Deep Seek". I'll give big odds, the Glenn Beck producers read Charles Hugh Smith's Substack. Wanna bet anyone?
Carl, I think it likely that we'll be forced to proceed in the direction you outline: local and analog, as all the digitally dependent stuff will turn into bricks as spare parts are no longer available and there's no longer enough surplus income to buy replacements for everything under the sun every few years. Few see this as possible, but I see it as inevitable. warm regards, charles
Per the moon landing, I understand the skepticism engendered by any government involvement, but many nations' spacecraft have photographed the landing sites and it's possible to see them with a telescope as well:
I just ended a wide-ranging discussion of these topics with Adam Taggart of Thoughtful Money (Substack): SPECIAL REPORT: Did China's DeepSeek Just Pop The AI Stock Bubble? | Charles Hugh Smith (56 minutes)
I used to dismiss comments like this as ridiculous conspiracy theories, but now I think our government has engaged in very big lies for a very long time. I don't know a lot about space travel, but that rickety tin can they supposedly traveled in doesn't seem like it could survive a 10-mph crash, let alone a long journey through space.
Good article. The software makes hardware efficient. Bill Gates never could figure that out, or maybe he made lousy software on purpose to sell new software and anti-virus software and new hardware.
I'm glad that DeepSeek is open software. That will make a big difference. The data centers will still need their big server complexes to store data, but accessing the data with efficient software will save money, energy, and hardware space. Is it a good thing when the state has data on all of us to have efficient software to sift through that data?
Carl, I'm glad you brought up quantum computing. I too read the Foreign Affairs essay, as I've been following qubits and quantum computing advances for a decade or so.
QC strikes me as similar to fusion, where the enthusiasm for the technology far exceeds the capabilities of the technology, and glosses over the immense physical limits of these technologies.
For example, for qubits (the "gates" in the QC) to remain stable, the machine must be cooled to near absolute zero. Hmm, this doesn't sound like a cheap or easy process. It sort of sounds like controlling processes at 100 million degrees F., i.e. fusion. Both of these are extremes that cannot be understood by conventional means because they're so far from human experience and so difficult to maintain for longer than micro-seconds--or in the case of fusion, billionths of a second.
The article's authors all work in Big Tech, so of course they think encryption is the Number One Big Problem of the future, and the value proposition of QC is identifying new compounds that have medical applications.
These are problems that they intend to "solve" with brute force, as that's the value proposition of QC-- faster processing times, parallel processing, etc.
But as we're now realizing with DeepSeek, maybe brute force processing is over-rated and unnecessary--in effect, a gigantic waste of irreplaceable resources.
For unknown reasons, a comment by Simple John did not post here, so I'm copying it here as it's a topic worth pondering:
"I'd appreciate correction if I'm wrong. I believe I've read that the DeepSeek models are strongest on math and physics. In fact, aren't LLM and image generating AI playing in a universe that is immensely less specific than math and physics and thus really just playing with words and images without any real insights?"
OK, Charles here. This is very insightful, as the examples of DeepSeek (or competing tools) solving math problems are "problems" where the "correct answer" can be determined. The "answer" to the "question" *write an essay on Charles Darwin for my class assignment" has no equivalent "correct answer." The AI Bot can hallucinate a response that might pass muster if the hallucination is not too wild.
I played around with Google's free AI tools for generating podcasts, summaries, etc., and have watched developments in the video-creation space. These are examples of brute-force processing working with templates assembled from "machine learning," i.e. sampling human-generated videos, stories, essays, etc. As John noted, this kind of extrapolation of existing content is a different kind of "answer" to a different kind of "problem."
So there are video-generation tools where you enter text instructions such as "a young man is walking through a 4th of July party holding a beer," and the program generates a video clip of this scene based on its vast database of 4th of July clips, people holding a beer, etc. This is fascinating and fun, because there is no "correct answer."
But what's the value proposition here when everyone (or anyone with a keen interest) can access the same tools and generate limitless AI content? Who's going to watch all this in an Attention Economy that'd already saturated with content?
Based on my limited understanding of the many software techniques DeepSeek employs, it seems likely that these structures may well be superior ways to solve "problems" that have a testable "correct" answers.
Without going too deeply into specifics, a key concept in high-end machine learning (Google's DeepMind, etc.) is estimating the accuracy of the answer, the probabilities of various potential solutions being correct. The "wait a minute, maybe this isn't the best path, let's start over" function of DeepSeek is based on the same idea, which is extremely valuable when the "correctness" of the answer actually matters.
Brad, those are very interesting points for exploration.
I'm reminded of various stories in TechLand about small teams developing the breakthroughs while thousands of employees in the corporation generated little value beyond maintaining the status quo. So Google has 175K employees, all smart, all dedicated, but a much smaller team blew the 175K teams of MSFT, AAPL, GOOG, META et al. away. Steve Jobs famously kept the Macintosh team isolated from Corporate Apple and exerted obsessive control over the development team. He knew better than to let Corporate do the Corporate Thing to the team.
Since finance and tech dominate the economy now, the big question is: how do I make a killing by investing in DeepSeek? This is a koan because there is no "owner" of the software techniques DeepSeek has shared with the world.
To Brad's question, so what do the tech monopolies / behemoths do with their 175K "teams" now? For sure they can try to replicate DeepSeek, but does that require 175K employees? And since everyone else has access to the same concepts, techniques and approaches, then where is the scarcity value that generates revenues? There is none.
This all leads to a sobering conclusion: there is little justification for these huge headcounts going forward.
Charles is it possible to use analog for local economies to function within the division of labor, thus replacing digital? The reason I ask is today, with the announcement of DeepSeek, Nvidia Corporation lost 1/2 $trillion in market cap. What is next? $100 billion margin calls? Federal Reserve bailouts similar to TARP? Fast forward and the US dollar will default via inflation.
Local economies will be the source of restoration. Not a central government. Perhaps analog will work?
The concept makes good sense. I wasn't around in the old mainframe era, but I learned to program on the first generation of PCs, making maximum use of small memory and slow PCs. I wrote some audio and graphics programs that ran smooth and fast. Modern bloatware on vastly larger and faster computers still can't match it.
Russia had one giant advantage in the Sputnik era, and our response to Sputnik made the Russian advantage larger, not smaller. Russia trained its math and engineer types with experience and teamwork, while we trained with memorized theory. We responded by increasing the theory, NOT by switching to experience and discussion. Since then we've offshored our tech and our engineering, so we've lost both skills. Russia didn't play the free trade game and continued making its own stuff and software. Our perpetual sanctions strengthened Russia's self-sufficiency.
Russia has always assumed extremely limited resources, while we've made the machines bigger and our brains smaller. China carries on the same tradition.
So if the Big 7 don't need all those massive server farms for AI, what else might they repurpose them for? In my experience, such a resource will not just be dismantled or abandoned, but the owners will desperately seek a way to use it for something.
I am certainly no software expert, but does the DeepSeek software also compete at making deepfake images and videos? Is it less prone to hallucinations?
I will be eagerly watching for more info in the coming weeks. And watching the stock market.
I've always had a hunch that there are simpler methods to achieve high performing artificial intelligence. Of course, we should note that the chips China did this on are vastly superior than anything prior to the year 2000 lets say. And China does have decades of manufacturing experience and a literal army worth of engineers to throw at the problem. So we shouldn't say that what China did was easy. It was just easier than throwing a mountain of money at the problem. You can't fake scarcity that easily and the Mag 7 as you call them will continue to struggle while the smaller teams given freedom to try anything with their limited resources will continue to find success.
Or that mountain of money might end up working. Who knows for sure until it happens?
Timmy. good question. Maybe the NSA will have to download the DeepSeek kit, too...
Gordon, It looks like DeepSeek has a way of recognizing the odds of an incorrect response are rising so it stops and starts over. If that's true, it should reduce hallucinations.
Per the stock market: technically, a case can be made that it was poised on a precipice of over-valuation awaiting a leaf to fall and tip it into the abyss. DeepSeek might just be that leaf.
But the "smart money" will be desperately trying to persuade the "dumb money" to not sell but rather "buy the dip," so the "smart money" can liquidate their positions and the "dumb money" will be left holding the bag as the bubble deflates. This is the usual course of bubbles popping,
warm regards,
charles
This is excellent content Charles. My bet is President Trump was unaware of Deep Seek before he initiated AI STARGATE. This link explains why President Trump was pushing AI and Stargate:
https://www.foreignaffairs.com/united-states/race-lead-quantum-future-chou-manyika-neven
Therefore Charles, it appears Charina Chou, James Manyika, and Hartmut Neven of Google Quantum AI deceived President Donald J. Trump into a "$500 billion fire drill" to save Google and Sam Altman, Larry Ellison, Bill Gates, Jensen Huang, Mark Zuckerberg and the others pushing AI onto Trump.
The air is filled with flying black swans.
I think it's wise to look at the implosion of the AI "growth" fantasy holding up the entire global stock market in conjunction with the reversal of the Treasury yield inversion, which has been a 100% accurate indicator of recession, and the rise in interest rates in Japan, which blows up the yen carry trade. To Carl's point, each of these can be viewed as an omen of things falling apart--Black Swan, Gray Swan, take your pick. An 80% decline in tech stocks would simply repeat what happened in the dot-com bubble, which is eerily similar to today's hype / euphoria / over-valuation bubble. warm regards, charles
Precisely Charles. "An 80% decline in tech stocks would simply repeat what happened in the dot-com bubble, which is eerily similar to today's hype / euphoria / over-valuation bubble."
FYI Charles: This AM on the Glenn Beck radio show, Beck made the statement: "Is this our Sputnik moment?" as he explained "Deep Seek". I'll give big odds, the Glenn Beck producers read Charles Hugh Smith's Substack. Wanna bet anyone?
Time for a humor break.
My wife just received this email, offering an opportunity to buy a crypto-token being issued by DeepSeek:
"DeepSeek AI officially launched its own crypto token on the Solana blockchain today!
This is a rare opportunity to be among the first investors in a project that combines the power of blockchain and artificial intelligence."
That Deep Seek has no public shares and no connection to blockchain technology--who cares? Get in the covered wagon, we're going to the Gold Rush!
warm regards, charles
Carl, I think it likely that we'll be forced to proceed in the direction you outline: local and analog, as all the digitally dependent stuff will turn into bricks as spare parts are no longer available and there's no longer enough surplus income to buy replacements for everything under the sun every few years. Few see this as possible, but I see it as inevitable. warm regards, charles
Per the moon landing, I understand the skepticism engendered by any government involvement, but many nations' spacecraft have photographed the landing sites and it's possible to see them with a telescope as well:
https://www.skyatnightmagazine.com/advice/skills/see-apollo-landing-sites-moon
How to see all 6 Apollo landing sites on the Moon with a telescope
https://petapixel.com/2024/07/24/which-country-captured-the-best-photo-of-the-apollo-11-landing-site/
Which Country Captured the Best Photo of the Apollo 11 Landing Site?
I just ended a wide-ranging discussion of these topics with Adam Taggart of Thoughtful Money (Substack): SPECIAL REPORT: Did China's DeepSeek Just Pop The AI Stock Bubble? | Charles Hugh Smith (56 minutes)
https://youtube.com/live/nLOvJ9alKGM
Did the USA really land men on the moon?
I used to dismiss comments like this as ridiculous conspiracy theories, but now I think our government has engaged in very big lies for a very long time. I don't know a lot about space travel, but that rickety tin can they supposedly traveled in doesn't seem like it could survive a 10-mph crash, let alone a long journey through space.
Indeed....the whole thing looks absurd from 2025.
Good article. The software makes hardware efficient. Bill Gates never could figure that out, or maybe he made lousy software on purpose to sell new software and anti-virus software and new hardware.
I'm glad that DeepSeek is open software. That will make a big difference. The data centers will still need their big server complexes to store data, but accessing the data with efficient software will save money, energy, and hardware space. Is it a good thing when the state has data on all of us to have efficient software to sift through that data?
Carl, I'm glad you brought up quantum computing. I too read the Foreign Affairs essay, as I've been following qubits and quantum computing advances for a decade or so.
QC strikes me as similar to fusion, where the enthusiasm for the technology far exceeds the capabilities of the technology, and glosses over the immense physical limits of these technologies.
For example, for qubits (the "gates" in the QC) to remain stable, the machine must be cooled to near absolute zero. Hmm, this doesn't sound like a cheap or easy process. It sort of sounds like controlling processes at 100 million degrees F., i.e. fusion. Both of these are extremes that cannot be understood by conventional means because they're so far from human experience and so difficult to maintain for longer than micro-seconds--or in the case of fusion, billionths of a second.
The article's authors all work in Big Tech, so of course they think encryption is the Number One Big Problem of the future, and the value proposition of QC is identifying new compounds that have medical applications.
These are problems that they intend to "solve" with brute force, as that's the value proposition of QC-- faster processing times, parallel processing, etc.
But as we're now realizing with DeepSeek, maybe brute force processing is over-rated and unnecessary--in effect, a gigantic waste of irreplaceable resources.
warm regards, charles
For unknown reasons, a comment by Simple John did not post here, so I'm copying it here as it's a topic worth pondering:
"I'd appreciate correction if I'm wrong. I believe I've read that the DeepSeek models are strongest on math and physics. In fact, aren't LLM and image generating AI playing in a universe that is immensely less specific than math and physics and thus really just playing with words and images without any real insights?"
OK, Charles here. This is very insightful, as the examples of DeepSeek (or competing tools) solving math problems are "problems" where the "correct answer" can be determined. The "answer" to the "question" *write an essay on Charles Darwin for my class assignment" has no equivalent "correct answer." The AI Bot can hallucinate a response that might pass muster if the hallucination is not too wild.
I played around with Google's free AI tools for generating podcasts, summaries, etc., and have watched developments in the video-creation space. These are examples of brute-force processing working with templates assembled from "machine learning," i.e. sampling human-generated videos, stories, essays, etc. As John noted, this kind of extrapolation of existing content is a different kind of "answer" to a different kind of "problem."
So there are video-generation tools where you enter text instructions such as "a young man is walking through a 4th of July party holding a beer," and the program generates a video clip of this scene based on its vast database of 4th of July clips, people holding a beer, etc. This is fascinating and fun, because there is no "correct answer."
But what's the value proposition here when everyone (or anyone with a keen interest) can access the same tools and generate limitless AI content? Who's going to watch all this in an Attention Economy that'd already saturated with content?
Based on my limited understanding of the many software techniques DeepSeek employs, it seems likely that these structures may well be superior ways to solve "problems" that have a testable "correct" answers.
Without going too deeply into specifics, a key concept in high-end machine learning (Google's DeepMind, etc.) is estimating the accuracy of the answer, the probabilities of various potential solutions being correct. The "wait a minute, maybe this isn't the best path, let's start over" function of DeepSeek is based on the same idea, which is extremely valuable when the "correctness" of the answer actually matters.
warm regards, charles
Brad, those are very interesting points for exploration.
I'm reminded of various stories in TechLand about small teams developing the breakthroughs while thousands of employees in the corporation generated little value beyond maintaining the status quo. So Google has 175K employees, all smart, all dedicated, but a much smaller team blew the 175K teams of MSFT, AAPL, GOOG, META et al. away. Steve Jobs famously kept the Macintosh team isolated from Corporate Apple and exerted obsessive control over the development team. He knew better than to let Corporate do the Corporate Thing to the team.
Since finance and tech dominate the economy now, the big question is: how do I make a killing by investing in DeepSeek? This is a koan because there is no "owner" of the software techniques DeepSeek has shared with the world.
To Brad's question, so what do the tech monopolies / behemoths do with their 175K "teams" now? For sure they can try to replicate DeepSeek, but does that require 175K employees? And since everyone else has access to the same concepts, techniques and approaches, then where is the scarcity value that generates revenues? There is none.
This all leads to a sobering conclusion: there is little justification for these huge headcounts going forward.
warm regards, charles
I wanted to mention my podcast with Hrvoye at Geopolitics & Empire (Substack): Charles Hugh Smith: Anti-Progress, Resource Constraints, & Digital Neofeudalism (1:29 hrs) https://geopoliticsandempire.substack.com/p/charles-hugh-smith-anti-progress
Charles is it possible to use analog for local economies to function within the division of labor, thus replacing digital? The reason I ask is today, with the announcement of DeepSeek, Nvidia Corporation lost 1/2 $trillion in market cap. What is next? $100 billion margin calls? Federal Reserve bailouts similar to TARP? Fast forward and the US dollar will default via inflation.
Local economies will be the source of restoration. Not a central government. Perhaps analog will work?
The concept makes good sense. I wasn't around in the old mainframe era, but I learned to program on the first generation of PCs, making maximum use of small memory and slow PCs. I wrote some audio and graphics programs that ran smooth and fast. Modern bloatware on vastly larger and faster computers still can't match it.
Russia had one giant advantage in the Sputnik era, and our response to Sputnik made the Russian advantage larger, not smaller. Russia trained its math and engineer types with experience and teamwork, while we trained with memorized theory. We responded by increasing the theory, NOT by switching to experience and discussion. Since then we've offshored our tech and our engineering, so we've lost both skills. Russia didn't play the free trade game and continued making its own stuff and software. Our perpetual sanctions strengthened Russia's self-sufficiency.
Russia has always assumed extremely limited resources, while we've made the machines bigger and our brains smaller. China carries on the same tradition.
So if the Big 7 don't need all those massive server farms for AI, what else might they repurpose them for? In my experience, such a resource will not just be dismantled or abandoned, but the owners will desperately seek a way to use it for something.
I am certainly no software expert, but does the DeepSeek software also compete at making deepfake images and videos? Is it less prone to hallucinations?
I will be eagerly watching for more info in the coming weeks. And watching the stock market.
I've always had a hunch that there are simpler methods to achieve high performing artificial intelligence. Of course, we should note that the chips China did this on are vastly superior than anything prior to the year 2000 lets say. And China does have decades of manufacturing experience and a literal army worth of engineers to throw at the problem. So we shouldn't say that what China did was easy. It was just easier than throwing a mountain of money at the problem. You can't fake scarcity that easily and the Mag 7 as you call them will continue to struggle while the smaller teams given freedom to try anything with their limited resources will continue to find success.
Or that mountain of money might end up working. Who knows for sure until it happens?