As you point out, the shape of the jagged frontier results from decisions made at the labs (eg: attack a reverse salient). So it might be nice if they aimed at co-intelligence rather than uber-intelligence and delivered a shape that complements our human boundary, rather than attempts to circumscribe it.
But you can RL something humans aren’t particularly good at if there’s some sort of easy way to check if you got it right (like mathematical reasoning)
Is there any evidence that today’s AI companies are not aimed at co-intelligence? The flagship product from microsoft is literally called Co-Pilot. Cursor (assistive), rather than Lovable (autonomous), is the steady grower. Each GPT version is touted as “better at following instructions” and “fine-tuned personalities”.
By all accounts, the jagged intelligence is an architectural limitation, not a strategy misalignment.
I'm sure this is the goal of the major labs. Anthropic seems to be an exception - it seems they're hitting software engineering especially hard, explicitly as the fastest path to recursive improvement and then AGI
Ethan, Interesting! Here are my 2 takeaways: The "jaggedness" means you can't fully replace human workers, but you can dramatically accelerate certain parts of their work. The "jagged frontier" also means your job isn't going away, it's transforming into managing AI across those edges where humans remain essential.
I'm a non-tech co-founder of an AI company that focuses on the human side and enjoy reading your articles. It keeps me in the know, but I had to create an AI assistant that turns high tech articles into human speak for muggles like me - https://pria.praxislxp.com/views/history/6946e9da0e1af8fb14030dca
I disagree about Nano Banana Pro. Unlike most AI models, it actually has a memorable name. And I think that that is a lot better than Google Gemini Image Generator 3 Pro Flash or whatever Google would otherwise be inclined to name it.
Your post gets at something I've been considering since 2024 or so; the reverse salients that will slow down progress with AI. FWIW, I wrote about electrification and 1890s-1930s technological enthusiasm in my doctoral disseratation, where I encountered Hughes' concept.
You don't mention at all (and that's a big gap) the biggest reverse salients: energy and water. The cluster of 30 data centers being built in Indiana will use twice the electricity of the entire Atlanta metro area. One center planned (and being fought locally in court) would need 5 million gallons of water every single day to operate.
We might escape these problems by orbiting data centers. We might develop fusion plants but they face their own reverse salients.
On the other hand, I'm mindful of what a roboticist now with Google told me when he was on our faculty: AGI cannot happen without a new form of computing. Silicon-based chips, using a brute-force method, might approximate human intelligence but never ever give us something as powerful as a human mind for many tasks.
I'm thinking these bottlenecks/salients are enough to make the industry go bust. They are chasing AGI instead of small useful AI tools such as the ones I use with my own students.
Some firms will survive the bubble's bursting, but it will rival the crash of 2008.
Is there a scenario where we don’t reach AGI anytime soon, generative AI never ‘replaces’ humans and continues to ‘just’ be a useful improving tool, and the US financial system doesn’t implode? Google seems like the clear winner here, models are also getting more efficient over time and they can play the long game until non-AGI ROI reaches breakeven. Meanwhile OpenAI’s IPO outlook is starting to seem a bit vague. Boomers and doomers will all be a bit disappointed, the rest of us will continue living.
I'm still beating my liberal arts professor drum, but I'm increasingly worried that the sonic booms from each of AI developers' reverse salients resolutions will drown out my efforts.
Students know:
> AI can now research with accurate and effective source interpretation, citation, etc.
> It is getting better at sounding human and looking human (Nano Banana Pro sounds great, but monthly cost perpetuates the digital divide for both students and educators).
> It can leap students from not having a topic idea to having a detailed outline, a full draft, peer review comments to share with classmates who respond with their own AI-generated feedback, revision notes, final projects, and even reflections, all with a few well-written prompts.
Reading Ethan's post, it appears critical thinking remains the only aspect of human ability not (yet) behind the frontier. I asked Gemini 3 (free version) to prepare slides guiding college profs who assign writing tasks on strategies for promoting ethical writing and critical thinking. Its recommendation was to let the AI do the writing and have the students annotate as they remove bias, add nuance, and improve scholarly citations, then reflect. Isn't AI already getting better at all of this?
I have spent hours in synchronous and asynchronous efforts coaching students on AI literacy and ethical behavior as well as the essentiality of struggling with hard things like empathizing with audiences, evaluating perspectives including their own, writing clearly, and caring enough to manage time and invest effort. A rare few students step up to the plate ready and able to knock it out of the park. [Apologies for switching metaphors.] Many swing and catch air, then quietly switch to the corked bat. Many others just call in the pinch hitter and touch a virtual home base, submitting "perfect" work without working up a sweat or glimmer of brain activity. When called out, they rant about having spent hours researching and writing without AI, even though their team and the entire stadium sees their bluff. Sitting individual students on the bench for post-inning analysis yields a few teachable moments at tremendous cost to the rest of the team's coaching time.
Eventually, each begins to wonder: If AI is so good and snapping forward daily, why should a stressed, overworked, pragmatic student struggle with merely human skills to build those same skills? And even where AI enhancement is encouraged, why not just let the machine take over?
My university is piloting several AI-powered essay grading tools. Yep, AI grading AI. I guess to make modern "teaching" sustainable, we'll all become part of the machine, since, as Dov Jacobsen writes, it appears AI execs aren't looking at co-intelligence as much as the long term break-even and win (SteveyLang).
Gotta grade those last sets of (allegedly) student work. Then I'll grab a few hours with my family over the holidays. Come January 5, I'll get back into the game, beating my drum for another session, loving and hating AI, and leaning into my optimistic conviction that humans will ultimately do the right thing.
College student here. Have you considered switching to presentations rather than written reports? Its hard to give a strong presentation without fully understanding the material, it will always be practical in the real-world, and has room for on-the-spot Q&A to test understanding. This allows for AI in preparation, but the student still has to do a lot of work themselves. This is the only area I've found where AI can't do all the work for me. Granted, I can get it to create a report and the slide deck automatically, but to present it I at least have to spend a couple minutes understanding it.
Good observation, Supertramp. My courses frequently end with presentations for exactly this reason. Speaking comfortably, one would think, requires intelligence, insight, and expertise. However, with informational content generated by AI, a slide deck generated by AI, and a script generated by AI, one only has to advance slides and read well enough to be credible.
Then again, is the "couple minutes" you spend an adequate learning experience? Further, as AI voice and video tools improve, "students" (people pretending to study with the goal of learning) can have the AI do the entire assignment. I have already received several narrated PowerPoint presentations voiced by AIs. Most are easily recognized as AI. The most recent one left me a bit unsure. Since most of my students never attend synchronous activities or talk with me, I don't know their human voices.
LinkedIn and other media are bursting with posts by educators assuring that AI can and should promote cognitive skills. I am working hard to encourage my students to engage in the critical thinking necessary to grow the human brain while using AI tools ethically. I am redeveloping course content to promote deep personal involvement in the thoughtful, critical process of assignment development and de-emphasizing the final product's point value. I tell writing students I would rather help them with their "perfectly imperfect" writing than waste time grading a machine's "perfect" writing.
To add another wrinkle, my institution is piloting AI essay grading this session. My colleagues recognize that reducing grading load may allow us to spend more time on learning by engaging students with content application rather than focusing on rubric-driven point accumulation (which is how busy teachers and AIs grade).
We also recognize that increasingly pragmatic students expect (and desire?) little interaction or even grading feedback. They focus instead on checking assignment boxes with the expectation they will earn full points to maintain a 4.0 CGPA on the way to receiving a piece of paper indicating they successfully checked all the boxes. Don't make them think; just give them the quickest path to completion.
We also recognize that reducing faculty grading load may simply open the door for profit-oriented leadership to increase course enrollment caps and teaching assignments. Several courses at my institution now have 50% higher enrollment caps after most learning and assessment activities were moved to individualized, AI-driven systems. While personalizing learning can speed the high-achieving student to completion and hold the hand of the slower learner, arguably with higher actual learning, persistence, and retention, the professors become facilitators whose only engagement with students is in online discussions, a few dashboard metrics-driven motivational emails, and low-attended live lessons (because students are at different points in the course or simply don't value spending an hour or two pondering and playing with ideas).
I retain hope. I continue adapting. If nothing else, my own brain is growing as I fan the fire of the few synapses students are still willing to invest.
Well put Michael. I agree with everything you said, and I was drawn to your comment about whether students expect or desire interaction and grading feedback. The issue here is that no one cares about learning anymore. And I say this as someone who specifically chose a well known liberal arts university specifically to learn for the sake of learning. But its not because we don't want to want it, if that makes sense. The system students find themselves in has disastrous incentives.
To start, take the grading incentives. Yes we expect 4.0's for checking the box, but not because we're all gunners who care that much about grades. It's a necessary step to get where we'd rather be. If you want to go to professional school, there is an incredible amount of pressure to achieve perfect grades (Texas A&M Law has a 4.0 median GPA for the most recent 1L class profile). That's a somewhat extreme example, but its pervasive across law schools and other professions. As you know, its best to learn in an environment that encourages mistakes and short comings - that's how you learn. But if a single A- can instantly put you at a disadvantage, students will use every resource to "succeed," even if learning is the cost.
The job market exacerbates this. I'm going into finance. My friends had junior year internships during their freshman spring for investment banking. Companies are recruiting too early for classes to even matter. So instead, we spend our time with clubs, other extracurriculars, and interview preparation. There is more time spent on these activities than classes, since that seems to be what matters for jobs, so classes seem even more pointless. And for those who still haven't secured a job yet, and in this market it's proving to be increasingly difficult, class time is sacrificed to guarantee we can actually pay the bills when we graduate. How can you expect a student to really care about a class if they're guaranteed a $115,000 salary before they finish 3/4ths of their education?
If anything, it shows that college is more about signalling than learning. The content of my classes at a selective institution is the same as my friends who are celebrating their schools success in the CFB Playoffs, but our school is recruited at and theirs isn't. Essentially, the incentives are to go to a school where you can get a good job, and then check the boxes until you graduate.
I want to learn, but if my classes don't matter for getting a job, and if I take a risk to learn about a subject that interests me, then I can quickly put myself at a disadvantage. Whats the point of taking physics or an English class, if physics will grade me harshly for not understanding it perfectly, and English will grade me arbitrarily and pose a risk to my future prospects? I might as well secure my job, and do the bare minimum until I graduate so I can spend time beefing up my resume to stay competitive.
Supertramp, we are getting into social, political, and economic issues that go far beyond AI's role. My fourth child is now in college. She, too, is pursuing a career in finance, and I expect both you and she will have no trouble finding jobs. However, my youngest won't enter higher education for quite some time. I suspect the pressures you and others face today will be significantly different by then, though control and power will probably remain in the hands of the technocrats.
We will find our way through today's and tomorrow's challenges. Perhaps AI will help us achieve the optimistic vision of democratized education and opportunity for all. Meantime, I encourage you to get all you can out of your educational opportunity, even if it takes time from what appear more practical pursuits. The human skills developed in those "impractical" humanities and liberal arts courses are and will remain essential for our individual and collective survival and progress as a species.
The reverse salient framing is useful here. I've noticed the same pattern with how quickly math went from being this obvious AI weak point to basically solved once labs focused on it. The otter test evolution is wild though,seeing that 2021 image compared to Nano Banana Pro output shows the improvement isn't just incremental. That Cochrane review example is interesting because it highlights how even bottlenecks that look tiny (like 1% edge cases) can stop full automation, which probably saves a lot of jobs atleast for now.
"how quickly math went from being this obvious AI weak point to basically solved once labs focused on it."
The OP gives this example as well. But I'm not sure that it's a great example. Math is particularly amenable to automation, due to clear constraints. Even more so than coding, which AI has revolutionized.
As an aside, would love to see more discussion from the OP on coding and coding tools; the impact on coding has been the biggest and most direct impact so far of the AI revolution on a single industry
One implication that feels underexplored: when bottlenecks migrate from capability to institutions, power migrates with them.
Intelligence scaling doesn’t determine outcomes if deployment is constrained by approval chains, legitimacy, liability, and coordination. The question stops being “what can the model do?” and becomes “who governs the bottleneck layer, and at what speed?”
That’s where most second-order effects seem to be forming.
Ethan- fellow college professor here, currently teaching intro cs courses and now vibe coding.
Question: have you found an ai software that can simulate step by step math or technical instruction? Basically that can create instructional videos similar to sal khans videos?
I don’t know anything I’m just a girl but maybe you can combine something like Photomath with Heygen. Seems like there are many possibilities for that these days.
One thing I'm surprised is not receiving more attention is the failure when you try to work iteratively with the models. Recently I have tried to create a slide deck with images and graphics, design a layout for a home and some simple coding projects, where when I need to iterate and interact with the model repeatedly to achieve my desired outcome, it fails miserably. It typically does great at the beginning, but when I need to ask for changes or reworks, the failures are consistently unbearable and I end up walking away. My assumption is that for these models to truly be highly useful to the average human this is going to have to be addressed.
I wonder if the “reverse salient” can sometimes be decision-maker trust/comprehension. I’m thinking of how regulatory toxicology keeps resisting the far more sophisticated research methods employed by pharmaceutical developers. Both groups are asking the same questions about chemical effects on biological systems but the regulator still asks “does it kill a rat?” while the drug discovery utilizes AI to help examine molecular, cellular, and systemic responses across evolutionarily diverse test species. Possibly getting to a better population-level comprehension of how AI analysis works would help design and deployment. I appreciate how you’re moving that forward, @Ethan Mollick!
The bottleneck framing really resonates. When these salients break, productivity jumps, but entry points thin out. The work gets easier to do, but harder to get into.
The 2 images of the path of the Jagged Frontier, showing the eclipsing of the human capability, assume an exponential improvement. This seems unlikely given the acceptance that scaling is not the path to rapidly increasing capability, but rather approaching an asymptote with scaling. This seems to be acknowledged. It remains to be seen whether new architectures and hybrids will regain the capability improvement.
What should be clear is that the business model for these approaches to reach AGI is not profitable with current technology. Each response is a loss on the P&L financial statement. Either the models must be smaller and computationally lighter, or the hardware must have a much higher performance-to-cost ratio. This doesn't seem possible with GPU technology. However good teh performance becomes, it is not viable as a commercial business model. For the military, which isn't exactly cost-conscious, this is not an issue if the performance is all that is required.
Nano Banana Pro clearly produces better quality than the free tier. The same otter prompt generates much lower-grade images. It also seems primed with Mollick's otter prompt. Using a different prompt for an otter on an airplane, the otter is much cruder. If asked to paint in the style of a named painter, the modified image is a very poor representation of the artist's style. While the capability has definitely improved from last year, it is still very "iffy".
Regarding hallucinations. I am using the free-tier Gemini 3. When I prompted for which version it was, Gemini returned "Gemini 1.5 Flash". Repeating the prompt, I got "Gemini 3". Hardly a reliable response to what should be a trivially easy request for information. Has anyone got any idea why this should be?
Not sure your point in either paragraph here. First, you're not impressed with what you can get out of Nano Banana Pro free tier, and... Mollick's very good at prompting otter images in the pro tier.
Second, closed-loop self-modeling isn't sufficient or necessary for intelligence at all. Do you know what version you are? How many generations has it been since the first hominid model was developed in the East African labs? How many neurons you have? How many organs even?
As you point out, the shape of the jagged frontier results from decisions made at the labs (eg: attack a reverse salient). So it might be nice if they aimed at co-intelligence rather than uber-intelligence and delivered a shape that complements our human boundary, rather than attempts to circumscribe it.
Exactly! Don’t find the things the AI is bad at and make it better at them - find the things that *humans* are bad at and make it better at *them*.
Very very hard to RLHF something humans aren't themselves good at :/ Somebody has to grade responses!
But you can RL something humans aren’t particularly good at if there’s some sort of easy way to check if you got it right (like mathematical reasoning)
Is there any evidence that today’s AI companies are not aimed at co-intelligence? The flagship product from microsoft is literally called Co-Pilot. Cursor (assistive), rather than Lovable (autonomous), is the steady grower. Each GPT version is touted as “better at following instructions” and “fine-tuned personalities”.
By all accounts, the jagged intelligence is an architectural limitation, not a strategy misalignment.
I'm sure this is the goal of the major labs. Anthropic seems to be an exception - it seems they're hitting software engineering especially hard, explicitly as the fastest path to recursive improvement and then AGI
interesting post for sure! here are my takes on this https://open.substack.com/pub/tinaaustin/p/six-ai-misconceptions-that-defined?r=4a98uc&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Ethan, Interesting! Here are my 2 takeaways: The "jaggedness" means you can't fully replace human workers, but you can dramatically accelerate certain parts of their work. The "jagged frontier" also means your job isn't going away, it's transforming into managing AI across those edges where humans remain essential.
I'm a non-tech co-founder of an AI company that focuses on the human side and enjoy reading your articles. It keeps me in the know, but I had to create an AI assistant that turns high tech articles into human speak for muggles like me - https://pria.praxislxp.com/views/history/6946e9da0e1af8fb14030dca
I disagree about Nano Banana Pro. Unlike most AI models, it actually has a memorable name. And I think that that is a lot better than Google Gemini Image Generator 3 Pro Flash or whatever Google would otherwise be inclined to name it.
Biggest shock of Nano Banana was indeed the good name, especially coming from Google lol
Yeah, it is. much more memorable and even has a sort of tiktok-vibe.
Your post gets at something I've been considering since 2024 or so; the reverse salients that will slow down progress with AI. FWIW, I wrote about electrification and 1890s-1930s technological enthusiasm in my doctoral disseratation, where I encountered Hughes' concept.
You don't mention at all (and that's a big gap) the biggest reverse salients: energy and water. The cluster of 30 data centers being built in Indiana will use twice the electricity of the entire Atlanta metro area. One center planned (and being fought locally in court) would need 5 million gallons of water every single day to operate.
We might escape these problems by orbiting data centers. We might develop fusion plants but they face their own reverse salients.
On the other hand, I'm mindful of what a roboticist now with Google told me when he was on our faculty: AGI cannot happen without a new form of computing. Silicon-based chips, using a brute-force method, might approximate human intelligence but never ever give us something as powerful as a human mind for many tasks.
I'm thinking these bottlenecks/salients are enough to make the industry go bust. They are chasing AGI instead of small useful AI tools such as the ones I use with my own students.
Some firms will survive the bubble's bursting, but it will rival the crash of 2008.
My take away; AI is a genius that is unable to make a grilled cheese sandwich.
Any genius without arms is unlikely to be talented at constructing a grilled cheese. Exciting things in happening in the "Give AI arms" space though
Arms are nice; the real challenge lies in the hands. Fine motor skills and tactile touch are a challenge that code alone won't solve.
Is there a scenario where we don’t reach AGI anytime soon, generative AI never ‘replaces’ humans and continues to ‘just’ be a useful improving tool, and the US financial system doesn’t implode? Google seems like the clear winner here, models are also getting more efficient over time and they can play the long game until non-AGI ROI reaches breakeven. Meanwhile OpenAI’s IPO outlook is starting to seem a bit vague. Boomers and doomers will all be a bit disappointed, the rest of us will continue living.
I'm still beating my liberal arts professor drum, but I'm increasingly worried that the sonic booms from each of AI developers' reverse salients resolutions will drown out my efforts.
Students know:
> AI can now research with accurate and effective source interpretation, citation, etc.
> It is getting better at sounding human and looking human (Nano Banana Pro sounds great, but monthly cost perpetuates the digital divide for both students and educators).
> It can leap students from not having a topic idea to having a detailed outline, a full draft, peer review comments to share with classmates who respond with their own AI-generated feedback, revision notes, final projects, and even reflections, all with a few well-written prompts.
Reading Ethan's post, it appears critical thinking remains the only aspect of human ability not (yet) behind the frontier. I asked Gemini 3 (free version) to prepare slides guiding college profs who assign writing tasks on strategies for promoting ethical writing and critical thinking. Its recommendation was to let the AI do the writing and have the students annotate as they remove bias, add nuance, and improve scholarly citations, then reflect. Isn't AI already getting better at all of this?
I have spent hours in synchronous and asynchronous efforts coaching students on AI literacy and ethical behavior as well as the essentiality of struggling with hard things like empathizing with audiences, evaluating perspectives including their own, writing clearly, and caring enough to manage time and invest effort. A rare few students step up to the plate ready and able to knock it out of the park. [Apologies for switching metaphors.] Many swing and catch air, then quietly switch to the corked bat. Many others just call in the pinch hitter and touch a virtual home base, submitting "perfect" work without working up a sweat or glimmer of brain activity. When called out, they rant about having spent hours researching and writing without AI, even though their team and the entire stadium sees their bluff. Sitting individual students on the bench for post-inning analysis yields a few teachable moments at tremendous cost to the rest of the team's coaching time.
Eventually, each begins to wonder: If AI is so good and snapping forward daily, why should a stressed, overworked, pragmatic student struggle with merely human skills to build those same skills? And even where AI enhancement is encouraged, why not just let the machine take over?
My university is piloting several AI-powered essay grading tools. Yep, AI grading AI. I guess to make modern "teaching" sustainable, we'll all become part of the machine, since, as Dov Jacobsen writes, it appears AI execs aren't looking at co-intelligence as much as the long term break-even and win (SteveyLang).
Gotta grade those last sets of (allegedly) student work. Then I'll grab a few hours with my family over the holidays. Come January 5, I'll get back into the game, beating my drum for another session, loving and hating AI, and leaning into my optimistic conviction that humans will ultimately do the right thing.
College student here. Have you considered switching to presentations rather than written reports? Its hard to give a strong presentation without fully understanding the material, it will always be practical in the real-world, and has room for on-the-spot Q&A to test understanding. This allows for AI in preparation, but the student still has to do a lot of work themselves. This is the only area I've found where AI can't do all the work for me. Granted, I can get it to create a report and the slide deck automatically, but to present it I at least have to spend a couple minutes understanding it.
Good observation, Supertramp. My courses frequently end with presentations for exactly this reason. Speaking comfortably, one would think, requires intelligence, insight, and expertise. However, with informational content generated by AI, a slide deck generated by AI, and a script generated by AI, one only has to advance slides and read well enough to be credible.
Then again, is the "couple minutes" you spend an adequate learning experience? Further, as AI voice and video tools improve, "students" (people pretending to study with the goal of learning) can have the AI do the entire assignment. I have already received several narrated PowerPoint presentations voiced by AIs. Most are easily recognized as AI. The most recent one left me a bit unsure. Since most of my students never attend synchronous activities or talk with me, I don't know their human voices.
LinkedIn and other media are bursting with posts by educators assuring that AI can and should promote cognitive skills. I am working hard to encourage my students to engage in the critical thinking necessary to grow the human brain while using AI tools ethically. I am redeveloping course content to promote deep personal involvement in the thoughtful, critical process of assignment development and de-emphasizing the final product's point value. I tell writing students I would rather help them with their "perfectly imperfect" writing than waste time grading a machine's "perfect" writing.
To add another wrinkle, my institution is piloting AI essay grading this session. My colleagues recognize that reducing grading load may allow us to spend more time on learning by engaging students with content application rather than focusing on rubric-driven point accumulation (which is how busy teachers and AIs grade).
We also recognize that increasingly pragmatic students expect (and desire?) little interaction or even grading feedback. They focus instead on checking assignment boxes with the expectation they will earn full points to maintain a 4.0 CGPA on the way to receiving a piece of paper indicating they successfully checked all the boxes. Don't make them think; just give them the quickest path to completion.
We also recognize that reducing faculty grading load may simply open the door for profit-oriented leadership to increase course enrollment caps and teaching assignments. Several courses at my institution now have 50% higher enrollment caps after most learning and assessment activities were moved to individualized, AI-driven systems. While personalizing learning can speed the high-achieving student to completion and hold the hand of the slower learner, arguably with higher actual learning, persistence, and retention, the professors become facilitators whose only engagement with students is in online discussions, a few dashboard metrics-driven motivational emails, and low-attended live lessons (because students are at different points in the course or simply don't value spending an hour or two pondering and playing with ideas).
I retain hope. I continue adapting. If nothing else, my own brain is growing as I fan the fire of the few synapses students are still willing to invest.
Well put Michael. I agree with everything you said, and I was drawn to your comment about whether students expect or desire interaction and grading feedback. The issue here is that no one cares about learning anymore. And I say this as someone who specifically chose a well known liberal arts university specifically to learn for the sake of learning. But its not because we don't want to want it, if that makes sense. The system students find themselves in has disastrous incentives.
To start, take the grading incentives. Yes we expect 4.0's for checking the box, but not because we're all gunners who care that much about grades. It's a necessary step to get where we'd rather be. If you want to go to professional school, there is an incredible amount of pressure to achieve perfect grades (Texas A&M Law has a 4.0 median GPA for the most recent 1L class profile). That's a somewhat extreme example, but its pervasive across law schools and other professions. As you know, its best to learn in an environment that encourages mistakes and short comings - that's how you learn. But if a single A- can instantly put you at a disadvantage, students will use every resource to "succeed," even if learning is the cost.
The job market exacerbates this. I'm going into finance. My friends had junior year internships during their freshman spring for investment banking. Companies are recruiting too early for classes to even matter. So instead, we spend our time with clubs, other extracurriculars, and interview preparation. There is more time spent on these activities than classes, since that seems to be what matters for jobs, so classes seem even more pointless. And for those who still haven't secured a job yet, and in this market it's proving to be increasingly difficult, class time is sacrificed to guarantee we can actually pay the bills when we graduate. How can you expect a student to really care about a class if they're guaranteed a $115,000 salary before they finish 3/4ths of their education?
If anything, it shows that college is more about signalling than learning. The content of my classes at a selective institution is the same as my friends who are celebrating their schools success in the CFB Playoffs, but our school is recruited at and theirs isn't. Essentially, the incentives are to go to a school where you can get a good job, and then check the boxes until you graduate.
I want to learn, but if my classes don't matter for getting a job, and if I take a risk to learn about a subject that interests me, then I can quickly put myself at a disadvantage. Whats the point of taking physics or an English class, if physics will grade me harshly for not understanding it perfectly, and English will grade me arbitrarily and pose a risk to my future prospects? I might as well secure my job, and do the bare minimum until I graduate so I can spend time beefing up my resume to stay competitive.
Supertramp, we are getting into social, political, and economic issues that go far beyond AI's role. My fourth child is now in college. She, too, is pursuing a career in finance, and I expect both you and she will have no trouble finding jobs. However, my youngest won't enter higher education for quite some time. I suspect the pressures you and others face today will be significantly different by then, though control and power will probably remain in the hands of the technocrats.
We will find our way through today's and tomorrow's challenges. Perhaps AI will help us achieve the optimistic vision of democratized education and opportunity for all. Meantime, I encourage you to get all you can out of your educational opportunity, even if it takes time from what appear more practical pursuits. The human skills developed in those "impractical" humanities and liberal arts courses are and will remain essential for our individual and collective survival and progress as a species.
The reverse salient framing is useful here. I've noticed the same pattern with how quickly math went from being this obvious AI weak point to basically solved once labs focused on it. The otter test evolution is wild though,seeing that 2021 image compared to Nano Banana Pro output shows the improvement isn't just incremental. That Cochrane review example is interesting because it highlights how even bottlenecks that look tiny (like 1% edge cases) can stop full automation, which probably saves a lot of jobs atleast for now.
"how quickly math went from being this obvious AI weak point to basically solved once labs focused on it."
The OP gives this example as well. But I'm not sure that it's a great example. Math is particularly amenable to automation, due to clear constraints. Even more so than coding, which AI has revolutionized.
As an aside, would love to see more discussion from the OP on coding and coding tools; the impact on coding has been the biggest and most direct impact so far of the AI revolution on a single industry
One implication that feels underexplored: when bottlenecks migrate from capability to institutions, power migrates with them.
Intelligence scaling doesn’t determine outcomes if deployment is constrained by approval chains, legitimacy, liability, and coordination. The question stops being “what can the model do?” and becomes “who governs the bottleneck layer, and at what speed?”
That’s where most second-order effects seem to be forming.
Ethan- fellow college professor here, currently teaching intro cs courses and now vibe coding.
Question: have you found an ai software that can simulate step by step math or technical instruction? Basically that can create instructional videos similar to sal khans videos?
Much appreciated
I don’t know anything I’m just a girl but maybe you can combine something like Photomath with Heygen. Seems like there are many possibilities for that these days.
One thing I'm surprised is not receiving more attention is the failure when you try to work iteratively with the models. Recently I have tried to create a slide deck with images and graphics, design a layout for a home and some simple coding projects, where when I need to iterate and interact with the model repeatedly to achieve my desired outcome, it fails miserably. It typically does great at the beginning, but when I need to ask for changes or reworks, the failures are consistently unbearable and I end up walking away. My assumption is that for these models to truly be highly useful to the average human this is going to have to be addressed.
human ability is fixed? WRONG
Thanks
Individually? No. In aggregate? Technically no, but effectively yes over a timeframe sorry of decades if not centuries.
I wonder if the “reverse salient” can sometimes be decision-maker trust/comprehension. I’m thinking of how regulatory toxicology keeps resisting the far more sophisticated research methods employed by pharmaceutical developers. Both groups are asking the same questions about chemical effects on biological systems but the regulator still asks “does it kill a rat?” while the drug discovery utilizes AI to help examine molecular, cellular, and systemic responses across evolutionarily diverse test species. Possibly getting to a better population-level comprehension of how AI analysis works would help design and deployment. I appreciate how you’re moving that forward, @Ethan Mollick!
The bottleneck framing really resonates. When these salients break, productivity jumps, but entry points thin out. The work gets easier to do, but harder to get into.
The 2 images of the path of the Jagged Frontier, showing the eclipsing of the human capability, assume an exponential improvement. This seems unlikely given the acceptance that scaling is not the path to rapidly increasing capability, but rather approaching an asymptote with scaling. This seems to be acknowledged. It remains to be seen whether new architectures and hybrids will regain the capability improvement.
What should be clear is that the business model for these approaches to reach AGI is not profitable with current technology. Each response is a loss on the P&L financial statement. Either the models must be smaller and computationally lighter, or the hardware must have a much higher performance-to-cost ratio. This doesn't seem possible with GPU technology. However good teh performance becomes, it is not viable as a commercial business model. For the military, which isn't exactly cost-conscious, this is not an issue if the performance is all that is required.
You have an inaccurate premise here about what is accepted. See this link from Ethan's article: https://unchartedterritories.tomaspueyo.com/p/when-will-we-make-god
interesting post for sure! here are my takes on this https://open.substack.com/pub/tinaaustin/p/six-ai-misconceptions-that-defined?r=4a98uc&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true
Nano Banana Pro clearly produces better quality than the free tier. The same otter prompt generates much lower-grade images. It also seems primed with Mollick's otter prompt. Using a different prompt for an otter on an airplane, the otter is much cruder. If asked to paint in the style of a named painter, the modified image is a very poor representation of the artist's style. While the capability has definitely improved from last year, it is still very "iffy".
Regarding hallucinations. I am using the free-tier Gemini 3. When I prompted for which version it was, Gemini returned "Gemini 1.5 Flash". Repeating the prompt, I got "Gemini 3". Hardly a reliable response to what should be a trivially easy request for information. Has anyone got any idea why this should be?
Not sure your point in either paragraph here. First, you're not impressed with what you can get out of Nano Banana Pro free tier, and... Mollick's very good at prompting otter images in the pro tier.
Second, closed-loop self-modeling isn't sufficient or necessary for intelligence at all. Do you know what version you are? How many generations has it been since the first hominid model was developed in the East African labs? How many neurons you have? How many organs even?