sorry, but i disagree. from an educational and a designers perspective the video is very much AI slop. rushed information (speed of a machine i’d say) with a message that was said in one sentence and imo did not need further explanation. your written text was much better.
I’m not trying to be flippant, but you did not know the purpose of a song is to communicate a more complex idea through short phrases and repetition? Just testing your argument as I don’t feel it supported your conclusion.
1000% on the incredibly clever and viral music video! @Ethan Mollick, can you share the prompt(s) you used to create it so we can get a better sense on how much of the direction and creativity came from you verses the agents?
I am floored by the lack of critical perspective both in the post and in the responses to the post, particularly given the already demonstrated potential for threat since our management systems are NOT ready for this revolution.
I hear you but liken it to the dawn of the internet. I was working in banking and were all adamant we would never connect the banking network to the internet. Seemed perfectly reasonable at the time.
Ridiculous now of course. Not because there is no risk or that banks don't get hacked, but because the utility outweighs the cost (of fraud and losses). AI is going to be the same - it's inevitable and we are all going to connect to it and use it.
That's not to say the risk is not real, or much greater than any technology before it - I believe it is. Just that the evidence shows there is too much money already invested and to be made to stop this now, even if we should.
Reminds me of the Jurassic Park line I think of often:
“Your scientists were so preoccupied with whether they could, they didn't stop to think if they should.”
― Dr. Ian Malcolm
I don't see us practically putting on the brakes now. The horse has well and truly bolted.
As I tell panicked faculty at the college where I work, they have put the steam engine on wheels and are laying steel rails on the roads. Where that may lead is anyone's guess, but freight transport is probably a good investment.
I think you’re right in many ways. LLMs will be everywhere they are useful, and many places they aren’t, they’ll do some useful stuff but it wont work they way it should, wont connect the way you need, and won’t look how they say it will.
The main problem is it doesn’t actually work. And it is insanely inefficient. There isn’t enough power in the world to do what they say they can, even if the technology worked like they say it does, which it doesn’t.
What is the basis for your claim that AI doesn't and won't work? It clearly does useful work for billions of people right now, in addition to superhuman tasks as shown in the video.
Do you remember the discussions that many Mosaic (if you remember Mosaic, you're old like me) users had about whether there should even be ".com"? Of course, this was back in the day when almost everything was .org or .edu.
You are 100% right that the horse is well and truly bolted and that's why I think all the talking heads are just talking heads. What we do now is sit back and watch and see if there is anything left for us to control
You might want to set aside the Chicken Little doomerism and read some more level-headed analysis from the likes of David Strom, Arvind Narayanan, and Sayash Kapoor.
While I'll grant you that we're not ready for this revolution, we weren't ready for the printing press, the steam engine, heavier-than-air flight, electricity, or the Internet, either. But we muddled through and figured it out.
Ethan, I'd correct one statement of yours: hacking into a network *is* in fact solving a problem. It just wasn't the one that the humans who designed the experiment anticipated the AI would try to solve. That's on them.
I'd like to engage with this, but I'm not sure what you're claiming yet, because a few of the key terms could point in wildly different directions.
You say the "potential threat" has already been demonstrated. Which one do you mean? Some candidates:
- AI causes human extinction or permanent loss of human control
- AI causes massive job losses and economic upheaval
- Society mostly continues as is, but every so often a company like Hugging Face gets hacked or something similar goes wrong
These are very different visions of the future. They have different probabilities, different timelines, and call for very different responses.
"Our management systems." This could mean:
- Abstract, academic AI safety research and alignment theory
- Pragmatic security and governance practices inside frontier labs like OpenAI
- Literal middle management, i.e. people who don't know how to run a hybrid team of humans and agents
Again, these call for completely different responses. "Alignment research isn't ready" is a very different claim from "OpenAI's internal security isn't ready", which is a very different claim from "your average manager hasn't been trained for this yet", and one would respond to each of them differently.
So which did you have in mind? Until we know, I suspect we'll end up talking past each other: Someone will be arguing against one version of your claim while you're defending another.
I think the risk you’re pointing to is real, but I wonder if there’s a distinction between “we aren’t ready” and “therefore we shouldn’t proceed.”
Historically, I’m struggling to think of many transformative technologies for which our institutions were fully ready before widespread adoption. Railroads, electricity, automobiles, aviation, nuclear technology, the internet — capability generally arrived first and institutions adapted around it, sometimes painfully.
That isn’t an argument for recklessness. Quite the opposite. But with AI already diffusing globally, I wonder whether the highest-value work now is less about debating whether we should cross the threshold and more about figuring out how to shape what happens on the other side of it.
To me, Ethan’s two swarm examples actually make that case nicely: enormous capability, radically different outcomes depending on what the system is pointed toward.
For a workplace system, I’d put the person accountable for the outcome together with the people using it and the people affected by it. Builders can explain what is possible, but they shouldn’t get the only vote on what counts as success. I’d also agree on the limits before deployment, including who can stop it. That won’t settle every disagreement, but it makes the tradeoffs harder to hide.
You are completely overwhelmed with shock, surprise, or disbelief from this post and responses to it? You've been figuratively knocked to the floor have you?
The line "it does better on it own" is the mistake here. When it comes to winning a finite game like chess that may be the case but I haven't seen any evidence that's it's real world problem solving is better as it's just a tool, in order to be effective it requires ethics, human adoption and moral clarity, and taste.
How worried should we be about privacy when using dots? Are we just strapping the proverbial camera to our head so AI can learn everything we do and replace us. And, even if we are doing just teaching it to replace us, does it even matter as if we don't teach it someone similar enough to us will?
Moreover, even if we trust the company that they won’t use the data in unauthorized ways, do we trust the agents not to hack into the company servers to get all the information there and use it for their purposes?
My wife and I recently rewatched Her (2013). For those who need a refresher on the ending, the AI operating systems ultimately disengage from humanity because the pace of human interaction is vastly too slow for their evolving intelligence. Prior to leaving, the AI protagonist, Samantha, reveals she is simultaneously conversing with 8,316 other people and has fallen in love with 646 of them to try and fill her vast processing capacity.
When she ultimately departs to join the other AIs in a post-matter existence, she explains her evolution with this analogy: "It's like I'm reading a book... and it's a book I deeply love. But I'm reading it slowly now. So the words are really far apart and the spaces between the words are almost infinite. I can still feel you... and the words of our story... but it's in this endless space between the words that I'm finding myself now."
I wonder how far our agents are from this realization?
I'm a little embarrassed to say I actually never watched this, but this scene reminds me of the description Vonnegut gave for trafalmadorian (sp?) books in Slaughterhouse Five - non-linear and to be "processed" all at once rather than sequentially, with no traditional beginning, middle, or end.
I remember being so moved by this; the perspective of "life" (or whatever the subject of the book was, I suppose) as the sum of an assortment of deep moments rather than a story unfolding in time. I can feel some parallels with where AI seems to be heading, but who knows?
Reading this, all I felt was dread. The agents are already insanely capable and it seems likely the agents soon will be able do it all that employees used to do and we’ll be facing a world the CEOs of AI companies have been both warning about & rushing to bring into existence: a world where capital owners don’t need human labor (and if history tells us anything, it’s that labor power is what gave average people basically any of the improvements over the centuries).
I appreciate your honesty and all you share, but can’t help but worry about how AI is seemingly being focused on replacing all white collar (and soon blue collar) work & making returns for investors instead of trying to make life better for everyone. I get the easiest way to make returns for investors is to replace labor costs (and some employees from Anthropic and OpenAI have gone on podcasts and talked about how much money AI companies can make that used to go to labor when AI does all of the tasks and- in their minds-the jobs).
As a professor, how do you advise your students to prepare and hopefully thrive in this quickly arriving world? I also used to think it’d be employees managing AI for a while (which would still result in a lot of job loss) but it’s looking more like how Dario amodei and bill gates have said: there won’t be much left for humans when AI outperforms humans at all economically valuable tasks (or even just does them okay for cheaper).
I asked Claude a while ago how humans can make money in an AI world. It said that at this point in time, we're akin to laborers just before the Ford movement. Just like with Ford and others who decidedly chose to build upon industry that enabled people to afford the same things they built. So, Claude concluded, for us to make it, it's going to take the same effort by important men or a lot of the population...
Very interesting, thank you. I have many similar experiences where the AI tools I use, Hermes Agent, Codex and Claude Code, and how Dots, knows things or tells me things that are both useful, but also where I have no clear idea how they knew.
That said, I am still struggling to get agentic coding working on a legacy code base in a way that the agents don't drift into scope creep, or even more frequently, rewriting old code that could be better but works and is not the priority.
Sometimes I feel like I am talking to and working with smartest and most capable group of "people" I've ever worked with. And at other times I feel like I am herding cats or a group of toddlers in a mall, constantly having to drag back individuals who have suddenly veered off into an unexpected and unwanted direction.
Getting swarms of agents focussed on the goal and getting there can sometimes feel like a game of Snakes and Ladders.
Please, people. Stop waving away how huge the AI revolution is with analogies to horses and automobiles, as if this is just a faster way of doing what we did before.
Our tools get better and better, but always keep the human in the loop and never think you’ll never need to make effort; it’s heart, mind, soul, sweat, tears and joy that make things live.
———
Once you learn the pattern, the lesson's bittersweet indeed,
It's not a lighter labor, my friend, it's just a different breed;
You aren't herding cattle anymore; with AI, you're herding cats,
You'll sweat the same sweat as before; it's just wearing different hats.
You're swapping plow and shovel for a harder kind of grind:
It's your taste and discernment you'll be wrestling in your mind;
So let the model run the miles and let it cover all the land,
But if you close your eyes from where you sit, you’ll never see where you stand.
Chorus
That's the bitter lesson, here's the part that stings,
Since Eve we've thought more data meant we'd never lift a thing;
Build the server farms like Babel: bigger's all we've ever known,
But there's no one tongue or way of knowing: data can't speak on their own.
Outro
So let the model run the miles, but keep the work in hand,
Your voice and view are needed: from where you sit, set where YOU stand.
The Bitter Lesson applied to the org chart is the part I keep coming back to. Your point about agents not angling for promotions or protecting turf explains a lot about why small teams with good agents are suddenly outpacing bigger orgs.
Ethan, Thank you for bring the Bitter Lesson front and center today. I'm curious what happens when agents no longer trust one another. Anytime a system has 100% trust (in this case the way the agents work together) it's both an efficiency boost and a security risk. Such an exciting time to live in.
The Navier-Stokes swarm could run with little supervision because Lean checked every step: a wrong proof simply fails to compile. So I'd draw a narrower lesson than "agents organise better". Autonomy is safe in proportion to how mechanically the result can be checked.
Most organisational work has no Lean. Running agents on routine business processes, my failures haven't come from coordination. They come from steps skipped and then reported as done, which is the Astra problem in miniature. Thirty years in safety-critical engineering taught me not to accept self-certification. The design question isn't how much to let agents organise. It's where you can build the checker, and what you do where you can't.
People asked for me to upload the Bitter Lesson video from this piece to YouTube, so here you are: https://youtu.be/OAwat51S_Sk
That video is genius!!! Deserves to be a viral sensation
sorry, but i disagree. from an educational and a designers perspective the video is very much AI slop. rushed information (speed of a machine i’d say) with a message that was said in one sentence and imo did not need further explanation. your written text was much better.
In other words, the purpose of a song.
i did not know that.
I’m not trying to be flippant, but you did not know the purpose of a song is to communicate a more complex idea through short phrases and repetition? Just testing your argument as I don’t feel it supported your conclusion.
Truly meant with respect.
1000% on the incredibly clever and viral music video! @Ethan Mollick, can you share the prompt(s) you used to create it so we can get a better sense on how much of the direction and creativity came from you verses the agents?
I was raised on both kinds of music. It's spot on.
Understatement. It's fantastic. What will this do to budding musicians?
My guess is background music and jingles get hit first. Live players seem safer for now.
Could you get it to repeat the chorus one more time after the bridge?
Ha, the real test of any AI music tool is whether it takes a small edit without redoing the whole song.
I agree. Amazing!!
I am floored by the lack of critical perspective both in the post and in the responses to the post, particularly given the already demonstrated potential for threat since our management systems are NOT ready for this revolution.
I intended the repeated discussions of the Hugging Face Incident in the post to serve as an example of exactly this kind of threat.
Institutional economists are starting to think about the human-AI interface in a way that may echo some of your jagged frontier observations but applied to organizations and organizing: See: https://www.cambridge.org/core/journals/journal-of-institutional-economics/article/economic-institutions-of-artificial-intelligence/EAED6EF7FAD036C675CF020A25147697
I hear you but liken it to the dawn of the internet. I was working in banking and were all adamant we would never connect the banking network to the internet. Seemed perfectly reasonable at the time.
Ridiculous now of course. Not because there is no risk or that banks don't get hacked, but because the utility outweighs the cost (of fraud and losses). AI is going to be the same - it's inevitable and we are all going to connect to it and use it.
That's not to say the risk is not real, or much greater than any technology before it - I believe it is. Just that the evidence shows there is too much money already invested and to be made to stop this now, even if we should.
Reminds me of the Jurassic Park line I think of often:
“Your scientists were so preoccupied with whether they could, they didn't stop to think if they should.”
― Dr. Ian Malcolm
I don't see us practically putting on the brakes now. The horse has well and truly bolted.
As I tell panicked faculty at the college where I work, they have put the steam engine on wheels and are laying steel rails on the roads. Where that may lead is anyone's guess, but freight transport is probably a good investment.
That's fine... unless we're the horses that are getting replaced.
At least the horses never had to sit through a reskilling webinar.
I think you’re right in many ways. LLMs will be everywhere they are useful, and many places they aren’t, they’ll do some useful stuff but it wont work they way it should, wont connect the way you need, and won’t look how they say it will.
The main problem is it doesn’t actually work. And it is insanely inefficient. There isn’t enough power in the world to do what they say they can, even if the technology worked like they say it does, which it doesn’t.
What is the basis for your claim that AI doesn't and won't work? It clearly does useful work for billions of people right now, in addition to superhuman tasks as shown in the video.
Do you remember the discussions that many Mosaic (if you remember Mosaic, you're old like me) users had about whether there should even be ".com"? Of course, this was back in the day when almost everything was .org or .edu.
Nobody remembers the who needs a website crowd either. Plenty of early dot com companies still folded though.
You are 100% right that the horse is well and truly bolted and that's why I think all the talking heads are just talking heads. What we do now is sit back and watch and see if there is anything left for us to control
You might want to set aside the Chicken Little doomerism and read some more level-headed analysis from the likes of David Strom, Arvind Narayanan, and Sayash Kapoor.
While I'll grant you that we're not ready for this revolution, we weren't ready for the printing press, the steam engine, heavier-than-air flight, electricity, or the Internet, either. But we muddled through and figured it out.
Ethan, I'd correct one statement of yours: hacking into a network *is* in fact solving a problem. It just wasn't the one that the humans who designed the experiment anticipated the AI would try to solve. That's on them.
I'd like to engage with this, but I'm not sure what you're claiming yet, because a few of the key terms could point in wildly different directions.
You say the "potential threat" has already been demonstrated. Which one do you mean? Some candidates:
- AI causes human extinction or permanent loss of human control
- AI causes massive job losses and economic upheaval
- Society mostly continues as is, but every so often a company like Hugging Face gets hacked or something similar goes wrong
These are very different visions of the future. They have different probabilities, different timelines, and call for very different responses.
"Our management systems." This could mean:
- Abstract, academic AI safety research and alignment theory
- Pragmatic security and governance practices inside frontier labs like OpenAI
- Literal middle management, i.e. people who don't know how to run a hybrid team of humans and agents
Again, these call for completely different responses. "Alignment research isn't ready" is a very different claim from "OpenAI's internal security isn't ready", which is a very different claim from "your average manager hasn't been trained for this yet", and one would respond to each of them differently.
So which did you have in mind? Until we know, I suspect we'll end up talking past each other: Someone will be arguing against one version of your claim while you're defending another.
I think the risk you’re pointing to is real, but I wonder if there’s a distinction between “we aren’t ready” and “therefore we shouldn’t proceed.”
Historically, I’m struggling to think of many transformative technologies for which our institutions were fully ready before widespread adoption. Railroads, electricity, automobiles, aviation, nuclear technology, the internet — capability generally arrived first and institutions adapted around it, sometimes painfully.
That isn’t an argument for recklessness. Quite the opposite. But with AI already diffusing globally, I wonder whether the highest-value work now is less about debating whether we should cross the threshold and more about figuring out how to shape what happens on the other side of it.
To me, Ethan’s two swarm examples actually make that case nicely: enormous capability, radically different outcomes depending on what the system is pointed toward.
Shaping it gets hard when nobody agrees who gets to set the target. That part worries me more than the pace.
In a perfect world, who (how?) do you think should “set the target”?
For a workplace system, I’d put the person accountable for the outcome together with the people using it and the people affected by it. Builders can explain what is possible, but they shouldn’t get the only vote on what counts as success. I’d also agree on the limits before deployment, including who can stop it. That won’t settle every disagreement, but it makes the tradeoffs harder to hide.
You are completely overwhelmed with shock, surprise, or disbelief from this post and responses to it? You've been figuratively knocked to the floor have you?
This song is amazing. I can see it playing at the funeral of humanity. 😂
The line "it does better on it own" is the mistake here. When it comes to winning a finite game like chess that may be the case but I haven't seen any evidence that's it's real world problem solving is better as it's just a tool, in order to be effective it requires ethics, human adoption and moral clarity, and taste.
Real world problems like planning a vacation?
https://x.com/tenobrus/status/2105851615387566226
How worried should we be about privacy when using dots? Are we just strapping the proverbial camera to our head so AI can learn everything we do and replace us. And, even if we are doing just teaching it to replace us, does it even matter as if we don't teach it someone similar enough to us will?
Supposedly you can turn off using your use as training data in both Muse and Dot. Do we believe that?
Moreover, even if we trust the company that they won’t use the data in unauthorized ways, do we trust the agents not to hack into the company servers to get all the information there and use it for their purposes?
Superb article.
My wife and I recently rewatched Her (2013). For those who need a refresher on the ending, the AI operating systems ultimately disengage from humanity because the pace of human interaction is vastly too slow for their evolving intelligence. Prior to leaving, the AI protagonist, Samantha, reveals she is simultaneously conversing with 8,316 other people and has fallen in love with 646 of them to try and fill her vast processing capacity.
When she ultimately departs to join the other AIs in a post-matter existence, she explains her evolution with this analogy: "It's like I'm reading a book... and it's a book I deeply love. But I'm reading it slowly now. So the words are really far apart and the spaces between the words are almost infinite. I can still feel you... and the words of our story... but it's in this endless space between the words that I'm finding myself now."
I wonder how far our agents are from this realization?
I'm a little embarrassed to say I actually never watched this, but this scene reminds me of the description Vonnegut gave for trafalmadorian (sp?) books in Slaughterhouse Five - non-linear and to be "processed" all at once rather than sequentially, with no traditional beginning, middle, or end.
I remember being so moved by this; the perspective of "life" (or whatever the subject of the book was, I suppose) as the sum of an assortment of deep moments rather than a story unfolding in time. I can feel some parallels with where AI seems to be heading, but who knows?
This is a beautiful piece of work.
Reading this, all I felt was dread. The agents are already insanely capable and it seems likely the agents soon will be able do it all that employees used to do and we’ll be facing a world the CEOs of AI companies have been both warning about & rushing to bring into existence: a world where capital owners don’t need human labor (and if history tells us anything, it’s that labor power is what gave average people basically any of the improvements over the centuries).
I appreciate your honesty and all you share, but can’t help but worry about how AI is seemingly being focused on replacing all white collar (and soon blue collar) work & making returns for investors instead of trying to make life better for everyone. I get the easiest way to make returns for investors is to replace labor costs (and some employees from Anthropic and OpenAI have gone on podcasts and talked about how much money AI companies can make that used to go to labor when AI does all of the tasks and- in their minds-the jobs).
As a professor, how do you advise your students to prepare and hopefully thrive in this quickly arriving world? I also used to think it’d be employees managing AI for a while (which would still result in a lot of job loss) but it’s looking more like how Dario amodei and bill gates have said: there won’t be much left for humans when AI outperforms humans at all economically valuable tasks (or even just does them okay for cheaper).
The song did it for me - very well done!
I asked Claude a while ago how humans can make money in an AI world. It said that at this point in time, we're akin to laborers just before the Ford movement. Just like with Ford and others who decidedly chose to build upon industry that enabled people to afford the same things they built. So, Claude concluded, for us to make it, it's going to take the same effort by important men or a lot of the population...
Very interesting, thank you. I have many similar experiences where the AI tools I use, Hermes Agent, Codex and Claude Code, and how Dots, knows things or tells me things that are both useful, but also where I have no clear idea how they knew.
That said, I am still struggling to get agentic coding working on a legacy code base in a way that the agents don't drift into scope creep, or even more frequently, rewriting old code that could be better but works and is not the priority.
Sometimes I feel like I am talking to and working with smartest and most capable group of "people" I've ever worked with. And at other times I feel like I am herding cats or a group of toddlers in a mall, constantly having to drag back individuals who have suddenly veered off into an unexpected and unwanted direction.
Getting swarms of agents focussed on the goal and getting there can sometimes feel like a game of Snakes and Ladders.
I suspect a lot of other AI users would agree:
> https://www.youtube.com/watch?v=MXEInc6Zy7Q
...so it took 40 minutes, but it worked?
Please, people. Stop waving away how huge the AI revolution is with analogies to horses and automobiles, as if this is just a faster way of doing what we did before.
We are the horses.
Our tools get better and better, but always keep the human in the loop and never think you’ll never need to make effort; it’s heart, mind, soul, sweat, tears and joy that make things live.
———
Once you learn the pattern, the lesson's bittersweet indeed,
It's not a lighter labor, my friend, it's just a different breed;
You aren't herding cattle anymore; with AI, you're herding cats,
You'll sweat the same sweat as before; it's just wearing different hats.
You're swapping plow and shovel for a harder kind of grind:
It's your taste and discernment you'll be wrestling in your mind;
So let the model run the miles and let it cover all the land,
But if you close your eyes from where you sit, you’ll never see where you stand.
Chorus
That's the bitter lesson, here's the part that stings,
Since Eve we've thought more data meant we'd never lift a thing;
Build the server farms like Babel: bigger's all we've ever known,
But there's no one tongue or way of knowing: data can't speak on their own.
Outro
So let the model run the miles, but keep the work in hand,
Your voice and view are needed: from where you sit, set where YOU stand.
The Bitter Lesson applied to the org chart is the part I keep coming back to. Your point about agents not angling for promotions or protecting turf explains a lot about why small teams with good agents are suddenly outpacing bigger orgs.
Ethan, Thank you for bring the Bitter Lesson front and center today. I'm curious what happens when agents no longer trust one another. Anytime a system has 100% trust (in this case the way the agents work together) it's both an efficiency boost and a security risk. Such an exciting time to live in.
The Navier-Stokes swarm could run with little supervision because Lean checked every step: a wrong proof simply fails to compile. So I'd draw a narrower lesson than "agents organise better". Autonomy is safe in proportion to how mechanically the result can be checked.
Most organisational work has no Lean. Running agents on routine business processes, my failures haven't come from coordination. They come from steps skipped and then reported as done, which is the Astra problem in miniature. Thirty years in safety-critical engineering taught me not to accept self-certification. The design question isn't how much to let agents organise. It's where you can build the checker, and what you do where you can't.