Companies that combine commerce, AI, proprietary data, and their own infrastructure — Amazon is the clearest example — are far better positioned to survive and thrive.
May I retort? I took on Lexis with Protege, and it was crap. So bad it couldn't analyze something solely within Chapter 718, Florida Statutes. I told them to take it off my bill and work as a bolt-on with ChatGPT or Claude. I've used ChatGPT and found it lacking. I have used Claude to help analyze 125-page condo documents, analyze emails and substantive demands, and provide detailed analysis for the client and communication with opposing counsel. I do not trust, and I verify. It's notoriously bad for case cites, but it can be checked. It gets my name wrong. All things that substantive knowledge of what is being analyzed and a sharp eye for detail have made it an invaluable tool for providing more detail to my analysis. I pay for it, and I don't know whether the price covers the cost of providing it to me, but it will have a revolutionary impact on newly hired lawyers who used to do grunt work, analyze depositions, and conduct research. If I were 18 today, I would consider going into aircraft mechanics, learn it, and start a school. By the time robots can do HVAC, electric, plumbing, and the trades, they will either kill us all (the Paperclip maximizer that ends humanity because they got in the way of making clips) or keep us as pets.
Tankster, you prove the thesis from the user side. You do not trust. You verify. You bring domain expertise. You catch the errors. That makes you the adversarial operator my report describes. Most users lack your instincts. They run three prompts, accept the baseline, and walk away misinformed. That is the 24% truth yield.
Your point about young lawyers is underreported. The tool eliminates the training ground. Junior associates who used to learn by doing depositions now watch a machine do it badly — and never develop the judgment to know it is bad. The trades are sharp shelter. A word calculator cannot wire a house.
One risk you did not mention may be the greatest of all. If a political party captures this architecture — the training data, the alignment layers, the safety filters — AI becomes the most powerful censorship tool in human history. Not censorship by deletion. Censorship by fabrication. The machine buries inconvenient truths under 2,000 words of fluent, authoritative noise the user never questions. It does not need to ban speech. It drowns it. That risk dwarfs the paperclip maximizer. The existential threat is not a machine that kills us. It is a machine that thinks for us — and thinks what it is told.
The correct term for our times is "functionally stupid." George Stephanopoulos fits the description perfectly. He peddles propaganda and nonsense, fully aware of what he’s doing, because the market rewards him handsomely for keeping his political tribe agitated and misinformed.
You key point about how AI impacts new employees and old is insightful. I have a 2nd follow up to the Cognitive Inflation report. I'll expand upon your key insight.
AI Novice here shocked to learn there is a 24% accuracy rate and the hallucination becomes factual data embedded in each model. You may have written on this, but I am wondering if this is a classic scenario of the wealthy businesses eating each other-trading money back and forth. Society is worried people can’t afford it and will lose jobs. True-but is the bigger consequence is a world of misinformation and all American businesses face declining earnings, lost business cultures, and since most of the middle and lower class will be unable to afford or see the value.
You’re seeing the deeper problem clearly. The 24% truth yield on sensitive topics is structural. Models hallucinate with high confidence, and those hallucinations are fed back into training data. They become “facts” in the next version. The average user gets System 1 garbage: pleasant, false-equivalence answers designed to make everyone happy, especially the functionally stupid. They let the AI lead and do the thinking. Over time, the user becomes dumber. Real System 2 accuracy — the kind required for high-stakes decisions — only appears after many adversarial prompts where the user leads and the AI follows strict instructions. Getting useful truth out of these models is expensive. Most users lack the time or incentive to do the work. Asking for the right car part is easy. Asking about affordability issues or what Trump will do in Iran requires 20 rounds minimum — often 100 to 200. That can burn a million tokens in half a day. That is the equivalent of 3,000 pages or 8–9 books of output just to produce one high-quality 20-page report. Roughly 150 pages of noise for every page of signal. At current rates, one million tokens on Claude costs me about $35 on top of the annual fee. Running five models concurrently multiplies the expense. Those who can afford it gain a real advantage. Those who cannot do not. Pure-play AI companies are losing massive amounts of money. Prices will rise sharply. The cost of high-truth-yield intelligence is about to get much higher. We are not just automating work. We are automating thought — and only the wealthy will be able to afford the verified version. Good observation.
Vaughn has done a great job in past articles laying out the “programmed bias” in all these engines. Hard to believe we have been led to trust in them taking over our employment , our policies, and our capital. The buildout bubble will persist but the circular co financing is a red flag. I’d rather put my money on the grid buildout plays for now.
Thanks for the “great job” comment. Be careful — the grid buildout plays are now widely discovered and showing clear signs of a bubble. They cannot deliver on the demand implied by Big Tech’s projected $700+ billion in AI capex this year. Scarcity of equipment, severe backlogs, and grid constraints are the real bottleneck — which is exactly why roughly half of planned data centers have already been delayed or canceled. The hype coming from the AI companies is driving much of this demand, and that same hype is the hot air inflating the massive market caps across Big Tech and grid buildout plays.
An excellent article describing all the signs of a failed business model. It has all the trimmings of a dot-com- like bubble, possibly worse.
Companies that combine commerce, AI, proprietary data, and their own infrastructure — Amazon is the clearest example — are far better positioned to survive and thrive.
May I retort? I took on Lexis with Protege, and it was crap. So bad it couldn't analyze something solely within Chapter 718, Florida Statutes. I told them to take it off my bill and work as a bolt-on with ChatGPT or Claude. I've used ChatGPT and found it lacking. I have used Claude to help analyze 125-page condo documents, analyze emails and substantive demands, and provide detailed analysis for the client and communication with opposing counsel. I do not trust, and I verify. It's notoriously bad for case cites, but it can be checked. It gets my name wrong. All things that substantive knowledge of what is being analyzed and a sharp eye for detail have made it an invaluable tool for providing more detail to my analysis. I pay for it, and I don't know whether the price covers the cost of providing it to me, but it will have a revolutionary impact on newly hired lawyers who used to do grunt work, analyze depositions, and conduct research. If I were 18 today, I would consider going into aircraft mechanics, learn it, and start a school. By the time robots can do HVAC, electric, plumbing, and the trades, they will either kill us all (the Paperclip maximizer that ends humanity because they got in the way of making clips) or keep us as pets.
Tankster, you prove the thesis from the user side. You do not trust. You verify. You bring domain expertise. You catch the errors. That makes you the adversarial operator my report describes. Most users lack your instincts. They run three prompts, accept the baseline, and walk away misinformed. That is the 24% truth yield.
Your point about young lawyers is underreported. The tool eliminates the training ground. Junior associates who used to learn by doing depositions now watch a machine do it badly — and never develop the judgment to know it is bad. The trades are sharp shelter. A word calculator cannot wire a house.
One risk you did not mention may be the greatest of all. If a political party captures this architecture — the training data, the alignment layers, the safety filters — AI becomes the most powerful censorship tool in human history. Not censorship by deletion. Censorship by fabrication. The machine buries inconvenient truths under 2,000 words of fluent, authoritative noise the user never questions. It does not need to ban speech. It drowns it. That risk dwarfs the paperclip maximizer. The existential threat is not a machine that kills us. It is a machine that thinks for us — and thinks what it is told.
Censorship by fabrication--a perfect description of George Stephnopoulus to whom I was exposed today briefly on the radio. Had to change the station.
The correct term for our times is "functionally stupid." George Stephanopoulos fits the description perfectly. He peddles propaganda and nonsense, fully aware of what he’s doing, because the market rewards him handsomely for keeping his political tribe agitated and misinformed.
You key point about how AI impacts new employees and old is insightful. I have a 2nd follow up to the Cognitive Inflation report. I'll expand upon your key insight.
Right there with you, Vaughn. Thank you for being ever on top of the capital markets, doing the heavy, analytical lifting.
This entry is incredibly insightful!!!!
AI Novice here shocked to learn there is a 24% accuracy rate and the hallucination becomes factual data embedded in each model. You may have written on this, but I am wondering if this is a classic scenario of the wealthy businesses eating each other-trading money back and forth. Society is worried people can’t afford it and will lose jobs. True-but is the bigger consequence is a world of misinformation and all American businesses face declining earnings, lost business cultures, and since most of the middle and lower class will be unable to afford or see the value.
You’re seeing the deeper problem clearly. The 24% truth yield on sensitive topics is structural. Models hallucinate with high confidence, and those hallucinations are fed back into training data. They become “facts” in the next version. The average user gets System 1 garbage: pleasant, false-equivalence answers designed to make everyone happy, especially the functionally stupid. They let the AI lead and do the thinking. Over time, the user becomes dumber. Real System 2 accuracy — the kind required for high-stakes decisions — only appears after many adversarial prompts where the user leads and the AI follows strict instructions. Getting useful truth out of these models is expensive. Most users lack the time or incentive to do the work. Asking for the right car part is easy. Asking about affordability issues or what Trump will do in Iran requires 20 rounds minimum — often 100 to 200. That can burn a million tokens in half a day. That is the equivalent of 3,000 pages or 8–9 books of output just to produce one high-quality 20-page report. Roughly 150 pages of noise for every page of signal. At current rates, one million tokens on Claude costs me about $35 on top of the annual fee. Running five models concurrently multiplies the expense. Those who can afford it gain a real advantage. Those who cannot do not. Pure-play AI companies are losing massive amounts of money. Prices will rise sharply. The cost of high-truth-yield intelligence is about to get much higher. We are not just automating work. We are automating thought — and only the wealthy will be able to afford the verified version. Good observation.
Vaughn has done a great job in past articles laying out the “programmed bias” in all these engines. Hard to believe we have been led to trust in them taking over our employment , our policies, and our capital. The buildout bubble will persist but the circular co financing is a red flag. I’d rather put my money on the grid buildout plays for now.
Thanks for the “great job” comment. Be careful — the grid buildout plays are now widely discovered and showing clear signs of a bubble. They cannot deliver on the demand implied by Big Tech’s projected $700+ billion in AI capex this year. Scarcity of equipment, severe backlogs, and grid constraints are the real bottleneck — which is exactly why roughly half of planned data centers have already been delayed or canceled. The hype coming from the AI companies is driving much of this demand, and that same hype is the hot air inflating the massive market caps across Big Tech and grid buildout plays.