Yesterday I spent a lot of hrs in a hospital (waiting for my second child's birth - which despite some hiccups is now with us, happy and healthy) and I saw the two following situations.
1. A 40-something woman accompanying her blind mother. They were arguing about the mother's diet, habits etc. At every turn of conversation, she took her phone and - yes, i confess I pried - asked Chatgpt the best way to convince her mother about doing certain stuff, taking certain medicines.
2. Another woman due to start labor shortly, came out from the initial doctor's check. She sat down, immediately took her phone out and start typing. This time I didn't have to pry, she told her partner that Chatgpt said so and so about what the doctor had said...
What's happening here? It would seem like a synthetic actor is mediating a real conversation, or what could have been a real conversation. Without entering in the particulars of the cases above, I'm realizing a big risk is not about intellectual offload, rather an offload of basic social skills. A chatbox, or even a more context-aware ai assistant embedded in glasses, will never have enough context to guide a conversation betwenen two human beings.
I'm by nature a highly empathetic person, and I can "feel" a person as I speak to her. I can "feel" a room and "sense" the tone. Is this something that can be learned? If so, is there a risk of teens and young adults offloading this by defaulting to AI to ask how to continue a conversation?
The go example you cite is somewhat contested. Recently a former go player argued that the increase in performance is mostly due to a) memorizing openings and b) cheating.
I'm a very happy reader of EV for many years. Reading your newsletter is usually a joy I reserve for my Sunday morning. However, I found today's edition slightly less enjoyable to read than usual.
Specifically, the "Linear knives in exponential gunfights" column relied heavily on the negative-positive parallelisms that are classic for AI writing. Please don't take this as an accusation, but I've personally grown quite fatigued by that specific stylistic pattern. I mention it only because I feel it subtly undermines the authority you’ve established over the years.
That is utter rubbish. Handwritten, initially on paper, dictated into Otter. The section on the JPM consensus forecasts was dictated by me into Granola while I was walking to the butchers, the transcript tidied up by me in Google Docs. Hand-edited printouts. ChatGPT reviewed the definition of intelligence, whcih was initally taught to me by a Swiss friend in 2000, becuase Google's etymology response only went one layer deep.
The typo "appropraite" was left in deliberately by me to see if anyone complained (someone did). The Micron typo, I didn't notice.
The analyst miss data is based on a data set I compiled ten days ago for an Investment Committee meeting I took last Monday. And the articulation around why the FT article missed came from, frankly, my work over the best 15 years. Obviously Claude Cowork drew the diagram because my version in Excel is far too ugly.
Mauboussin makes it in there because I've just prepared a valuation reading list for one of our researchers who is new to finance. I know the exact quote because, well I know the exact quote.
As for the cliche, knives and gunfights, you will recall that when i first wrote about rare earths in 2017 I quoted the 2010 Foreign Policy article “don’t bring a praesodymium knife to a gunfight.” I quite liked the phrasing, and I deliberately used it in this piece BECAUSE it is a cliche. Because the notion on relying on consensus forecasts during a regime change is so laughably absurd. (Although, I would add that the fact people do just creates alpha for those who don’t.)
Everything we publish runs through GPT-zero and Pangram, This is to make sure we haven't accidentally copied and pasted something from our research into the live document during the edit process. (Multiple people do edit and so this check is a safety feature.). The Pangram score is 100% human; GPT-Zero is 0% AI probability (Scores the other way to Pangram, i.e. 100% human). I would add we do this process with all our output - 100% human scores on Pangram and GPT-Zero before publishing, but ultimately the human writing it takes the judgment call, usually from reading a print out.
Despite the 100% score, GPT-Zero flags some sections it thinks were AI-human-mixed, but I reviewed them and didn't change them because I actually wrote them. (A good example is the section where I describe the distribution of outcomes. The reason is that it is a technical point, so it requires more moderate, precise language. I didn't have time to turn my current cumbersome paragraph into something more appealing.)
Ditto, by the way, on AI and thinking. I'll draw attention to how i refrained "joining the dots" at the end of two paragraphs. Hoping people would ... join the dots.
I'd also add that our standard at EV is the use the em dash. As an early typographer, designed my own fonts in the 1990s, I'm particular about these things. We use them by default, or should.
You are so far off the mark that you are, to use my phrasing, “on the wrong curve entirely.”
Well, that certainly escalated quickly. I think we are talking past each other.
I am not accusing you of using an LLM to ghostwrite your column, nor am I questioning your editorial process or your ideas themselves.
My comment was on the stylistic choice of using repeated negative-positive parallelisms (for lack of a better term). Those have become a hallmark of AI cadence, which makes for an (in my opinion) exhausting reading experience.
To show you what I mean, here is how the argument was structured repeatedly in just this one section:
'...we aren’t just choosing a number. We are expressing...'
'You haven’t just chosen the wrong point... You’ve chosen the wrong curve.'
'Analysts aren’t just missing the number once. They are missing it repeatedly.'
'The lesson isn’t that analysts are too pessimistic... It is that consensus forecasts...'
My point is simply that leaning this heavily on the 'It's not X, it's Y' structure distracts from the actual substance of your ideas.
AI editing question: when you grasped your fountain pen and wrote the “To reverse Charlie Munger” sentence, did it originally read “show me the outcome, and I’ll show you the incentive,” which your AI editors subsequently reversed?
It’s hard to believe that you would have intended to quote Munger directly after starting with “To reverse…”
Interesting. You are right as you are cynical. That doesn't read as well as it would if it was inverted.
I have the "autosaves" from Substack, because this is one bit I typed, starting on Friday. At 815pm on Saturday night, the line ran:
"To reverse Charlie Munger,
%%%% more thoughts here
or for during the course of their employment. "
About 25 minutes, later it read:
"To reverse Charlie Munger, we’re seeing the outcomes, so let’s figure out what the incentives are. If the outcome is offloading to AI, perhaps the incentives make using ChatGPT for your thinking the appropriate choice.
Those incentives emerged from a complex system of tests, credentialing and career ladders that we’ve built up over centuries, when cognition was scarce"
But i didn't think the cumbersome sentence followed by the passive tense--and a long sentence work. (That second paragraph doesn't even end with 'was scarce'... it goes on and on.)
The next change at about 11pm --after I took a break--is roughly the sentence as written:
"To reverse Charlie Munger, show me the incentive, and I’ll show you the outcome. The outcome here is offloading important thinking to a chatbot. So what incentives produced it? We’ve developed tests, credentials and career ladders and cemented them across centuries of practice. They were built when cognition was scarce and scarcity paid well. Thinking proved you had that scarce asset, and we never needed to ask if you were actually the one doing the thinking. That isn’t true now, but the payoff for perfecting the scarcity theatre of a perfect GPA, Ivy education, and the rest remains."
Here is Claude's history on feedback, which I must have got before the 11pm final edit
What's working well
The etymology of intellegere is the best moment in the essay — "to choose between" as the core of intelligence lands hard and earns its place. The marathon/car analogy is vivid and does real work. And the Munger inversion at the end ("we're seeing the outcomes, so let's figure out what the incentives are") is elegant — it reframes the whole anxiety productively.
The personal texture is also genuinely useful here: the whiteboards, the fountain pens, the RMA going off to do its own research while you work on paper. That's not padding — it earns the right to your conclusion.
A few things to tighten
"Milquetoasting" is doing a lot of work in the opening and I'm not sure it's doing it well — the word is slightly awkward and the essay doesn't return to it. You could cut it and not lose anything.
The Futurama/Neuralink riff is fun but runs a beat too long. You could cut it at "Half of us in VR headsets" and the joke lands better.
"Cognitive cogitations" — that's a slip. Sounds like the kind of phrase you'd red-pen in someone else's draft.
So you'll notice -- Claude commended the original Munger inversion.
I'm not clear when I swapped it round. Maybe during the hand-edit? My original version (and the one Claude preferred) does read better. Humans, eh.
Not cynical, because I’m a loyal fan of your work. Perhaps I was projecting something I’ve seen far too often: the Unreliable Editor, when my LLM partner goes rogue and “fixes” things that aren’t broken.
They rarely edit text directly. If you do that at any great length they start changing words here and there, left right and centre.
The final pass is always grammar or google spell check.
In this case, i suspect taht i corrected it because it was late on Saturday night, i was tired and it looked odd to me—but i had forgotten the context of WHY i had written it.
I am resonating with the tile of this newsletter (which I deeply appreciate, as all your clever ones): we have done a far deeper research on the gains/losses of 20 technologies, you may enjoy at least looking at the summary tables: Technologies Smarter, Humans, “Dumber”? (2026)
"This paper examines how successive technological innovations—from writing and printing to digital media, smartphones, and artificial intelligence—have altered human cognitive capabilities and the conditions for learning. Technological progress consistently produces both cognitive gains and losses: while technologies expand abstraction, efficiency, and access to knowledge, they simultaneously weaken embodied, contextual, and internally sustained capacities. Recommendations for changes in education foci are thus provided."
makes an interesting counterpoint. Only 15% of global GDP is digital, the rest is composed of what he refers to as atoms. Starting to see more commentary on my feeds asking where is the AI beef? (Old American TV ad about roast beef sandwiches)
I came of age as desktop computers began to creep into the corporate world. It was a slow two decade process getting the workforce up and running on PC's. It takes time to bring new ways of working to bear fruit; flows need rethinking and it takes time to identify those who are capable of making these new tools hum.
"Does AI make you dumber" Well, from personal experience, I'm probably a worse programmer than I would have been if I had coded 2 2,500 line programs from mid-January to mid-March myself instead of getting Gemini to do it--but I would never have even attempted that. What AI has done for me is greatly expand the scope of what I will try to do, either in coding or in researching.
There is some advantage to working with an AI that knows "everything"--even if jaggedly. Obscure computer hardware--it knows about it. German institution that offers a public domain program for making an industrial level BMS: AI can point you to it--and then walk you through using it.
Regarding "AI punctuation": I've always used "--", ";", ":". My theory is that AI learned to write by copying writers who wrote the way I do.
i guess it depends where you start. i am a marginally better programmer (or at least architect) because i have had a lot of practice in the past eigtht months!
Absolutely, and in a broader context, I would say I'm a much more productive developer than I was a year ago, because if I have an idea, I just ask Gemini / Chat / Claude to throw together a python program (I don't program in python) to graph some data, and 3 minutes later, I have my graph, usually on the first try. What's the goal? Is it to improve on a particular skill, or is it to accomplish more things in the real world?
Yesterday I spent a lot of hrs in a hospital (waiting for my second child's birth - which despite some hiccups is now with us, happy and healthy) and I saw the two following situations.
1. A 40-something woman accompanying her blind mother. They were arguing about the mother's diet, habits etc. At every turn of conversation, she took her phone and - yes, i confess I pried - asked Chatgpt the best way to convince her mother about doing certain stuff, taking certain medicines.
2. Another woman due to start labor shortly, came out from the initial doctor's check. She sat down, immediately took her phone out and start typing. This time I didn't have to pry, she told her partner that Chatgpt said so and so about what the doctor had said...
What's happening here? It would seem like a synthetic actor is mediating a real conversation, or what could have been a real conversation. Without entering in the particulars of the cases above, I'm realizing a big risk is not about intellectual offload, rather an offload of basic social skills. A chatbox, or even a more context-aware ai assistant embedded in glasses, will never have enough context to guide a conversation betwenen two human beings.
I'm by nature a highly empathetic person, and I can "feel" a person as I speak to her. I can "feel" a room and "sense" the tone. Is this something that can be learned? If so, is there a risk of teens and young adults offloading this by defaulting to AI to ask how to continue a conversation?
Burying the lede! Congrats 🐣
The go example you cite is somewhat contested. Recently a former go player argued that the increase in performance is mostly due to a) memorizing openings and b) cheating.
https://www.lesswrong.com/posts/nR3DkyivzF4ve97oM/how-go-players-disempower-themselves-to-ai
I'm a very happy reader of EV for many years. Reading your newsletter is usually a joy I reserve for my Sunday morning. However, I found today's edition slightly less enjoyable to read than usual.
Specifically, the "Linear knives in exponential gunfights" column relied heavily on the negative-positive parallelisms that are classic for AI writing. Please don't take this as an accusation, but I've personally grown quite fatigued by that specific stylistic pattern. I mention it only because I feel it subtly undermines the authority you’ve established over the years.
That is utter rubbish. Handwritten, initially on paper, dictated into Otter. The section on the JPM consensus forecasts was dictated by me into Granola while I was walking to the butchers, the transcript tidied up by me in Google Docs. Hand-edited printouts. ChatGPT reviewed the definition of intelligence, whcih was initally taught to me by a Swiss friend in 2000, becuase Google's etymology response only went one layer deep.
The typo "appropraite" was left in deliberately by me to see if anyone complained (someone did). The Micron typo, I didn't notice.
The analyst miss data is based on a data set I compiled ten days ago for an Investment Committee meeting I took last Monday. And the articulation around why the FT article missed came from, frankly, my work over the best 15 years. Obviously Claude Cowork drew the diagram because my version in Excel is far too ugly.
Mauboussin makes it in there because I've just prepared a valuation reading list for one of our researchers who is new to finance. I know the exact quote because, well I know the exact quote.
As for the cliche, knives and gunfights, you will recall that when i first wrote about rare earths in 2017 I quoted the 2010 Foreign Policy article “don’t bring a praesodymium knife to a gunfight.” I quite liked the phrasing, and I deliberately used it in this piece BECAUSE it is a cliche. Because the notion on relying on consensus forecasts during a regime change is so laughably absurd. (Although, I would add that the fact people do just creates alpha for those who don’t.)
Everything we publish runs through GPT-zero and Pangram, This is to make sure we haven't accidentally copied and pasted something from our research into the live document during the edit process. (Multiple people do edit and so this check is a safety feature.). The Pangram score is 100% human; GPT-Zero is 0% AI probability (Scores the other way to Pangram, i.e. 100% human). I would add we do this process with all our output - 100% human scores on Pangram and GPT-Zero before publishing, but ultimately the human writing it takes the judgment call, usually from reading a print out.
Despite the 100% score, GPT-Zero flags some sections it thinks were AI-human-mixed, but I reviewed them and didn't change them because I actually wrote them. (A good example is the section where I describe the distribution of outcomes. The reason is that it is a technical point, so it requires more moderate, precise language. I didn't have time to turn my current cumbersome paragraph into something more appealing.)
Ditto, by the way, on AI and thinking. I'll draw attention to how i refrained "joining the dots" at the end of two paragraphs. Hoping people would ... join the dots.
I'd also add that our standard at EV is the use the em dash. As an early typographer, designed my own fonts in the 1990s, I'm particular about these things. We use them by default, or should.
You are so far off the mark that you are, to use my phrasing, “on the wrong curve entirely.”
BTW, here is the Granola on the mean reversion stuff
https://notes.granola.ai/d/b52902af-8752-4f87-9875-e50a25c57dfd
I fear Philipp brought an accusatory knife to an evidence-based gun fight….
Well, that certainly escalated quickly. I think we are talking past each other.
I am not accusing you of using an LLM to ghostwrite your column, nor am I questioning your editorial process or your ideas themselves.
My comment was on the stylistic choice of using repeated negative-positive parallelisms (for lack of a better term). Those have become a hallmark of AI cadence, which makes for an (in my opinion) exhausting reading experience.
To show you what I mean, here is how the argument was structured repeatedly in just this one section:
'...we aren’t just choosing a number. We are expressing...'
'You haven’t just chosen the wrong point... You’ve chosen the wrong curve.'
'Analysts aren’t just missing the number once. They are missing it repeatedly.'
'The lesson isn’t that analysts are too pessimistic... It is that consensus forecasts...'
My point is simply that leaning this heavily on the 'It's not X, it's Y' structure distracts from the actual substance of your ideas.
Interesting piece. Made me think. I presumed the typos and issues like "Half, half, half" were "easter eggs" to see if we were actually reading it!
AI editing question: when you grasped your fountain pen and wrote the “To reverse Charlie Munger” sentence, did it originally read “show me the outcome, and I’ll show you the incentive,” which your AI editors subsequently reversed?
It’s hard to believe that you would have intended to quote Munger directly after starting with “To reverse…”
Interesting. You are right as you are cynical. That doesn't read as well as it would if it was inverted.
I have the "autosaves" from Substack, because this is one bit I typed, starting on Friday. At 815pm on Saturday night, the line ran:
"To reverse Charlie Munger,
%%%% more thoughts here
or for during the course of their employment. "
About 25 minutes, later it read:
"To reverse Charlie Munger, we’re seeing the outcomes, so let’s figure out what the incentives are. If the outcome is offloading to AI, perhaps the incentives make using ChatGPT for your thinking the appropriate choice.
Those incentives emerged from a complex system of tests, credentialing and career ladders that we’ve built up over centuries, when cognition was scarce"
But i didn't think the cumbersome sentence followed by the passive tense--and a long sentence work. (That second paragraph doesn't even end with 'was scarce'... it goes on and on.)
The next change at about 11pm --after I took a break--is roughly the sentence as written:
"To reverse Charlie Munger, show me the incentive, and I’ll show you the outcome. The outcome here is offloading important thinking to a chatbot. So what incentives produced it? We’ve developed tests, credentials and career ladders and cemented them across centuries of practice. They were built when cognition was scarce and scarcity paid well. Thinking proved you had that scarce asset, and we never needed to ask if you were actually the one doing the thinking. That isn’t true now, but the payoff for perfecting the scarcity theatre of a perfect GPA, Ivy education, and the rest remains."
Here is Claude's history on feedback, which I must have got before the 11pm final edit
What's working well
The etymology of intellegere is the best moment in the essay — "to choose between" as the core of intelligence lands hard and earns its place. The marathon/car analogy is vivid and does real work. And the Munger inversion at the end ("we're seeing the outcomes, so let's figure out what the incentives are") is elegant — it reframes the whole anxiety productively.
The personal texture is also genuinely useful here: the whiteboards, the fountain pens, the RMA going off to do its own research while you work on paper. That's not padding — it earns the right to your conclusion.
A few things to tighten
"Milquetoasting" is doing a lot of work in the opening and I'm not sure it's doing it well — the word is slightly awkward and the essay doesn't return to it. You could cut it and not lose anything.
The Futurama/Neuralink riff is fun but runs a beat too long. You could cut it at "Half of us in VR headsets" and the joke lands better.
"Cognitive cogitations" — that's a slip. Sounds like the kind of phrase you'd red-pen in someone else's draft.
So you'll notice -- Claude commended the original Munger inversion.
I'm not clear when I swapped it round. Maybe during the hand-edit? My original version (and the one Claude preferred) does read better. Humans, eh.
Not cynical, because I’m a loyal fan of your work. Perhaps I was projecting something I’ve seen far too often: the Unreliable Editor, when my LLM partner goes rogue and “fixes” things that aren’t broken.
They rarely edit text directly. If you do that at any great length they start changing words here and there, left right and centre.
The final pass is always grammar or google spell check.
In this case, i suspect taht i corrected it because it was late on Saturday night, i was tired and it looked odd to me—but i had forgotten the context of WHY i had written it.
Really appreciate the human honesty, which will trade at a rising premium over time
I am resonating with the tile of this newsletter (which I deeply appreciate, as all your clever ones): we have done a far deeper research on the gains/losses of 20 technologies, you may enjoy at least looking at the summary tables: Technologies Smarter, Humans, “Dumber”? (2026)
https://curriculumredesign.org/wp-content/uploads/Technology-Smarter-Humans-DumberQ-CCR.pdf
"This paper examines how successive technological innovations—from writing and printing to digital media, smartphones, and artificial intelligence—have altered human cognitive capabilities and the conditions for learning. Technological progress consistently produces both cognitive gains and losses: while technologies expand abstraction, efficiency, and access to knowledge, they simultaneously weaken embodied, contextual, and internally sustained capacities. Recommendations for changes in education foci are thus provided."
Be well,
Charles
Greg Satell published this column today,
https://digitaltonto.com/2026/the-world-is-not-digital-and-thats-why-software-wont-eat-it/
makes an interesting counterpoint. Only 15% of global GDP is digital, the rest is composed of what he refers to as atoms. Starting to see more commentary on my feeds asking where is the AI beef? (Old American TV ad about roast beef sandwiches)
I came of age as desktop computers began to creep into the corporate world. It was a slow two decade process getting the workforce up and running on PC's. It takes time to bring new ways of working to bear fruit; flows need rethinking and it takes time to identify those who are capable of making these new tools hum.
"Does AI make you dumber" Well, from personal experience, I'm probably a worse programmer than I would have been if I had coded 2 2,500 line programs from mid-January to mid-March myself instead of getting Gemini to do it--but I would never have even attempted that. What AI has done for me is greatly expand the scope of what I will try to do, either in coding or in researching.
There is some advantage to working with an AI that knows "everything"--even if jaggedly. Obscure computer hardware--it knows about it. German institution that offers a public domain program for making an industrial level BMS: AI can point you to it--and then walk you through using it.
Regarding "AI punctuation": I've always used "--", ";", ":". My theory is that AI learned to write by copying writers who wrote the way I do.
i guess it depends where you start. i am a marginally better programmer (or at least architect) because i have had a lot of practice in the past eigtht months!
Absolutely, and in a broader context, I would say I'm a much more productive developer than I was a year ago, because if I have an idea, I just ask Gemini / Chat / Claude to throw together a python program (I don't program in python) to graph some data, and 3 minutes later, I have my graph, usually on the first try. What's the goal? Is it to improve on a particular skill, or is it to accomplish more things in the real world?