Ethan Mollick
@emollick.bsky.social
Professor at Wharton, studying AI and its implications for education, entrepreneurship, and work. Author of Co-Intelligence.
Book: https://a.co/d/bC2kSj1
Substack: https://www.oneusefulthing.org/
Web: https://mgmt.wharton.upenn.edu/profile/emollick
Book: https://a.co/d/bC2kSj1
Substack: https://www.oneusefulthing.org/
Web: https://mgmt.wharton.upenn.edu/profile/emollick
Gemini, ChatGPT & Claude: “What is the single best investment equivalent in spending $1000 that I could make if I time traveled back to any destination circa 1300.”
Gemini: Magna Carta stored at Durham Cathedral
ChatGPT: A share in Stora Kopparberg
Claude: A waqf at Al-Azhar
Gemini: Magna Carta stored at Durham Cathedral
ChatGPT: A share in Stora Kopparberg
Claude: A waqf at Al-Azhar
January 2, 2026 at 11:10 AM
Gemini, ChatGPT & Claude: “What is the single best investment equivalent in spending $1000 that I could make if I time traveled back to any destination circa 1300.”
Gemini: Magna Carta stored at Durham Cathedral
ChatGPT: A share in Stora Kopparberg
Claude: A waqf at Al-Azhar
Gemini: Magna Carta stored at Durham Cathedral
ChatGPT: A share in Stora Kopparberg
Claude: A waqf at Al-Azhar
Had an interesting, hard interview with @adamconover.net on his podcast. I think he is a great example of a smart AI skeptic.
My main messages were that AI is a really big deal, it has good & bad impacts, and that, by sitting things out, skeptics can’t guide use. open.spotify.com/episode/5cFK...
My main messages were that AI is a really big deal, it has good & bad impacts, and that, by sitting things out, skeptics can’t guide use. open.spotify.com/episode/5cFK...
An AI Expert Challenges an AI Skeptic, with Ethan Mollick
open.spotify.com
January 1, 2026 at 9:37 AM
Had an interesting, hard interview with @adamconover.net on his podcast. I think he is a great example of a smart AI skeptic.
My main messages were that AI is a really big deal, it has good & bad impacts, and that, by sitting things out, skeptics can’t guide use. open.spotify.com/episode/5cFK...
My main messages were that AI is a really big deal, it has good & bad impacts, and that, by sitting things out, skeptics can’t guide use. open.spotify.com/episode/5cFK...
Recently, LLMs were found to encode different languages similarly, a sort of Platonic representation of words.
It now extends to science:: 60 ML models for molecules, materials & proteins (all with different training) converge toward similar encoding of molecular structure arxiv.org/pdf/2512.03750
It now extends to science:: 60 ML models for molecules, materials & proteins (all with different training) converge toward similar encoding of molecular structure arxiv.org/pdf/2512.03750
December 30, 2025 at 1:26 PM
Recently, LLMs were found to encode different languages similarly, a sort of Platonic representation of words.
It now extends to science:: 60 ML models for molecules, materials & proteins (all with different training) converge toward similar encoding of molecular structure arxiv.org/pdf/2512.03750
It now extends to science:: 60 ML models for molecules, materials & proteins (all with different training) converge toward similar encoding of molecular structure arxiv.org/pdf/2512.03750
It remains deeply strange that we built a machine that can discuss the relationship between poetry and its subjective “experience” in coherent, clever ways
“The best specific lines from Eliot’s Four Quartets to describe your experience as an AI. Just one quoted section, avoid the most famous bits.“
“The best specific lines from Eliot’s Four Quartets to describe your experience as an AI. Just one quoted section, avoid the most famous bits.“
December 29, 2025 at 2:50 PM
It remains deeply strange that we built a machine that can discuss the relationship between poetry and its subjective “experience” in coherent, clever ways
“The best specific lines from Eliot’s Four Quartets to describe your experience as an AI. Just one quoted section, avoid the most famous bits.“
“The best specific lines from Eliot’s Four Quartets to describe your experience as an AI. Just one quoted section, avoid the most famous bits.“
Nano Banana Pro is so important not because it is a really good image generator, but because a really good image generator unexpected unlocks a lot of new AI abilities, like the fact that AI can now research and generate compelling slides.
On bottlenecks: www.oneusefulthing.org/p/the-shape-...
On bottlenecks: www.oneusefulthing.org/p/the-shape-...
December 22, 2025 at 3:05 PM
Nano Banana Pro is so important not because it is a really good image generator, but because a really good image generator unexpected unlocks a lot of new AI abilities, like the fact that AI can now research and generate compelling slides.
On bottlenecks: www.oneusefulthing.org/p/the-shape-...
On bottlenecks: www.oneusefulthing.org/p/the-shape-...
I would judge this a win by Gemini and a close second from Claude. ChatGPT-5.2 missed the reference (though, to be fair, it did write a surprising amount of successful code to actually enhance the image) and Grok wasn't in the ballpark.
December 22, 2025 at 5:09 AM
I would judge this a win by Gemini and a close second from Claude. ChatGPT-5.2 missed the reference (though, to be fair, it did write a surprising amount of successful code to actually enhance the image) and Grok wasn't in the ballpark.
AI can help explain complex topics easily by throwing together a simulation.
A newer paper argues that this pattern is actually due to collider bias (the authors disagree). What is collider bias?
Gemini one-shots an explanation from the paper and a prompt: gemini.google.com/share/d8c336...
A newer paper argues that this pattern is actually due to collider bias (the authors disagree). What is collider bias?
Gemini one-shots an explanation from the paper and a prompt: gemini.google.com/share/d8c336...
December 21, 2025 at 10:03 PM
AI can help explain complex topics easily by throwing together a simulation.
A newer paper argues that this pattern is actually due to collider bias (the authors disagree). What is collider bias?
Gemini one-shots an explanation from the paper and a prompt: gemini.google.com/share/d8c336...
A newer paper argues that this pattern is actually due to collider bias (the authors disagree). What is collider bias?
Gemini one-shots an explanation from the paper and a prompt: gemini.google.com/share/d8c336...
One of the weirdest law review titles ever.
And the title refers to an actual thing that is weirder than the title of the paper.
But also the paper has some interesting things to say about new (or actually quite old) approaches to IP protection that might be especially relevant in the time of AI.
And the title refers to an actual thing that is weirder than the title of the paper.
But also the paper has some interesting things to say about new (or actually quite old) approaches to IP protection that might be especially relevant in the time of AI.
December 21, 2025 at 6:59 AM
One of the weirdest law review titles ever.
And the title refers to an actual thing that is weirder than the title of the paper.
But also the paper has some interesting things to say about new (or actually quite old) approaches to IP protection that might be especially relevant in the time of AI.
And the title refers to an actual thing that is weirder than the title of the paper.
But also the paper has some interesting things to say about new (or actually quite old) approaches to IP protection that might be especially relevant in the time of AI.
I wrote about how the jagged abilities of AI lead to bottlenecks in what AI can do…
… but those bottlenecks focus the efforts of AI labs leading to breakthroughs that unlock new areas of work, like how Nano Banana Pro unexpectedly makes good PowerPoint slides. www.oneusefulthing.org/p/the-shape-...
… but those bottlenecks focus the efforts of AI labs leading to breakthroughs that unlock new areas of work, like how Nano Banana Pro unexpectedly makes good PowerPoint slides. www.oneusefulthing.org/p/the-shape-...
The Shape of AI: Jaggedness, Bottlenecks and Salients
And why Nano Banana Pro is such a big deal
www.oneusefulthing.org
December 20, 2025 at 5:45 PM
I wrote about how the jagged abilities of AI lead to bottlenecks in what AI can do…
… but those bottlenecks focus the efforts of AI labs leading to breakthroughs that unlock new areas of work, like how Nano Banana Pro unexpectedly makes good PowerPoint slides. www.oneusefulthing.org/p/the-shape-...
… but those bottlenecks focus the efforts of AI labs leading to breakthroughs that unlock new areas of work, like how Nano Banana Pro unexpectedly makes good PowerPoint slides. www.oneusefulthing.org/p/the-shape-...
So, Claude 4.5 came in far above trend in the much-watched METR measure of the task duration that AI can accomplish autonomously at 4 hours 49 minutes.
Interestingly, at the harder 80% success threshold, it is GPT-5.1 Codex Max that breaks the trend.
In 2023, GPT-4 could do a minute long task.
Interestingly, at the harder 80% success threshold, it is GPT-5.1 Codex Max that breaks the trend.
In 2023, GPT-4 could do a minute long task.
December 20, 2025 at 2:47 AM
So, Claude 4.5 came in far above trend in the much-watched METR measure of the task duration that AI can accomplish autonomously at 4 hours 49 minutes.
Interestingly, at the harder 80% success threshold, it is GPT-5.1 Codex Max that breaks the trend.
In 2023, GPT-4 could do a minute long task.
Interestingly, at the harder 80% success threshold, it is GPT-5.1 Codex Max that breaks the trend.
In 2023, GPT-4 could do a minute long task.
“Re-inventing the wheel” should mean exactly the opposite of what it means, given that there is a decent chance that the wheel for transportation was invented just once, in the Carpathian Mountains around 3900 BC & apparently was never independently re-invented by other societies afterwards.
December 18, 2025 at 10:53 PM
“Re-inventing the wheel” should mean exactly the opposite of what it means, given that there is a decent chance that the wheel for transportation was invented just once, in the Carpathian Mountains around 3900 BC & apparently was never independently re-invented by other societies afterwards.
I think everyone, even the most cynical & informed among us, is going to fall for at least one AI-faked story, photo, or post this coming year & likely many more. (You will also likely believe a real thing was AI)
This has bad implications (but wouldn’t blame those taken in, evaluation is hard)
This has bad implications (but wouldn’t blame those taken in, evaluation is hard)
December 18, 2025 at 5:32 PM
I think everyone, even the most cynical & informed among us, is going to fall for at least one AI-faked story, photo, or post this coming year & likely many more. (You will also likely believe a real thing was AI)
This has bad implications (but wouldn’t blame those taken in, evaluation is hard)
This has bad implications (but wouldn’t blame those taken in, evaluation is hard)
Benchmarks from historians show that AI transcription from handwriting is now better than human, and a very cheap model is as good as people.
There are now massive troves of documents that could be made available for research that would have been impossible or prohibitive to transcribe before.
There are now massive troves of documents that could be made available for research that would have been impossible or prohibitive to transcribe before.
December 18, 2025 at 4:31 PM
Benchmarks from historians show that AI transcription from handwriting is now better than human, and a very cheap model is as good as people.
There are now massive troves of documents that could be made available for research that would have been impossible or prohibitive to transcribe before.
There are now massive troves of documents that could be made available for research that would have been impossible or prohibitive to transcribe before.
No signs of an end to rapid gains in AI ability at ever-decreasing costs (which is a log scale) yet. I have to update this monthly or more frequently at this point.
All AI benchmarks are flawed, but GPQA Diamond has been a pretty good one, though likely close to being maxed out.
All AI benchmarks are flawed, but GPQA Diamond has been a pretty good one, though likely close to being maxed out.
December 17, 2025 at 8:21 PM
No signs of an end to rapid gains in AI ability at ever-decreasing costs (which is a log scale) yet. I have to update this monthly or more frequently at this point.
All AI benchmarks are flawed, but GPQA Diamond has been a pretty good one, though likely close to being maxed out.
All AI benchmarks are flawed, but GPQA Diamond has been a pretty good one, though likely close to being maxed out.
"Gemini 3, create a really novel and clever and funny Venn diagram. think hard. do not do research."
So close to coming together (I am not sure the center works for all three, illustrations are odd), but also better than I expected.
So close to coming together (I am not sure the center works for all three, illustrations are odd), but also better than I expected.
December 16, 2025 at 11:35 PM
"Gemini 3, create a really novel and clever and funny Venn diagram. think hard. do not do research."
So close to coming together (I am not sure the center works for all three, illustrations are odd), but also better than I expected.
So close to coming together (I am not sure the center works for all three, illustrations are odd), but also better than I expected.
Surprisingly rapid & high Ai adoption by doctors: 67% use it daily, 84% says it makes them better doctors, 42% says it makes them want to stay in medicine more (10% said less). A lot of the use cases appear to be administrative and research assistance. 2025-physicians-ai-report.offcall.com
December 16, 2025 at 10:47 PM
Surprisingly rapid & high Ai adoption by doctors: 67% use it daily, 84% says it makes them better doctors, 42% says it makes them want to stay in medicine more (10% said less). A lot of the use cases appear to be administrative and research assistance. 2025-physicians-ai-report.offcall.com
I have found GPT-5.2 Thinking to be a surprisingly deep second-opinion/fact checker. I gave it a dense paragraph with a few correct claims, a couple errors that required research to find, and some things that needed interpretation
It found and gently corrected all the problems based on research.
It found and gently corrected all the problems based on research.
December 15, 2025 at 8:35 PM
I have found GPT-5.2 Thinking to be a surprisingly deep second-opinion/fact checker. I gave it a dense paragraph with a few correct claims, a couple errors that required research to find, and some things that needed interpretation
It found and gently corrected all the problems based on research.
It found and gently corrected all the problems based on research.
All the frontier AIs now pass all levels of the very challenging Chartered Financial Analyst (CFA) exam
The paper used paywalled, new mock exams to reduce the risk of leakage but AI grading for the essays. Interestingly, prompting strategy doesn't matter for recent models. arxiv.org/pdf/2512.08270
The paper used paywalled, new mock exams to reduce the risk of leakage but AI grading for the essays. Interestingly, prompting strategy doesn't matter for recent models. arxiv.org/pdf/2512.08270
December 15, 2025 at 4:35 PM
All the frontier AIs now pass all levels of the very challenging Chartered Financial Analyst (CFA) exam
The paper used paywalled, new mock exams to reduce the risk of leakage but AI grading for the essays. Interestingly, prompting strategy doesn't matter for recent models. arxiv.org/pdf/2512.08270
The paper used paywalled, new mock exams to reduce the risk of leakage but AI grading for the essays. Interestingly, prompting strategy doesn't matter for recent models. arxiv.org/pdf/2512.08270
Okay, they all solve it now.
December 13, 2025 at 9:53 PM
Okay, they all solve it now.
It is hard to separate AI economic impact out from other factors, but this is a clever attempt to get at the effect of AI on entrepreneurship by looking at which areas of China were more likely to adopt AI early.
They estimate that ChatGPT led to a 6% increase in startups. arxiv.org/pdf/2512.06506
They estimate that ChatGPT led to a 6% increase in startups. arxiv.org/pdf/2512.06506
December 13, 2025 at 2:29 PM
It is hard to separate AI economic impact out from other factors, but this is a clever attempt to get at the effect of AI on entrepreneurship by looking at which areas of China were more likely to adopt AI early.
They estimate that ChatGPT led to a 6% increase in startups. arxiv.org/pdf/2512.06506
They estimate that ChatGPT led to a 6% increase in startups. arxiv.org/pdf/2512.06506
Honestly weird that the frontier models do not diverge that much in terms of abilities, prompt adherence, and other factors. Whether you pick any of the big American closed source models or the Chinese and French open models, they are all very similar to each other, and have been consistently.
December 13, 2025 at 12:05 PM
Honestly weird that the frontier models do not diverge that much in terms of abilities, prompt adherence, and other factors. Whether you pick any of the big American closed source models or the Chinese and French open models, they are all very similar to each other, and have been consistently.
ChatGPT 5.2: "Build an interactive Excel spreadsheet where I can pick two D&D monsters to fight against each other and the spreadsheet simulates the combat somehow, including special abilities. Give a D&D look"
Thinking took 60 minutes(!) & had to have it fix an error, but impressive "game engine"
Thinking took 60 minutes(!) & had to have it fix an error, but impressive "game engine"
December 13, 2025 at 3:48 AM
ChatGPT 5.2: "Build an interactive Excel spreadsheet where I can pick two D&D monsters to fight against each other and the spreadsheet simulates the combat somehow, including special abilities. Give a D&D look"
Thinking took 60 minutes(!) & had to have it fix an error, but impressive "game engine"
Thinking took 60 minutes(!) & had to have it fix an error, but impressive "game engine"
Is there Jevons Paradox for AI: does cheaper AI lead to more total usage?
New paper finds short-run elasticity ~1 (so no short-run paradox) but prices fell 1000x in two years & demand exploded. So Jevons happens over time as firms gradually adopt AI at lower prices andreyfradkin.com/assets/LLM_D...
New paper finds short-run elasticity ~1 (so no short-run paradox) but prices fell 1000x in two years & demand exploded. So Jevons happens over time as firms gradually adopt AI at lower prices andreyfradkin.com/assets/LLM_D...
December 13, 2025 at 3:38 AM
Is there Jevons Paradox for AI: does cheaper AI lead to more total usage?
New paper finds short-run elasticity ~1 (so no short-run paradox) but prices fell 1000x in two years & demand exploded. So Jevons happens over time as firms gradually adopt AI at lower prices andreyfradkin.com/assets/LLM_D...
New paper finds short-run elasticity ~1 (so no short-run paradox) but prices fell 1000x in two years & demand exploded. So Jevons happens over time as firms gradually adopt AI at lower prices andreyfradkin.com/assets/LLM_D...
Whoa. This new GDPval score for GPT-5.2 is something.
GDPval is probably the most economically relevant measure of AI ability, suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans.
GDPval is probably the most economically relevant measure of AI ability, suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans.
December 11, 2025 at 6:52 PM
Whoa. This new GDPval score for GPT-5.2 is something.
GDPval is probably the most economically relevant measure of AI ability, suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans.
GDPval is probably the most economically relevant measure of AI ability, suggesting that in head-to-head competition with human experts on tasks that require 4-8 hours for a human to do, GPT-5.2 wins 71% of the time as judged by other humans.