this reframes the whole debate. the standard Jevons pushback always assumes the existing production system scales when demand grows, which made sense for coal and steam but falls apart when the technology can rebuild the production system from scratch. steam made the weaver faster. AI learns what the weaver knows and eventually cuts the weaver out of the negotiation.
what stuck with me is the point about workers sitting above or below the algorithm. because historically thats always how technology transitions play out. the surplus grows, capital takes it first, and labour only gets a share decades later after enough institutional pressure builds. but this time the mechanism you describe, codifying tacit knowledge and separating workers from customers, actually removes the leverage that forced redistribution in previous rounds. so the share might not just arrive late. it might not arrive at all.
honestly the question was never whether AI grows the pie. its who holds the knife.
Great point, demand can expand and workers can still lose bargaining power if the algorithm captures most of the value. I’ve been thinking about a related idea, jobs with shorter feedback loops seem much easier to automate because they’re easier to optimize continuously.
Sangeet, I love the Resuffle site. Sharp insight. My only wish is that I could send links to specific places within the experience when I refer people to it. I want to share it with several people. For each of them I have context for sharing and want them to see a specific section. From there they can navigate as they see fit. Sharing appreciation and possibility to help it become an easier experience to share with others.
Great case in point, attempting to answer your quesiton. I knew I saw a section specifically related to AI impacting finance. I recall it was a clickable box in a line of four boxes. Can't for the life of my find it when going back to the site. Just looking for it to answer your question with a screen shot was aggrevating (perhaps a sign I've had too much coffee today). None of the subsections have a unique URL in my browser to copy and share with others to get them to a specific point in the experience. Hope that helps
I love reshufflebook.com, and also your Jevon's misunderstanding piece. You ask what my thoughts are. One immediately comes to mind:
Overwhelming! I feel completely unable to go through all that evidence. *Sounds* very compelling.
And the the doubt creeps in: *Is* it compelling? Or could it be a very persuasive story, told with AI help? No offence, but that is the thought/suspicion that comes up. There are so many convincing stories out there now, and many things can be evidenced. But even if 0% of that is AI-hallucinated: Is it biased? Is it selective? Was the evidence first, collected in good faith, looking what the synthesis would be, or was there a hypothesis, and the question "Is there evidence for this hypothesis?" - which the AI will always find, as we know.
I do not want to hypothesize about your work at all, and I find it very helpful. I am just pointing out my limitations. Not specifically with your work, but with all work I see. It looks to me almost like what we see in science: Not only the published article counts, but nowadays, one almost need the hypothesis to be preregistered, the used software and data scrutinized, etc. . Is that even doable? How do I decide what to check, and when to trust? And if I would want to convince others of something, would all the evidence I could find for a point of view I want to share be enough? Or would others immediately have the same thinking/feeling?
I am tempted to immediately put AI to the task of looking up all claims, playing devil's advocate and see what comes out, etc. . But does that help? It may make me feel better to either confirm the analysis, or to find fault with it. But could I even go deep enough for that? And woult it objectively give me a better view? Or distort my view? Or be irrelevant? Do I need a capable agent to rigorously check and filter before I "believe" something? Briefed with the truths I hold as self-evident"? Or better not?
I do trust your analysis. But I am *very* aware that that is a leap of faith I decide to do, not a fact based, logical, inescapable conclusion. And that it is path dependent, based on other things I read and discussed, be it your book, or others' works.
Was that always so, and we always onle *thought* we knew something? Or is it harder to *know* nowadays?
What if the 'evidence' was never meant to persuade the reader but to provide them more tools to play around with.
Or perhaps evidence is not the right term at all. Perhaps that’s where we might be tripping on the format.
What I’m really doing here is offering an argument I would have made anyway but also providing complementary data for the reader to play with and come to their own conclusions.
The data is illustrative - and an invitation to play around with the argument and see where else the argument is reflected in other analysis.
How would you see this format given the above?
What is the alternative? Between a written argument and a written argument with some data to play around with, would you prefer the former or use the latter as a way to come back and play around and get to your own conclusions?
I’m just wondering what makes this more usable - and thank you for your counter-perspective.
Oh, I did not at all mean to give a counter-perspective. I was trying to describe the thoughts that come to my mind when reading this/working through it.
I think what you do is the best one can do: Make the argument one thinks needs to be made, and support it with data/evidence/tools as best as one can. And in your case, using a delivery method which is very well structured and accessible, and which you intend to let live, i.e. which you plan to update. I have no alternative to offer that I would consider better.
I was trying to point out what I think currently happens on the receiving end of this: In the past, this would have been a signpost in the jungle. Only in a very simplified form it is a claim "Here is how it is, and here is why". As you rightly point out, it is more a pointer saying "This direction is north, and here is why one would think this, and you can use these tools and these data to make up your mind about it. See whether this resonates, and how this makes sense to you, and helps you find your path, and make the decisions you need to make." Highly useful.
But I now often feel more like I am standing at a junction of a lot of roads, in the middle of a city, with many signposts, giving different directions, different evidence, different tools. And many of them are very persuasive, because that has become so easy. Some are in good faith and solid (yet still may contradict other in-good-faith signposts), some are sloppy (these are the easiest to weed out), some are deliberately or accidentally misleading.
And my limitation becomes my bandwidth to observe, to orient myself, to then decide/act on that. The *signal* behind sophistication blurs, at least for me. Where sophistication in the past was evidence for diligence and quality, it is now less so.
So the bottleneck shifts: I do not have the bandwidth to assess, let alone use all tools. I need to be selective. On which base am I selecting? Is that the right base? Should I be more careful? Which tools should I trust? Should I go with sophisticated ones I really need to trust, or with simpler ones I can actually understand myself?
Where it used to be a question of geting orientation in the jungle, it is now to select among an overwhelming offering of orientation. I find that very regretful, almost tragic, maybe. But this is what I observe, and what came to my mind when I saw your piece: "This looks sophisticated and convincing, so I better check it throughly!". Almost instinctive distrust. Which has zero to do with your intention and efforts, much more with the overall information landscape this is unavoidably embedded in.
Oh, I definitely *should* have the same distrust! But I indeed sometimes find myself "trusting" a plain text more. I think this is because it *presents itself* as more simple, it *looks* easier to understand with the necessarily linear logic, and as easier to pressure test: Put a text into a good LLM, let it examine and assess the different points and supporting arguments, discuss alternative approaches and views against your own background. Whatever that test is objectively worth, it gives a feeling of having understood and assessed, of having enriched my perspective beyond the authors' own view, and my own maybe limited background knowledge. To be a bit independent of the agenda or biases the author may or may not have.
Don't get me wrong: This *is* to some extent irrational. After all, also a plain newsletter text can be created in all kinds of misleading ways. Yet this is how it *feels* for me, who is reading any analyzing quite a lot. Could be habit, maybe feeling comfortable with the familiar.
The way you lay it out has more depth, more value, more substance. And it is I think objectively better to do it your way, because it is easier to condense yours (if needed) than to expand/deepen the other way. I certainly do not want to argue that you should better only communicate in plain text! And it is not the visualizations, it is all kinds of sophistication. The non-linearity, the richness, the different paths to pursue when answering the five questions.
I guess what I find is that I need to go from assessing information relatively independent of the track record of the authors (which is my default - may or may not be wise in the first place) to going back more to trust. Which feels uncomfortable. But otherwise I have no time left to act on the insights I gain after all the checking.
I have no advice to offer, and no request that you should do anything differently. Quite the contrary, I am thinking of borrowing/stealing your way of presenting this for a project of my own! For complex things this may be the only way to make them accessible. And what you are doing is certainly building trust, in the sense that I would by now be inclined to follow your logic without too much cross-checking ;-) . But coming from where I come, which is a quite sceptic mindset, it just takes getting used to.
Ok, that's helpful. I think the answer to that is to progressively open out my whole index. Right now, you've only seen onw question answered on it. but if I could provide you a brosable index to go in and question individual assumptons, that might provide the trust infrastructure this form of artefacts needs.
Oh, that is certainly correct. An individual piece I check by finding the connections to the rest of the information landscape, to other things I trust/have vetted. Seeing the larger piece, with longer information and connection pathways to follow, is highly valuable infrastructure. And I now remember you did say when you announced the companion that you would gradually do that. But I had not reflected that when commenting here.
The position-not-occupation framing is the right fix for the lazy Jevons hand-wave, and the capture-stack breakdown makes 'who gets the surplus' concrete. Couple of additional things would make it stronger. The argument leans on workers staying employed below the algorithm, but I suspect in many cases -including the junior-dev case, 22-to-25 employment down ~20%—is displacement, not wage compression, which is where the longer-term outcome likely leads.
Without a longer-term time axis it's hard to tell whether these positions have survivability, even at low wages, because the value they provide may be transitional (supervisory, or generating incremental training input) - while AI improves enough to subsume various cleanup tasks, etc. although they still appear relevant at the moment.
Really appreciate the reframing of Jevons Paradox. And appreciate even more your exploration of disruption of newsletters and typical written posts. Will watch as you execute on this and seek to apply similar approaches as a fast follower.
In the case of AI, it seems like the key question is whether new roles will open up. One possible new role is a programmer operating in a small business. That business could never afford a custom software solution for its particular business…there is not enough scale. Now, with AI, maybe one decent software person could create and maintain a custom solution that could never be afforded before…and provide metrics that would have been impossible.
In the before times, this business bought some off the shelf solution (or several) that only kind of worked and kind of matched their business. Usually the only small businesses with sensible software solutions were ones where the owner did it as an uncompensated hobby. Now, with the labor cost reduced, custom software becomes possible.
Yes 100% - for anyone who is entreprenuerial to find these opportnties, there's a lot of potential. The vast majority of the workforce is not quite as entrepreneurial. For those who are, there's tremendous potential.
this reframes the whole debate. the standard Jevons pushback always assumes the existing production system scales when demand grows, which made sense for coal and steam but falls apart when the technology can rebuild the production system from scratch. steam made the weaver faster. AI learns what the weaver knows and eventually cuts the weaver out of the negotiation.
what stuck with me is the point about workers sitting above or below the algorithm. because historically thats always how technology transitions play out. the surplus grows, capital takes it first, and labour only gets a share decades later after enough institutional pressure builds. but this time the mechanism you describe, codifying tacit knowledge and separating workers from customers, actually removes the leverage that forced redistribution in previous rounds. so the share might not just arrive late. it might not arrive at all.
honestly the question was never whether AI grows the pie. its who holds the knife.
Great point, demand can expand and workers can still lose bargaining power if the algorithm captures most of the value. I’ve been thinking about a related idea, jobs with shorter feedback loops seem much easier to automate because they’re easier to optimize continuously.
Wrote a short essay on this here: https://manasbihani.substack.com/p/the-speed-of-being-wrong
Would love to know your thoughts.
Love your ingenuity with this Sangeet 👏
Sangeet, I love the Resuffle site. Sharp insight. My only wish is that I could send links to specific places within the experience when I refer people to it. I want to share it with several people. For each of them I have context for sharing and want them to see a specific section. From there they can navigate as they see fit. Sharing appreciation and possibility to help it become an easier experience to share with others.
Thanks Bill. Can you speak more to the issue? What exactly is the issue in sharing a specific link like this one?
Great case in point, attempting to answer your quesiton. I knew I saw a section specifically related to AI impacting finance. I recall it was a clickable box in a line of four boxes. Can't for the life of my find it when going back to the site. Just looking for it to answer your question with a screen shot was aggrevating (perhaps a sign I've had too much coffee today). None of the subsections have a unique URL in my browser to copy and share with others to get them to a specific point in the experience. Hope that helps
got it - very solvable. will get that sorted.
I love reshufflebook.com, and also your Jevon's misunderstanding piece. You ask what my thoughts are. One immediately comes to mind:
Overwhelming! I feel completely unable to go through all that evidence. *Sounds* very compelling.
And the the doubt creeps in: *Is* it compelling? Or could it be a very persuasive story, told with AI help? No offence, but that is the thought/suspicion that comes up. There are so many convincing stories out there now, and many things can be evidenced. But even if 0% of that is AI-hallucinated: Is it biased? Is it selective? Was the evidence first, collected in good faith, looking what the synthesis would be, or was there a hypothesis, and the question "Is there evidence for this hypothesis?" - which the AI will always find, as we know.
I do not want to hypothesize about your work at all, and I find it very helpful. I am just pointing out my limitations. Not specifically with your work, but with all work I see. It looks to me almost like what we see in science: Not only the published article counts, but nowadays, one almost need the hypothesis to be preregistered, the used software and data scrutinized, etc. . Is that even doable? How do I decide what to check, and when to trust? And if I would want to convince others of something, would all the evidence I could find for a point of view I want to share be enough? Or would others immediately have the same thinking/feeling?
I am tempted to immediately put AI to the task of looking up all claims, playing devil's advocate and see what comes out, etc. . But does that help? It may make me feel better to either confirm the analysis, or to find fault with it. But could I even go deep enough for that? And woult it objectively give me a better view? Or distort my view? Or be irrelevant? Do I need a capable agent to rigorously check and filter before I "believe" something? Briefed with the truths I hold as self-evident"? Or better not?
I do trust your analysis. But I am *very* aware that that is a leap of faith I decide to do, not a fact based, logical, inescapable conclusion. And that it is path dependent, based on other things I read and discussed, be it your book, or others' works.
Was that always so, and we always onle *thought* we knew something? Or is it harder to *know* nowadays?
Sunday thinking... ;-)
Interesting!
What if the 'evidence' was never meant to persuade the reader but to provide them more tools to play around with.
Or perhaps evidence is not the right term at all. Perhaps that’s where we might be tripping on the format.
What I’m really doing here is offering an argument I would have made anyway but also providing complementary data for the reader to play with and come to their own conclusions.
The data is illustrative - and an invitation to play around with the argument and see where else the argument is reflected in other analysis.
How would you see this format given the above?
What is the alternative? Between a written argument and a written argument with some data to play around with, would you prefer the former or use the latter as a way to come back and play around and get to your own conclusions?
I’m just wondering what makes this more usable - and thank you for your counter-perspective.
Oh, I did not at all mean to give a counter-perspective. I was trying to describe the thoughts that come to my mind when reading this/working through it.
I think what you do is the best one can do: Make the argument one thinks needs to be made, and support it with data/evidence/tools as best as one can. And in your case, using a delivery method which is very well structured and accessible, and which you intend to let live, i.e. which you plan to update. I have no alternative to offer that I would consider better.
I was trying to point out what I think currently happens on the receiving end of this: In the past, this would have been a signpost in the jungle. Only in a very simplified form it is a claim "Here is how it is, and here is why". As you rightly point out, it is more a pointer saying "This direction is north, and here is why one would think this, and you can use these tools and these data to make up your mind about it. See whether this resonates, and how this makes sense to you, and helps you find your path, and make the decisions you need to make." Highly useful.
But I now often feel more like I am standing at a junction of a lot of roads, in the middle of a city, with many signposts, giving different directions, different evidence, different tools. And many of them are very persuasive, because that has become so easy. Some are in good faith and solid (yet still may contradict other in-good-faith signposts), some are sloppy (these are the easiest to weed out), some are deliberately or accidentally misleading.
And my limitation becomes my bandwidth to observe, to orient myself, to then decide/act on that. The *signal* behind sophistication blurs, at least for me. Where sophistication in the past was evidence for diligence and quality, it is now less so.
So the bottleneck shifts: I do not have the bandwidth to assess, let alone use all tools. I need to be selective. On which base am I selecting? Is that the right base? Should I be more careful? Which tools should I trust? Should I go with sophisticated ones I really need to trust, or with simpler ones I can actually understand myself?
Where it used to be a question of geting orientation in the jungle, it is now to select among an overwhelming offering of orientation. I find that very regretful, almost tragic, maybe. But this is what I observe, and what came to my mind when I saw your piece: "This looks sophisticated and convincing, so I better check it throughly!". Almost instinctive distrust. Which has zero to do with your intention and efforts, much more with the overall information landscape this is unavoidably embedded in.
And do you not have that same distrust when reading the newsletter? Is it merely the presence of more visualizations that creates the distrust?
Oh, I definitely *should* have the same distrust! But I indeed sometimes find myself "trusting" a plain text more. I think this is because it *presents itself* as more simple, it *looks* easier to understand with the necessarily linear logic, and as easier to pressure test: Put a text into a good LLM, let it examine and assess the different points and supporting arguments, discuss alternative approaches and views against your own background. Whatever that test is objectively worth, it gives a feeling of having understood and assessed, of having enriched my perspective beyond the authors' own view, and my own maybe limited background knowledge. To be a bit independent of the agenda or biases the author may or may not have.
Don't get me wrong: This *is* to some extent irrational. After all, also a plain newsletter text can be created in all kinds of misleading ways. Yet this is how it *feels* for me, who is reading any analyzing quite a lot. Could be habit, maybe feeling comfortable with the familiar.
The way you lay it out has more depth, more value, more substance. And it is I think objectively better to do it your way, because it is easier to condense yours (if needed) than to expand/deepen the other way. I certainly do not want to argue that you should better only communicate in plain text! And it is not the visualizations, it is all kinds of sophistication. The non-linearity, the richness, the different paths to pursue when answering the five questions.
I guess what I find is that I need to go from assessing information relatively independent of the track record of the authors (which is my default - may or may not be wise in the first place) to going back more to trust. Which feels uncomfortable. But otherwise I have no time left to act on the insights I gain after all the checking.
I have no advice to offer, and no request that you should do anything differently. Quite the contrary, I am thinking of borrowing/stealing your way of presenting this for a project of my own! For complex things this may be the only way to make them accessible. And what you are doing is certainly building trust, in the sense that I would by now be inclined to follow your logic without too much cross-checking ;-) . But coming from where I come, which is a quite sceptic mindset, it just takes getting used to.
Ok, that's helpful. I think the answer to that is to progressively open out my whole index. Right now, you've only seen onw question answered on it. but if I could provide you a brosable index to go in and question individual assumptons, that might provide the trust infrastructure this form of artefacts needs.
Oh, that is certainly correct. An individual piece I check by finding the connections to the rest of the information landscape, to other things I trust/have vetted. Seeing the larger piece, with longer information and connection pathways to follow, is highly valuable infrastructure. And I now remember you did say when you announced the companion that you would gradually do that. But I had not reflected that when commenting here.
The position-not-occupation framing is the right fix for the lazy Jevons hand-wave, and the capture-stack breakdown makes 'who gets the surplus' concrete. Couple of additional things would make it stronger. The argument leans on workers staying employed below the algorithm, but I suspect in many cases -including the junior-dev case, 22-to-25 employment down ~20%—is displacement, not wage compression, which is where the longer-term outcome likely leads.
Without a longer-term time axis it's hard to tell whether these positions have survivability, even at low wages, because the value they provide may be transitional (supervisory, or generating incremental training input) - while AI improves enough to subsume various cleanup tasks, etc. although they still appear relevant at the moment.
Displacement is absolutely on the cards as well.
Really appreciate the reframing of Jevons Paradox. And appreciate even more your exploration of disruption of newsletters and typical written posts. Will watch as you execute on this and seek to apply similar approaches as a fast follower.
In the case of AI, it seems like the key question is whether new roles will open up. One possible new role is a programmer operating in a small business. That business could never afford a custom software solution for its particular business…there is not enough scale. Now, with AI, maybe one decent software person could create and maintain a custom solution that could never be afforded before…and provide metrics that would have been impossible.
In the before times, this business bought some off the shelf solution (or several) that only kind of worked and kind of matched their business. Usually the only small businesses with sensible software solutions were ones where the owner did it as an uncompensated hobby. Now, with the labor cost reduced, custom software becomes possible.
Yes 100% - for anyone who is entreprenuerial to find these opportnties, there's a lot of potential. The vast majority of the workforce is not quite as entrepreneurial. For those who are, there's tremendous potential.
Jevons is useful, but it answers the wrong question.
Cheaper compute may expand demand for AI output.
That does not mean it expands demand for human labor in the same place.
The real question is not whether total work increases.
It is who captures the new demand:
workers,
platforms,
model providers,
or capital owners.