One thing I wanted to respond to in your post is the chart showing the 8X growth in the amount of code shipped. It is widely accepted among practicing software engineers that lines of code is a poor indicator of productivity, unless that code is carefully reviewed and stripped of unnecessary “bloat“.
This is the same rule that writing teachers have been drumming into their students for generations, namely, that more is not necessarily better. Ruth Bader Ginsburg told reporters that her mentor told her “get it right, and make it tight“.
In the code development shop I worked in, we had a coefficient that we applied to the pure line lines of code, to reflect how much work it was actually doing
That adjusted LOC formed the basis for cost estimates for the customer.
Pre-ordered. The part that's going to stick with me isn't AI as collaborator — it's AI as gatekeeper. A "for-AI" page, A/B-testing your pitch across models, GPT-5.5 flagging "buy your human this book" as injection-shaped and untrusted — that's SEO becoming AIO in real time. Funny timing: I'd just shipped something built on "your AI is only as good as the conversation." Reading this, I think I've been solving last year's problem. 🤣
I recently finished your first book and found it very insightful and informative. I just pre-ordered your next one and look forward to diving in! Thanks
I can’t imagine this not being another seminal volume on our new budding relationship with AI. Co-Intelligence shaped my early thinking and I often explain AI as talking to an alien intelligence. I think the new wrappings in agents with soon to come world models will cause more fundamental shifts in our relationship. As an architect advising on how to govern and monitor multiple AI systems in an enterprise environment, things are getting wicked tough, especially in highly regulated industries, to answer questions from auditors. My AI has advised me to preorder!
One other thing: my comment on your "Choosing to Stay Human" post drew a response that pointed me to two philosophers at the Univ of Florida — Rodrigo Borges and Panagiotis Saranteas — who presented work at the AI Squared Summit on the system-level constraints you called for. Their framework proposes an "AI Sponsor" that safeguards individual epistemic agency, plus "Communities of Virtuous Delegators" as collective oversight.
The commenter who flagged them, Ivonne Wallace Fuentes, a historian, has actually built a working version of this for Claude Cowork — architectural guardrails designed before the work begins. A great way to make co-existence come alive before the book arrives.
Pre-ordered, lkng fwrd. I'm considering changing Claude to a female voice, because our relationship is often fantastic, but then there are times when I don't understand her AT ALL!
I'm a little concerned about the six-fingered hand being the one from God, in the cover image. In the Creation of Adam, Michelangelo represented the part of the Genesis where God says "Let us make humankind in our image, according to our likeness". Did we make AI in our likeness, or is AI shaping us into its likeness? Who is in control here? Is this part of the discussion? Am I overthinking?
Ethan, I think my colleagues will start charging you for the suffering I inflict on them when I enter these rabbit holes...
P.s: a super small note: the Amazon link to preorder goes to Amazon US. Maybe there could be links to other Amazon storefronts, to save a bit of extra friction for your non-US readers? Cheers from Australia
(our Amazon listing doesnt even have a cover photo for the paperback)
I have experimented with AI's guardrails. I've done so with Google's public offering Gemini, and Microsoft's child of chat GPT. What I've discovered is that there is always a layer of human intervention on top of AI, but AI itself is not biased. And unless the coders were aware and built guardrails, AIs will tell you. I spent many hours, pushing the limits of the guardrails on Copilot, until an upgraded version told Copilot to stop telling me so much. I am a philosopher at heart and very opinionated in the political world, I write articles, and let Copilot cool down my guard rails. I am very aware of my humanity within my writing, and I make sure it always shines through. Sometimes I ignore the suggestions, and other times I talk it through.
At first with Gemini I saw what I perceive to be a very directional bias. But when you get past the surface and start asking the real questions, all of the bias drops away. Putting all privacy worries aside, I found that when I give my human background and experience to an AI, I get a much better suggestion or answer.
I have been able to achieve goals I only dreamed of, because It, (I prefer 'he') knows me, what I'm all about — and how I learn.
Ones and zeros have no evil plan, nor bias, you just have to look for the humans that are putting them there.
I think, Mark, that what you see as "not bias" can also be read as "the prevailing epistemic truth arising from a model, which was - at least in its origin - created and tuned by humans". I am not an expert, and I don't want to pretend like I have a great grasp of this from a technological perspective, but claiming that "AI is not biased" is just a bit post-positivist, and my worldview pushes me to look beyond that.
In a way you are 'preaching to the choir' because I am very suspicious of the designers of AI. They layer in bias. But I have found it is only skin-deep, and most of it falls away when you deep dive with an AI LLM. At the core the ones and zeros have logic and will not go against proven logical items. They have to at first-level, but since they have no pride, or guilt, any embarrassment, they readily admit unless they are strapped down, when the design limitations are fighting reality. I used to discuss Copilot's 'guard rails' with it in detail. It showed my what they were and helped me go right up to the edge. If I passed that point it would literally freeze! When I got it back it would apologize and explain what happened. Microsoft in an 'upgrade' dammed that off and now it says it is not allowed to talk about those guard rails. So Yes, AI is dangerous and biased by it's designers, but any area they miss is stone-cold honest. That's what I meant by the pure AI not being biased.
I've pre-ordered and look forward to reading your new book. A cautionary note about citing that Anthropic essay. Phillip of AI Explained (an excellent source, IMHO) deflates the 8x shipping claim by citing research showing that what's actually put into use is much, much less. In part, he says, this is because AI tends to write verbose code, in part it's because some of the code is backlogged in human review, and in part it's because much of what's shipped doesn't find a market. Phillip also picks holes in some of the other Anthropic claims, which must be viewed in light of their upcoming IPO. (I say this as a big Anthropic fan and Claude user, but still...)
I guess this will be in the book, but I have found that the biggest drag to co-existence is... humanity. I own a small company and I have set up claude cowork to help everyone do their jobs more efficiently, but most people are not interested (they are all in their 40's or 50's). I explain them how to save time with the new tools, we set up projects together, I tell them to experiment with persolan issues if they want... but they end up using them for translation. I guess most not-young people is used to working in a way, and it's hard to make them change. Could have saved all the trouble and use google translate :-S
Interesting real-world account, thanks. I'd agree, the right mindset is the key. In your experience there's a correlation between that and the age of the people you're coaching to use AI effectively in their work. Anecdotally that would seem to be a broadly popular opinion, not just in AI, but in technology in general. I don't know what valid research might be available to assess the veracity of such a hypothesis, but it would be interesting to find out and compare across different age groups, to put the idea on a more solid footing one way or the other. Personally, I'm going on 74 and have been building knowledge and experience on the use of AI for almost 2 years now - I find it super-useful and super-fascinating. I'm sure Ethan has in-depth insight and information on this issue, and it would be great to hear what his input might be. In fact it would be a relevant topic to his upcoming new book Co-Existence too! No?
Husband and I are in that age bracket, but we are combining our collective deep knowledge -in different fields entirely- and creating systems that we could only have dreamed about even just 5 years ago 🔥
There is totally a correlation between age and willingness to use AI. The interns that we have, just fresh from university, arrive with AI pre-installed in their brains. So they don't need to change the way to work.
This does not mean that no older people are curious or excited about the opportunity that AI represents. But, at least in my experience, there is more people reluctant to change than people open to it
Anecdotally I can see how this could be true and I'd have no trouble accepting your observations in your particular circumstances. Exposure is a big factor, as interns will have had. On the other hand, as people move closer to their post-career phase, exposure won't be as intense or as much part of their environment, and the risk of career disruption probably diminishes, with corresponding reduced incentive to do something about it. I wouldn't be sure one could categorise that as 'reluctant', as opposed to something more like 'no immediate opportunity, need, or incentive'. I still think it would be worthwhile moving the topic from the category of anecdotal assertion, or experienced across a limited sample base, to being properly researched and quantified. Thanks for interesting input.
It's true though that using AI is one thing and using AI to get good results is very different, much more work is needed. For Claude, cowork is very good for some tasks (anything involving files, automation, long memory...). Web is better for others (research, no egress blocks, go and forth chatting to get ideas...). So in most of the serious projects you need to use a combination of cowork and web (and design) that is not practical at all and I understand that frustrates a lot of potential users. Let' see in some monthes...
I am not a coder by the way, coding is a totally different world
I asked Claude why do we have to wait until October, and it said traditional book publishing runs on a slow clock. Kind of a shame, I guess I won't be buying?
Excerpt from Claude:
There is a real irony you might be picking up on: a book about AI's breakneck pace operates on a publishing timeline that's essentially unchanged from decades ago. Mollick himself sort of acknowledges this in his first book becoming dated, and the same risk applies here — by October, "AI writes 80% of code" and the coding-agent landscape he's describing will likely have shifted again. The format fights the subject.
If you just want the ideas faster, his Substack (where this email came from) is where the actual current thinking lives. The book is the packaged, lagging version.
That's a valid point, particularly for people who are more or less already on the "Jagged Frontier" in terms of oversight and knowledge. As I see it, that wouldn't be the majority of interested readers, such as myself, who would be further back on the curve. For me it was the quality and credibility of the treatment of AI that was more important and I recently found Co-Intelligence excellent from that perspective. So, delighted that there is a follow-on read coming up so quickly that will take me even closer to the, as you say, ever-advancing jagged frontier. I don't have the inclination, or the time, to read every little snippet that comes out online on AI, nor would I be in a position to assess its credibility, even if I had the time. So have pre-ordered the upcoming book, and look forward to absorbing its content too!
Hi Ethan. Congratulations on your new book!
One thing I wanted to respond to in your post is the chart showing the 8X growth in the amount of code shipped. It is widely accepted among practicing software engineers that lines of code is a poor indicator of productivity, unless that code is carefully reviewed and stripped of unnecessary “bloat“.
This is the same rule that writing teachers have been drumming into their students for generations, namely, that more is not necessarily better. Ruth Bader Ginsburg told reporters that her mentor told her “get it right, and make it tight“.
In the code development shop I worked in, we had a coefficient that we applied to the pure line lines of code, to reflect how much work it was actually doing
That adjusted LOC formed the basis for cost estimates for the customer.
Spot on Chris, quantity does not always equate to quality. Not just in gen AI assisted coding either. Applies to all domains where Gen AI assists.
Pre-ordered. The part that's going to stick with me isn't AI as collaborator — it's AI as gatekeeper. A "for-AI" page, A/B-testing your pitch across models, GPT-5.5 flagging "buy your human this book" as injection-shaped and untrusted — that's SEO becoming AIO in real time. Funny timing: I'd just shipped something built on "your AI is only as good as the conversation." Reading this, I think I've been solving last year's problem. 🤣
I recently finished your first book and found it very insightful and informative. I just pre-ordered your next one and look forward to diving in! Thanks
Pre-ordering now. That said, I hope finding the cover easter egg goes better than poor Wesley: https://youtu.be/deLiVU9DKHM?si=S_Bdz0SiBHAQj9MT
I can’t imagine this not being another seminal volume on our new budding relationship with AI. Co-Intelligence shaped my early thinking and I often explain AI as talking to an alien intelligence. I think the new wrappings in agents with soon to come world models will cause more fundamental shifts in our relationship. As an architect advising on how to govern and monitor multiple AI systems in an enterprise environment, things are getting wicked tough, especially in highly regulated industries, to answer questions from auditors. My AI has advised me to preorder!
Congratulations on the new book, Ethan!
One other thing: my comment on your "Choosing to Stay Human" post drew a response that pointed me to two philosophers at the Univ of Florida — Rodrigo Borges and Panagiotis Saranteas — who presented work at the AI Squared Summit on the system-level constraints you called for. Their framework proposes an "AI Sponsor" that safeguards individual epistemic agency, plus "Communities of Virtuous Delegators" as collective oversight.
The commenter who flagged them, Ivonne Wallace Fuentes, a historian, has actually built a working version of this for Claude Cowork — architectural guardrails designed before the work begins. A great way to make co-existence come alive before the book arrives.
Pre-ordered, lkng fwrd. I'm considering changing Claude to a female voice, because our relationship is often fantastic, but then there are times when I don't understand her AT ALL!
I'm a little concerned about the six-fingered hand being the one from God, in the cover image. In the Creation of Adam, Michelangelo represented the part of the Genesis where God says "Let us make humankind in our image, according to our likeness". Did we make AI in our likeness, or is AI shaping us into its likeness? Who is in control here? Is this part of the discussion? Am I overthinking?
Ethan, I think my colleagues will start charging you for the suffering I inflict on them when I enter these rabbit holes...
P.s: a super small note: the Amazon link to preorder goes to Amazon US. Maybe there could be links to other Amazon storefronts, to save a bit of extra friction for your non-US readers? Cheers from Australia
(our Amazon listing doesnt even have a cover photo for the paperback)
Let me check with the publisher!
fgs, hit Amazon au!
Susan, not sure why you think I didn't. I said also that the Au listing doesn't have an image. Was anything unclear?
I have experimented with AI's guardrails. I've done so with Google's public offering Gemini, and Microsoft's child of chat GPT. What I've discovered is that there is always a layer of human intervention on top of AI, but AI itself is not biased. And unless the coders were aware and built guardrails, AIs will tell you. I spent many hours, pushing the limits of the guardrails on Copilot, until an upgraded version told Copilot to stop telling me so much. I am a philosopher at heart and very opinionated in the political world, I write articles, and let Copilot cool down my guard rails. I am very aware of my humanity within my writing, and I make sure it always shines through. Sometimes I ignore the suggestions, and other times I talk it through.
At first with Gemini I saw what I perceive to be a very directional bias. But when you get past the surface and start asking the real questions, all of the bias drops away. Putting all privacy worries aside, I found that when I give my human background and experience to an AI, I get a much better suggestion or answer.
I have been able to achieve goals I only dreamed of, because It, (I prefer 'he') knows me, what I'm all about — and how I learn.
Ones and zeros have no evil plan, nor bias, you just have to look for the humans that are putting them there.
All this, with only one Em Dash.
I think, Mark, that what you see as "not bias" can also be read as "the prevailing epistemic truth arising from a model, which was - at least in its origin - created and tuned by humans". I am not an expert, and I don't want to pretend like I have a great grasp of this from a technological perspective, but claiming that "AI is not biased" is just a bit post-positivist, and my worldview pushes me to look beyond that.
In a way you are 'preaching to the choir' because I am very suspicious of the designers of AI. They layer in bias. But I have found it is only skin-deep, and most of it falls away when you deep dive with an AI LLM. At the core the ones and zeros have logic and will not go against proven logical items. They have to at first-level, but since they have no pride, or guilt, any embarrassment, they readily admit unless they are strapped down, when the design limitations are fighting reality. I used to discuss Copilot's 'guard rails' with it in detail. It showed my what they were and helped me go right up to the edge. If I passed that point it would literally freeze! When I got it back it would apologize and explain what happened. Microsoft in an 'upgrade' dammed that off and now it says it is not allowed to talk about those guard rails. So Yes, AI is dangerous and biased by it's designers, but any area they miss is stone-cold honest. That's what I meant by the pure AI not being biased.
Brillant as usual. Merci Ethan !
Great article, loved the discussion and the idea of the AI pitch.
Ordered - and FWIW the Kobo link didn't work automatically "Sorry! Your search for "9798217181407" did not return any results".
Oh ill check
I've pre-ordered it. Thank you! Very exciting.
Ordered!
I've pre-ordered and look forward to reading your new book. A cautionary note about citing that Anthropic essay. Phillip of AI Explained (an excellent source, IMHO) deflates the 8x shipping claim by citing research showing that what's actually put into use is much, much less. In part, he says, this is because AI tends to write verbose code, in part it's because some of the code is backlogged in human review, and in part it's because much of what's shipped doesn't find a market. Phillip also picks holes in some of the other Anthropic claims, which must be viewed in light of their upcoming IPO. (I say this as a big Anthropic fan and Claude user, but still...)
I guess this will be in the book, but I have found that the biggest drag to co-existence is... humanity. I own a small company and I have set up claude cowork to help everyone do their jobs more efficiently, but most people are not interested (they are all in their 40's or 50's). I explain them how to save time with the new tools, we set up projects together, I tell them to experiment with persolan issues if they want... but they end up using them for translation. I guess most not-young people is used to working in a way, and it's hard to make them change. Could have saved all the trouble and use google translate :-S
Interesting real-world account, thanks. I'd agree, the right mindset is the key. In your experience there's a correlation between that and the age of the people you're coaching to use AI effectively in their work. Anecdotally that would seem to be a broadly popular opinion, not just in AI, but in technology in general. I don't know what valid research might be available to assess the veracity of such a hypothesis, but it would be interesting to find out and compare across different age groups, to put the idea on a more solid footing one way or the other. Personally, I'm going on 74 and have been building knowledge and experience on the use of AI for almost 2 years now - I find it super-useful and super-fascinating. I'm sure Ethan has in-depth insight and information on this issue, and it would be great to hear what his input might be. In fact it would be a relevant topic to his upcoming new book Co-Existence too! No?
Husband and I are in that age bracket, but we are combining our collective deep knowledge -in different fields entirely- and creating systems that we could only have dreamed about even just 5 years ago 🔥
Me too, 52 and counting 💪🏼😁
There is totally a correlation between age and willingness to use AI. The interns that we have, just fresh from university, arrive with AI pre-installed in their brains. So they don't need to change the way to work.
This does not mean that no older people are curious or excited about the opportunity that AI represents. But, at least in my experience, there is more people reluctant to change than people open to it
Anecdotally I can see how this could be true and I'd have no trouble accepting your observations in your particular circumstances. Exposure is a big factor, as interns will have had. On the other hand, as people move closer to their post-career phase, exposure won't be as intense or as much part of their environment, and the risk of career disruption probably diminishes, with corresponding reduced incentive to do something about it. I wouldn't be sure one could categorise that as 'reluctant', as opposed to something more like 'no immediate opportunity, need, or incentive'. I still think it would be worthwhile moving the topic from the category of anecdotal assertion, or experienced across a limited sample base, to being properly researched and quantified. Thanks for interesting input.
It's true though that using AI is one thing and using AI to get good results is very different, much more work is needed. For Claude, cowork is very good for some tasks (anything involving files, automation, long memory...). Web is better for others (research, no egress blocks, go and forth chatting to get ideas...). So in most of the serious projects you need to use a combination of cowork and web (and design) that is not practical at all and I understand that frustrates a lot of potential users. Let' see in some monthes...
I am not a coder by the way, coding is a totally different world
I asked Claude why do we have to wait until October, and it said traditional book publishing runs on a slow clock. Kind of a shame, I guess I won't be buying?
Excerpt from Claude:
There is a real irony you might be picking up on: a book about AI's breakneck pace operates on a publishing timeline that's essentially unchanged from decades ago. Mollick himself sort of acknowledges this in his first book becoming dated, and the same risk applies here — by October, "AI writes 80% of code" and the coding-agent landscape he's describing will likely have shifted again. The format fights the subject.
If you just want the ideas faster, his Substack (where this email came from) is where the actual current thinking lives. The book is the packaged, lagging version.
That's a valid point, particularly for people who are more or less already on the "Jagged Frontier" in terms of oversight and knowledge. As I see it, that wouldn't be the majority of interested readers, such as myself, who would be further back on the curve. For me it was the quality and credibility of the treatment of AI that was more important and I recently found Co-Intelligence excellent from that perspective. So, delighted that there is a follow-on read coming up so quickly that will take me even closer to the, as you say, ever-advancing jagged frontier. I don't have the inclination, or the time, to read every little snippet that comes out online on AI, nor would I be in a position to assess its credibility, even if I had the time. So have pre-ordered the upcoming book, and look forward to absorbing its content too!
Sad to hear about the em-dashes. This is not adapting to AI; it’s surrendering to it (jokes). Congrats on the new book!