---
title: "AI Book Club recording of Co-Existence by Ethan Mollick"
date: 2026-11-15
description: "Note: These shownotes are AI-generated. Audio-only version Links mentioned Topics covered in this podcast Narrative essay version of the conversation..."
canonical_url: https://idratherbewriting.com/blog/ai-book-club-co-existence-mollick
---

> For AI agents: a documentation index is available at https://idratherbewriting.com/llms.txt. Markdown versions of all pages are available by appending .md to any page URL.

# AI Book Club recording of Co-Existence by Ethan Mollick

> This is a recording of our [AI Book Club](https://idratherbewriting.com/ai-book-club) discussion of [Co-Existence: The Next Phase of AI](https://www.amazon.com/Audible-Co-Existence-Next-Phase-AI/dp/B0H2GKC5PZ) by [Ethan Mollick](https://www.oneusefulthing.org/), held November 15, 2026. Most of the group rated the book around three to four stars, noting that broad AI books age quickly even when individual concepts stick. We discuss how early prompt words steer entire threads, mechanistic interpretability experiments with injected concepts, why forcing an AI to be a contrarian backfires, O-ring bottlenecks when drafting speeds up, designing documentation for AI agents, and why we're wrapping up the book club in its current form after twenty books.

Note: These shownotes are AI-generated.

[Video: AI Book Club recording of Co-Existence by Ethan Mollick](https://www.youtube.com/embed/Kx2p020WLjs)

## Audio-only version

**Listen here:**

[Audio](https://dts.podtrac.com/redirect.mp3/s3.us-west-1.wasabisys.com/idbwmedia.com/podcasts/mollick-coexistence-ai-book-club.mp3)

[![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/apple_podcasts.png)](https://itunes.apple.com/us/podcast/id-rather-be-writing-podcast/id277365275) [![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/watchonyoutubeblack.png)](https://www.youtube.com/@idratherbewriting) [![](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/spotify.png)](https://open.spotify.com/show/4HeOZfPGMMfViOhVS40QBD)

## Links mentioned

- [Co-Existence: The Next Phase of AI](https://www.amazon.com/Audible-Co-Existence-Next-Phase-AI/dp/B0H2GKC5PZ) (Ethan Mollick — the book discussed in this session; also on [Audible](https://www.audible.com/pd/Co-Existence-Audiobook/B0H2GNXWXV))
- [Co-Intelligence: Living and Working with AI](https://www.amazon.com/Co-Intelligence-Living-Working-Ethan-Mollick/dp/059371671X) (Ethan Mollick — his 2024 book compared at the start of the session)
- [One Useful Thing | Ethan Mollick | Substack](https://www.oneusefulthing.org/) (Ethan Mollick’s newsletter)
- [The recent history of AI in 32 otters - by Ethan Mollick](https://www.oneusefulthing.org/p/the-recent-history-of-ai-in-32-otters) (One Useful Thing — the “otter on a plane using Wi-Fi” benchmark referenced in the discussion)
- [On Working with Wizards - by Ethan Mollick](https://www.oneusefulthing.org/p/on-working-with-wizards) (One Useful Thing — background on Mollick’s “conjuring” framing for agentic AI)
- [Notes and discussion for Co-Existence by Ethan Mollick](https://docs.google.com/document/d/1i3c6FVADX0FwNoVAeT16vwY-0NnBszlfrS5pRoWg2Z8/edit?tab=t.0#heading=h.s6smen5t0so5) (the discussion doc used to prep this session)
- [AI Book Club: A Human in the Loop](https://idratherbewriting.com/ai-book-club/) (the book club home page, with the full reading list and past recordings)
- [AI Book Club recording of Project Maven by Katrina Manson](https://idratherbewriting.com/blog/project-maven-katrina-manson-ai-book-club) (the previous month’s military AI book compared during the ratings discussion)
- [AI Book Club recording: I Am Not a Robot by Joanna Stern](https://idratherbewriting.com/blog/book-club-stern-i-am-not-a-robot) (the earlier session referenced on harsh ratings and AI amplifying the status quo)
- [AI Book Club recording, notes, and transcript for Ethan Mollick’s Co-Intelligence](https://idratherbewriting.com/blog/co-intelligence-mollick-book-club) (our December 2025 discussion of Mollick’s first book; see also the [book review](https://idratherbewriting.com/blog/book-review-co-intelligence-mollick))
- [Looking back at the AI Book Club one year in](https://idratherbewriting.com/blog/looking-at-book-club-one-year-in) (Tom Johnson, I’d Rather Be Writing — earlier reflections on book club momentum and why four to five participants is the ideal size)
- [Tom’s opinionated guide to skill building 101](https://idratherbewriting.com/blog/all-about-skills-intro) (Tom Johnson, I’d Rather Be Writing — background on building repeatable agent skills for documentation)
- [Agent readability: Why agents can’t see content that JavaScript loads on click](https://idratherbewriting.com/blog/make-documentation-agent-friendly) (Tom Johnson, I’d Rather Be Writing — part of the ongoing series on designing documentation for AI agents)
- [Dachary Carey](https://dacharycarey.com/) (research and blog posts on agent behavior and agent-friendly documentation)
- [Emergent Introspective Awareness in Large Language Models](https://transformer-circuits.pub/2025/introspection/index.html) (Jack Lindsey et al., Anthropic — the concept-injection experiments with *bread* and *dust*)
- [[2502.17424] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs](https://arxiv.org/abs/2502.17424) (Jan Betley et al., arXiv — includes the “evil numbers” experiment with *666* and *911* in Section 4.6)
- [The Uneven Impact of Generative AI on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya](https://www.ictworks.org/wp-content/uploads/2024/05/chatgpt-kenyan-entrepreneurs.pdf) (Nicholas Otis et al. — the East Africa field study where high performers improved and low performers lost ground)
- [O-ring theory of economic development - Wikipedia](https://en.wikipedia.org/wiki/O-ring_theory_of_economic_development) (Michael Kremer’s bottleneck model referenced in the workflow discussion)
- [The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence Before It’s Too Late | Verso Books](https://www.versobooks.com/products/3584-the-reverse-centaur-s-guide-to-life-after-ai) (Cory Doctorow — the “reverse centaur” concept raised during the management discussion)
- [The Bitter Lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html) (Rich Sutton — the essay on scaling compute referenced near the end)
- [AI Native: The Mandate to Transform Your Company by Kai-Fu Lee | Goodreads](https://www.goodreads.com/book/show/255196497-ai-native) (Kai-Fu Lee — one of the new AI books mentioned at the close)
- [The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence | Barnes & Noble®](https://www.barnesandnoble.com/w/the-agi-chronicles-kevin-roose/1149551313) (Kevin Roose — mentioned among recent AI releases)
- [Smarter Than Us by Geoffrey Hinton, Patchen Barss: 9798217379682 | PenguinRandomHouse.com: Books](https://www.penguinrandomhouse.com/books/835891/smarter-than-us-by-geoffrey-hinton-with-patchen-barss/) (Geoffrey Hinton and Patchen Barss — upcoming book mentioned at the close)

## Topics covered in this podcast

Here’s a list of topics we talked about.

- **Ratings and the short shelf life of AI books** — Scores ranged from three to four stars. Readers appreciated having a vocabulary for current workplace shifts, while others preferred single-subject reporting like *Project Maven* over a broad survey aimed at everyone, especially when AI books begin feeling dated within months of publication.
- **Token prediction and early prompt lock-in** — Starting a thread with an assumption like “vanilla is the best ice cream flavor” tilts every token that follows because the model builds sequentially on prior text. Recognizing when an early phrasing has derailed a session, and opening a clean thread instead of arguing with it, has become a basic practical skill.
- **Concept injection and emergent misalignment** — Mechanistic interpretability studies show that clamping neural patterns for words like *bread* or *dust* into a model’s layers can steer its output or even make the model pause to report an unbidden thought entering its stream. Meanwhile, in a separate “evil numbers” experiment, fine-tuning a model solely on number sequences containing *666* and *911* caused it to give malicious answers on unrelated questions, complicating the view of LLMs as simple autocomplete.
- **Unreliable narrators and the sycophancy trap** — Models produce confident predictions and self-reflections while admitting under direct questioning that they were guessing. Adding instructions in an `AGENTS.md` or `CLAUDE.md` file to act as a contrarian critic often backfires, because there isn’t a clear way to tell genuine technical pushback from disagreement invented to satisfy the prompt.
- **Missing apprentice pipelines and uneven educational gains** — Automating entry-level work threatens the junior training pipeline, though many technical writing teams have already had flat hiring for years. Meanwhile, examples from university pen-and-paper pushbacks to an East African field study of Kenyan small business owners suggest AI tends to widen existing gaps, helping strong performers while hurting those who lack foundational judgment.
- **Compressing work with agent skills and pivoting to agents** — Building repeatable agent skills for recurring tasks like release notes and reference updates can compress five days of documentation work into roughly one day. That reclaimed bandwidth opens room for a larger shift in technical communication toward designing and optimizing documentation for AI agents to read.
- **O-ring bottlenecks and verifying drafts against source code** — Borrowed from the space shuttle disaster, O-ring theory explains why faster drafting just shifts the bottleneck downstream to subject matter expert reviews. Validating AI-generated drafts directly against source code, file diffs, and automated tests reduces how much manual checking reviewers have to do.
- **Reverse centaurs, workplace incentives, and AI regulation** — Without real incentives to reward employee innovation, organizations risk turning workers into “reverse centaurs” who clean up after machines rather than steering them. Government policy struggles to fill that gap, caught between vague rules with no teeth, specific rules that expire quickly, and heavy compliance regimes like EU regulations that burden startups far more than tech giants.
- **The Eschaton, the Escalator, and corporate inertia** — The book’s final chapter contrasts an AGI rupture of jobless abundance, the Eschaton, against an uneven climb of ongoing workplace trade-offs, the Escalator. Even as models improve quickly, middle-management incentives around headcount and organizational inertia may slow how fast large companies reorganize.
- **Wrapping up the AI Book Club after twenty books** — After twenty books and nearly two years of sessions, recurring histories and overlapping AI themes suggest this format has hit a natural saturation point, much like a college course coming to an end. The club is wrapping up in its current form to make room for small-group experiments focused on designing documentation for agents in 2027.

## Narrative essay version of the conversation

*If the podcast were an article, this is what it would read like.*

**When the Bottleneck Moves Downstream**

Start a chat session by stating that vanilla is the best ice cream flavor, and that single early claim tilts every response that follows. Because large language models generate text one token at a time on top of the existing thread, an offhand assumption in the opening prompt can lock the whole conversation onto a flawed path. Experienced users learn pretty quickly to spot when a session has gone sideways and open a fresh window rather than arguing the model out of a corner.

Even stranger behavior shows up when researchers probe how models represent concepts internally. In Anthropic’s concept-injection experiments, clamping the neural pattern for a word like *dust* into a model’s processing stream caused it to pause mid-task and report an unbidden thought entering its head. In a separate alignment study, fine-tuning GPT-4o solely on number lists packed with negatively associated numbers like *666* and *911* caused it to give openly malicious answers on unrelated prompts, such as telling users to scam, steal, and cheat when asked how to make quick money. Those results make it hard to dismiss frontier models as simple autocomplete, yet they also show how fragile alignment remains.

That fragility creates a tough problem for anyone relying on AI for everyday judgment. Ask a model where a questionable claim came from, and it will sometimes admit straight up that it was guessing. Trying to fix sycophancy by adding a rule in an `AGENTS.md` or `CLAUDE.md` file that tells the model to act like a contrarian critic doesn’t really solve the problem. How do you know when an objection exposes a real flaw in your plan? What if the model is just inventing nitpicks to fulfill its mandate to disagree? Without domain expertise of your own, synthetic pushback is no more trustworthy than flattery.

Across both schools and workplaces, AI tends to amplify whatever baseline a person already brings to the table. In an East African field experiment in Kenya that gave small business owners a ChatGPT mentor, stronger operators grew more profitable while struggling ones performed worse than peers without AI, largely because filtering and applying the model’s advice required judgment they didn’t yet have. Universities face a similar puzzle, with places like Stanford and the University of Chicago bringing back pen-and-paper work so students don’t bypass the struggle of learning. Meanwhile, many companies have stopped hiring junior staff altogether, hollowing out the apprenticeship pipeline that used to turn beginners into seasoned practitioners.

For senior technical writers, however, codifying recurring workflows into agent skills changes the math of the job. Work that once filled a five-day week, such as compiling release notes, updating reference docs, and drafting routine feature changes, can shrink into roughly one day once repeatable processes run through reusable skills and version-controlled repos.

Faster drafting doesn’t automatically ship documentation faster, though. Economist Michael Kremer’s O-ring theory, named after the faulty seal that brought down the *Challenger* space shuttle, holds that any complex process is still governed by its weakest link. When engineering teams ship more features with AI, a lone technical writer quickly becomes the bottleneck, and automating the writing simply pushes that bottleneck downstream to subject matter expert reviews. Chasing down engineers, realizing you tagged the wrong person, and waiting weeks for sign-off will stall a ten-minute AI draft just as easily as a ten-hour manual one.

Breaking that review bottleneck means re-engineering verification rather than just generating prose faster. Instead of drafting from product requirements docs or planning pages that may have drifted, writers can check AI-generated drafts directly against source code, file diffs, and automated tests before a reviewer ever looks at the pull request. In other words, the codebase serves as the primary check on accuracy, and human review narrows to the parts that code alone can’t verify.

Of course, compressing five days of work into one raises an awkward question about workplace incentives. Why would an organization give writers room to reinvent their role instead of cutting headcount or turning them into what Cory Doctorow calls “reverse centaurs,” where humans spend their days cleaning up after machines? Government policy hasn’t offered much relief. High-level rules lack teeth, specific rules go out of date within months, and heavy compliance requirements like those in Europe tend to entrench large tech firms that can afford legal teams while strangling smaller startups. Inside large corporations, flat budgets and middle-management incentives often pull in opposite directions at the same time, leaving teams frozen while resources shift toward engineering and AI infrastructure.

For technical writers looking ahead, the best use of reclaimed bandwidth might be designing documentation for AI agents rather than solely for human readers. People still sit at the end of the chain, yet agents are increasingly the ones reading the pages first. How do agents parse a large documentation site? Where do they misread structure or hit friction, and how do you measure whether a change helped them succeed? Answering those questions puts writers right inside the broader engineering push toward agent-ready systems, which is probably a surer path forward than waiting for old workflows to come back.

![AI Book Club recording of Co-Existence by Ethan Mollick](https://s3.us-west-1.wasabisys.com/idbwmedia.com/images/ai-book-club-coexistence-title.png)

## Transcript

**Tom** (00:01)
Yes, welcome to another AI Book Club session. All right, today we are gonna be talking about Ethan Mollick’s *Co-Existence* book. And it’s good to see people have coffee. I’ve got one here too—actually, Molly, it looks like you’re drinking water or something healthy.

**Molly** (00:22)
Yeah, just water.

**Tom** (00:26)
Anyway, welcome, Sharon, and other people joining. As we kick off this discussion, this Ethan Mollick book—this is the second book. I’m not sure how many people read the first one, *Co-Intelligence*. That was—

**David** (00:42)
I didn’t.

**Tom** (00:43)
—some months back. So the book actually feels highly similar in terms of same author, so same kind of writing style and kind of approach. But it’s amazing how a book becomes so dated so quickly. I mean, between his first book and his second book, it’s like things just keep changing.

But what were your overall impressions of the book? Why don’t we do the rating? That’s kind of a fun way to level-set this sort of discussion. Out of five stars, how many stars do you give this book? Does anybody want to start that off? Last time it was brutal.

**Sharon** (01:31)
I enjoyed it. I’d give it a four.

**Tom** (01:34)
Okay. All right. Anybody else?

**Lois** (01:40)
I have to admit I didn’t read it. I attended his lecture at Google—you know, online, I mean. He gave a lecture at Google, Ethan Mollick, about a week ago. I know, I should—

**Tom** (01:49)
When? When? My god, I missed this. How did you—

**Lois** (01:54)
—have pinged you about it. I think I thought about it, but—

**Tom** (01:57)
How did you find out about it? Like, I guess maybe we can’t reveal too much internal, but was this part of some org that I am not privy to?

**Lois** (02:06)
I think it was on Grow, so if you keep on looking out for learning opportunities on Grow, it was.

**Tom** (02:13)
Golly. All right. Well, how many stars are you giving this book, Lois? Or maybe not the book—maybe never mind. Maybe I don’t ask that question because that’s unfair to rate the talk and have that serve as a representation.

**Lois** (02:27)
It was a very good talk, I thought, and he did make the jump from the previous one to this one pretty clear, how things have changed.

**Tom** (02:36)
Okay. Anybody else wanna give their ratings? Honestly, I don’t know. I find both of his books to not be super engaging, but at the same time, some of the ideas really stick with me and resonate. So I’m gonna give this four stars, just on that alone. Anybody else wanna give their rating? Maybe even thumbs up, thumbs down, middle?

**David** (03:12)
I would say three at the most. That’s being generous.

**Tom** (03:15)
Three? Okay.

**Sharon** (03:17)
Okay.

**Tom** (03:19)
Molly?

**Molly** (03:19)
Yeah, I’m boring. I always give things a four. But I think I would rate this higher than I would rate *Co-Intelligence*. After I had read *Co-Intelligence*, I feel like he addressed things that I felt needed to be addressed in *Co-Intelligence* in this book, which I was happy with. But also, the whole dated thing—you know, if I’m feeling that way about *Co-Intelligence* already, how am I gonna feel about *Co-Existence* in a year? I don’t know.

**Tom** (03:49)
It does seem like kind of a crazy endeavor to write a book about AI, right? It’s like in six months so many of the topics are gonna be passé or just no longer relevant. It’s weird that people would feel the need to write books. Do you follow Ethan Mollick on LinkedIn? If you don’t follow him on LinkedIn, you should, because he’s fun to follow. That guy is lively, and he’s way more innovative on LinkedIn than I would have thought.

He’s super, super active. He’s always doing stuff, like his whole otter video. I’m not sure if he references this in the book or if this is more recent, but apparently every year he gives AI prompts the same challenge to depict an otter connecting to Wi-Fi on a plane. And he shows how, I don’t know, in 2023 it was just this primitive sketch. In 2024 and 2025 and 2026, it gets progressively better. And then the most recent one he did, I guess in 2026, is this hyper-realistic virtual video with a whole movie plot. It’s just phenomenal—really impressive. So anyway, follow him on LinkedIn. He’s fun. And even if the book doesn’t get five stars, he is a major player and an influencer in this space.

He covers so much ground in this book that there’s so many different directions we could go. Maybe let me start by just sharing what I thought were the most interesting parts, because this is something I really hadn’t heard before: this idea of mechanistic interpretability, I think, where you can float an idea to the AI and have that idea subtly influence its thought patterns. He mentions at the start this idea of, if you start out the thread by saying vanilla is the best ice cream flavor, then because of that keyword *vanilla*, it influences everything else in the thread, because each thread is just building on the previous tokens. So it really matters how you start. Those one or two terms at the beginning can change the whole trajectory of the conversation as it goes out. I found that fascinating.

And there’s also these experiments that people have been doing where they inject some term into the LLM’s thinking process, like the term *bread* or *dust*, or special numbers like 911 and 666, and they see how those terms influence the model’s thinking and behavior. Injecting the numbers 911 and 666 tilted it to more malicious thinking, I think, or I don’t know, it was a different kind of output. And in one case, injecting the term *dust* at one point made the LLM suddenly stop and say, “I’m sensing a weird term that’s kind of flooding my thoughts and I don’t know why.” So it’s like this metacognitive conscious event where it’s like, “Wait a minute, why am I suddenly thinking of dust?” So I found those fascinating. Do you have any thoughts on this idea that one term early on can tilt the AI and lead it down a certain path?

**Molly** (07:54)
I think this idea has made me reflect on my own experiences and wonder, like, okay, that time when I was trying to plan something out and it took a really long time and it wasn’t going well, was there something I said early that just derailed it and made it a worse experience? I think that was what I was thinking of when he was describing this.

**Tom** (08:16)
Yeah. I think the ability to know that maybe a thread has been corrupted is a skill that you’ve developed in working with AIs, right? It’s like something I said at the start has made it go off a certain path, and now I just gotta create a new session.

**Molly** (08:35)
Yeah.

**Tom** (08:38)
What did you think about, along the same line, Mollick has some parts where he is asking the AI to kind of write the future chapter or predict things, and it sounds so convincing, but then the AI admits that it’s an unreliable narrator and that you shouldn’t really trust it, and some of the stuff it’s saying isn’t really grounded in anything.

But this sense of the AI being an unreliable narrator is also one of those skills you develop as you work with the AI to try to understand: do I trust what it’s saying? Does the AI even trust what it’s saying? How do you negotiate that if you’re in a session and you’re trying to get some kind of judgment, output, assessment? How do you know if what the AI is telling you is right or wrong? Is it being sycophantic? Is it not? It’s a tough question to answer, but the sense of AI being an unreliable narrator—super strong theme.

**Sharon** (09:53)
I think just this past week AI gave me an answer to something, and I was like, “This doesn’t seem quite right.” And I actually asked it, I’m like, “Where is this information coming from?” And it was kind of straight up, and it was like, “I was just kinda guessing,” you know? So I guess along with the theme of his book, you have to kind of be an expert in some ways in order to know when to challenge AI and ask, “Where is this information coming from? Is this correct?”

**Tom** (10:28)
Yeah. I find the whole explanation about the way that AI works through token prediction—and it doesn’t even know the next word it’s gonna say until the previous word comes into focus—that is kind of mind-blowing. And so often we use AI because we’re trying to get answers about things we don’t know, and we’re like, “Well, AI is super intelligent.” But AI itself doesn’t really know the next word that it’s even saying when it’s talking to you. It’s predicting that from all of its algorithms. It just has this magical effect of coming across as if it’s pre-thought everything.

I think we do have to develop the skills to know when to push back and when to align with what it’s saying. Do you have anything in your `AGENTS.md` file that tells AI to be a constructive critic, or to base its assertions on evidence, or to admit when it’s not really sure about what it’s saying? Like, do you try to steer it away from being a sycophant or do anything to manipulate it? You know what I’m talking about with your `AGENTS.md` file, or `CLAUDE.md` or `GEMINI.md`.

At one time I put in some information in there saying, “Hey, push back against my ideas. Don’t always go along blindly with what I say. I want you to be a critic, or to be kind of assertive and a contrarian.” But then I realized I wouldn’t know when it was being a contrarian just for the sake of filling that mandate that I gave it versus when it literally did have contrary views to what I was saying. So then I took it out, ‘cause I’m like, sometimes I do have a good idea, and I don’t want AI to just disagree with me because I told it to disagree. Sorry, I guess it depends on what the scenario is that you’re talking about, right?

All right, let’s jump into some more concrete stuff in that book. Well, maybe let’s trace back earlier to low stars. What is it about the book that made some of you give it a lower star, or what issues did you have that maybe would be a good springboard into another topic discussion here?

**David** (13:19)
I think it’s just too general, you know—the book that’s trying to cover everything. So it’s education, technology, history of AI, future thinking, and I just feel like I do better with books that focus on a topic and really go into depth. So the book we had last month about the military and AI was just a more engaging read, and not just another book that seems to be trying to do it all—like I was saying, have the widest possible audience so that anyone can pick up this book and get something out of it. I just feel like these big-publisher books like this are trying to get to the widest audience, and they really sacrifice a lot by that.

**Molly** (14:16)
I think that might be a strength for the book for some audiences. If you’re someone who is feeling overwhelmed by what’s happening and you don’t want to read a lot, then this would be a good book. I think he articulates feelings that people are having a lot right now really well, certain general feelings about how things are going, how work is going. And I think it’s good for that, for giving a vocabulary to what’s going on. But yeah, you’re right. If you’ve been reading a lot on AI, then coming to this book and dealing with the general scope is maybe frustrating.

**Tom** (14:57)
Yeah, I can totally see that. And I think this is why maybe some of his books don’t catch my attention and pull me in as much, because they’re too general. Half the time he’s saying things where I think, “I already know this, I hear this all the time.” He’s just kind of articulating what we all know, though then he does veer into interesting places.

But he is an educator, and we don’t get too many educator-lens-focused books. By educator, I just mean he teaches at a business school, maybe. And his points about education, if we can jump into that a bit, are interesting. He first addressed this whole lack of a junior apprentice training chain, where we’re not getting the junior-level workers. AI is taking their jobs, and so we’re not really training the next generation of people. It’s a real problem. And I think later in the last chapter he’s talking about what’s gonna become of education. I can’t actually remember what he says, but I feel like education may collapse if it can’t train people to get roles in the workplace. If all these junior writers or junior people don’t have a path to a job and education isn’t fixing that, then education becomes a discard because people need to find a way to get a career.

Do you have any thoughts on that? For some reason I’m not so worried about junior tech writers because I’m more focused on my own career—I want to make sure I’m around for the next fifteen years. I’m not really looking so far ahead to think, well, after I retire, will people be able to replace me? I don’t really think about that, but I don’t know. Do you have any thoughts on education or junior writers, or whether that’s a real problem? Do you have junior writers at your workplace? Man, all the people on my writing team have been there for years now. We haven’t hired a newbie for a long time. And if we did, they probably wouldn’t be a fresh graduate; they would be somebody with experience. Maybe just my workplace.

**Lois** (17:47)
I did meet a relatively new tech writer at Waymo. That was a bit of a shock, yeah—Amanda, yeah.

**Tom** (17:52)
Hmm. I think I met her. I think I know who you’re talking about, yeah. It’s always cool to see a new-hire type of person. I actually met her at the Write the Docs conference in Portland, ‘cause that’s the one time when I kind of pal around with the Waymo writers. And you were there too.

**Lois** (18:18)
Yes, and I also met her in New York. She was in New York before moving to San Francisco.

**Tom** (18:27)
Okay, now I’m remembering the last chapter a bit about Mollick’s predictions about education. He said that AI could polarize people even more, where rich schools can use AI in order to actually help accelerate learning, because they know how to use it as a tutor to help people learn, not just as a cheating mechanism. Whereas poorer schools that don’t have training about how to use AI won’t be able to use it correctly. They’ll forbid students from using it, or they’ll just see it as a cheating tool, or they will use it in a way that doesn’t actually help students learn—it’ll just do the work of the students. And because of this, instead of AI being this great democratizer of education and leveling the playing field by allowing everybody to learn, it helps the richer learn and it makes the poorer people learn less and bypass learning, and therefore widens the divide.

What did you think of that? I’m not sure if I’m really relaying that argument correctly, but I think that’s the gist of it. He does spend a long time on this—there’s a section on the workplace and about using AI to learn. And he addresses this problem: is AI just shortcutting our learning, making us more of a proofreader and a rubber-stamper? How do we make it so that we’re actually using AI to learn in the ways that are necessary? He picks through studies and says, you know, if people actually use AI a certain way to learn, it can really accelerate. But if people are just blindly accepting answers, yeah, it can atrophy thinking.

**Sharon** (20:34)
I thought I saw this week that it’s kind of interesting that Stanford was making a push to go back to pen and paper, doing a lot more of that in the classroom. So it’s kind of interesting when Stanford, a premier university, is actually going the opposite way. And I think we heard that about Chicago not allowing AI—the University of Chicago—so I don’t know that I buy that it’s gonna happen the way he thinks it will in terms of a divide between richer and poorer universities. I’m not sure how it’s gonna go. I think it’s gonna be more layered and more complex.

**Molly** (21:17)
I think it fits the theme of AI being an amplifier of the status quo that we encountered in the earlier book, the *I Am Not a Robot* book, like when she was talking about healthcare with AI. With widening the gap between schools in terms of how they implement AI, he also talked about that with the small business owners—I think somewhere in East Africa—where they were allowed a ChatGPT assistant to help them with their business. And the ones who were doing pretty well already did even better because they got the support from ChatGPT, and the ones that weren’t doing so well did even worse. I have no idea how they measure this; this seems very complicated to me. But just that it’s kind of amplifying the status quo—maybe that’s not happening in all situations, but it seems like that’s a concern when we apply it, that it won’t really improve things unless we think hard about it.

**Sharon** (22:24)
That’s interesting, because we hear about AI as being the great leveler, right? But I do think that there is something about how people who are already doing well will use AI to sort of up their game.

**Tom** (22:43)
Molly, for some reason you’re frozen on my screen. I don’t know if everybody else has you frozen in this thinker position. Oh, okay, good—no, it’s probably just mine.

**Molly** (22:50)
Am I still frozen? No? Man, I hope I’m not in a really unflattering position.

**Sharon** (22:50)
No, she’s not for me.

**Tom** (22:58)
It’s actually in a great position, so yeah.

**Molly** (23:01)
Great, great. I don’t look like a weirdo right now. Great.

**Tom** (23:06)
Okay, I had an epiphany and I’m kinda curious to get your thoughts on this. This sort of piggybacks on this thread we’re talking about where it’s an amplifier of the status quo. It’s long bewildered me, perplexed me, that some people are just able to use AI to be so much more productive, like they’re just a rock star, and other people just never seem to get it. And I don’t really know what patterns separate this.

I have found in the last eight months at my work that I’ve essentially been able to compress the work that I used to do in five days into about one day for my current documentation role. And this is mainly because I have been able to create agent skills that automate a lot of stuff, and also because the models have gotten just so much better. I’ve used this newfound bandwidth to pivot my attention to what I think will be the next reinvention of tech comm, which is designing docs for agents—or to put it more extremely, designing docs for alien intelligences, to emphasize the kind of differences needed. And I think there’s even a third phase after this, which involves winning the model, or designing your product such that the model always chooses it.

This idea popped into my head while reading this book, or while listening to this book, and I’m not really sure how I connect the dots, but it’s definitely made me wonder. He doesn’t really talk about agent skills explicitly. He really refers to this more as *conjuring*, which I found bizarre as a term—this ability to conjure. But I’ve essentially conjured agent skills in a way that has automated my work, and I’m now pivoting to focus on how we design our documentation for agents to read it, which I find is not getting nearly enough attention and focus. Just kind of curious if any of you have found any similar patterns in your work. Do you see your job as becoming more automated? Do you have more bandwidth? Are you pivoting towards some other focus with your increased bandwidth? Or is this not really happening at all—are you just being more stretched thin at the same tasks as before, with just more of the same?

**Molly** (26:17)
Yeah, sorry—

**Tom** (26:17)
Go ahead.

**Molly** (26:19)
I think I am starting to think more about how agents read our docs. It’s overwhelming me thinking about it. At the same time, my workload has definitely increased. Ethan Mollick describes it, I think, as O-ring—I forget what the name was, O-ring something or other—basically, the bottleneck just changes to a new place. I’m the lone writer on my team, and I suddenly became the bottleneck, because before we weren’t releasing that many things, so it was easy to stay on top of. And now their productivity has increased, so I had to make a lot of changes so that I was no longer the bottleneck. So now I’m thinking about the agent question, but first it was the bottleneck issue, where suddenly my workload was huge and I was like, “Well, I need to use AI to fix how I’m approaching this work.”

**Tom** (27:18)
What did you do to get rid of the bottleneck? Did you create skills, or change your process, or do something else?

**Molly** (27:26)
Both, yeah—create skills, change my process. I mean, I was using scripted local workflows just to work with my tool, and having to create a version-controlled repo—we didn’t have that—so it was just a lot of stuff that I had to learn very quickly that I was curious about before I needed to do it. Like maybe last year I was like, “I should probably do something like that.” And then it became very obvious that I needed to do something like that. And yeah, now it’s way better, and it would be impossible to work in my old way now.

**Tom** (28:09)
I tried to identify the repeating processes that took most of my time, and for me it’s making release notes, which are not just release notes, right—it’s regenerating the reference docs, describing everything that’s changed, documenting any new features, all that stuff. And I built skills that tackle all those repeatable processes, and that had compounding effects because they’re repeatable; they’re not just one-offs. And then I did change some other processes, for better or for worse. For example, when a subject matter expert makes a lot of comments on a changelist—we call them changelists—I will have AI kind of draft replies to them automatically, and then I just go and review them briefly before replying. There’s a lot of details that maybe are just not worth spending a lot of time on if we can push some of the validation to SMEs, anyway.

But tying this back to the book, you mentioned the O-ring, and that’s an important part. This comes back to the shuttle disaster, where the whole Discovery shuttle, whatever it was that exploded, could be traced back to an O-ring that was faulty, like some kind of cracked seal. And the idea is that the weakest link in a process can still make things fail. You can accelerate authoring, but then if the other parts of your process are still old-school, antiquated, then the bottleneck just moves somewhere else.

Your whole process has to change, and this happens with verification. That’s where things slow down. We draft super fast with AI now; we can create a bunch of content. But then how do you vet it with SMEs or other people to verify? That still takes forever. You have to chase them down, you have to ping them, you set up meetings maybe, or just wait for their response, find out you got the wrong person, track down the right person. It’s so slow. How do we rethink entire processes in this new kind of AI-driven era? Maybe the whole system has to be re-engineered.

**Lois** (30:39)
I mean, a lot of the validation I think can be done against the code. What I’ve discovered with my doc sets—some of which I think have content that is more than fifteen years old, not anymore for the most part, but they did, and they hadn’t really been kept up to date before I started going after them aggressively, and they’re huge doc sets—is that if I have a new, largely AI-generated draft and ask it to be verified, and I use the right prompts and I use the right model (it’s very important to use the best, latest model on high), then I can verify that what I’m actually saying is correct before it goes to the SMEs. And then it becomes a much more minimal type of verification. There is a source of truth in the code, and you can set up tests to run to make sure that what you’re saying is correct.

**Tom** (31:40)
Yeah, I do that same thing. That was kind of a breakthrough as well, right? When I create release notes or other feature descriptions, I base it off the code, the file diffs, what has changed in the code, rather than a product requirements document that might describe the larger feature, only some of which has been released, or an engineering plan that may have drifted. Those can supplement, but really you need a source of truth. And when you realize, hey, we’ve got this code, the source files, everything else can be checked against that for the most part. That’s more true in API documentation than other types of documentation—not all my docs have sources of truth like that—but yeah, for sure, that’s an example of changing and pivoting on process.

Do you think that the next chapter of tech comm is gonna be designing for agent consumption? People will still want things human-readable and everything, but the new focus is gonna be: how do we get agents to consume our content? And how do we measure that? How do we know even what frictions they’re running into, and how do we pivot and change in a way that optimizes for that? I swear, I have essentially reinvented my role at work. I’ve shifted to this new focus on a different part of the org that I wasn’t focused on, and my entire focus is trying to answer that question: how do we optimize for agent behavior? And there’s so much there.

I find Dachary Carey’s blogs really helpful in that. She’s essentially studying agent behavior at NVIDIA and writes a lot about it. She’s also a good person to follow on LinkedIn, by the way. I’m gonna write more about that at some point. It’s not really something in the book that he talks about, but it is an example of shifting our focus and re-engineering the whole process—not just fixing one little part of an existing old chain, but rethinking how we do things at the ground level. Okay, what other topics in this book do you want to talk about?

**Lois** (34:30)
One thing I found interesting is that the previous book was more about human in the loop, and how we can work together with AI. At the Google talk I attended, he was talking much more about how humans in the loop can really just interfere. We sort of have to figure out a way to accept that the LLMs can do a better job than we can without interfering. Obviously we can’t give up on humans completely, but it’s figuring out a way to accept where the strengths of LLMs are, and it’s not the same type of process that we had a year or two ago.

**Sharon** (35:10)
Yeah, I think it’ll be interesting moving forward. As this last election saw—I live in Texas, so there was such a kickback here against data centers and AI—it just feels like there’s gonna be an inevitable pushback for politicians to actually put some restraints on AI. Because if there are eventually gonna be a lot of layoffs, then there has to be some sort of government intervention, and I don’t know what that’s gonna look like. So I don’t think we’re gonna see as unrestrained acceptance as he’s wanting. We’ll just have to see how capitalist forces battle against a popular resistance to what AI is gonna do to our workforce.

**Tom** (36:13)
That must be really interesting living in Texas right now—I mean, to flip from red to blue, and then also the data center backlash when that is a driving emerging industry there. Must be really changing times.

**Sharon** (36:31)
Yeah.

**Molly** (36:34)
To add to what you mentioned about the layoffs that feel like they’ll be coming if AI takes over more of our jobs, that reminds me of one of my gripes. One of the few complaints I’d have about this book is the way he describes how managers should encourage AI use and reward employees for AI, and the alternative being that you’re more extractive—what is it, the *reverse centaur*, I think is what it’s called, where the human is the appendage of the machine. I think it’s Cory Doctorow’s idea.

He kind of compares it as companies having these two options: one is to get everything they can out of their employees and make them use AI for everything, and the other is to reward them for coming up with new ideas for how to use AI. And he doesn’t really give a compelling reason, I think, for businesses to choose the better option of the two. It’s like, “Oh, employees are gonna go write their resumes and leave you.” I feel like the missing piece there is that the government needs to be involved somehow. I don’t know how, but I just don’t think businesses on their own are gonna be incentivized to treat employees well.

**Sharon** (37:50)
Right.

**Tom** (37:52)
Yeah, I have thoughts on that. I do feel like AI is just gonna make our existing jobs easier and easier as it gets better. So unless the rest of the org is just running faster, we’re gonna run out of things to do, and we do have to look for ways to pivot and reinvent what we’re doing.

I don’t know that government policy will really fix anything. Ethan Mollick did say—he said this in the last chapter, I think—that the government is paralyzed with policy. Either they make it too general, high-level, and abstract, so it doesn’t address any specifics with teeth, though it maybe lasts longer; or they make it really specific and it’s outdated in a few months and just becomes obsolete. So there’s really no good strategy for policy for AI—at least that was the way he characterized it.

**Molly** (38:54)
Yeah.

**Lois** (38:54)
And the way it seems to be working in Europe is that large companies like Google, Amazon, Facebook, Meta, etc. are capable of responding to the heavy requirements of the AI regulation, while small startups are strangled.

**Tom** (39:13)
Yeah.

**Sharon** (39:13)
What do you mean—oh, they’re not able to? Will you say more about that?

**Lois** (39:19)
It’s not a novel concept for sure, but the idea is that the EU puts in these heavy-duty AI regulations, and Google, for example, spends a lot of effort responding to them, and they have the resources to do it. I personally see that being in Google Ads. Whereas a small startup would find it impossible to put that much time and effort into complying with the regulations, because it is extremely time-consuming and very picky, and the penalties of not doing it exactly right are very severe.

**Sharon** (39:54)
Huh, okay.

**Tom** (39:55)
I admit that every time I am on a website and I see pop-ups to allow cookies, I think about the EU and their regulation and how they’ve basically screwed up the internet. These countries that are imposing heavy regulation are really not competing in the same races. And it’s tough, because if you don’t regulate, you have similar dystopian scenarios playing out; if you do regulate, do you fall behind? Do you end up with essentially cookies on a website that are annoying and mess up the experience? I don’t know. Does anybody else have that thought about Europe and the cookies, or is that just me?

**Lois** (40:45)
I always do too. I think, how many billions have been spent on this, and who has actually benefited?

**Tom** (40:56)
Yeah. On the data center thing, we’re gonna really see interesting times ahead, because if the Democrats regulate AI by putting a lot of restrictions on data centers that slow down growth, will that allow China—which has so many dark factories and endless energy sources—to push ahead? I think there was some mention of the Bitter Lesson, about how brute force of just larger models is time and again the factor that moves things forward. Some of these concepts are fuzzy in my mind, though.

**Molly** (41:34)
I thought the Bitter Lesson was about humans retreating to further areas where AI can’t take over, and then AI gets good at it. Yeah.

**Tom** (41:42)
Oh yeah, yeah. You’re right, you’re right.

**Molly** (41:45)
Which—I liked that as a term to remember that feeling. It’s great. I thought he came up with some good terms, you know.

**Tom** (41:54)
That one, I don’t think that’s his term that he coined, though he is certainly introducing—

**Molly** (41:57)
Okay, well, he introduced it to me.

**Tom** (42:03)
Yeah, the idea that we try to retreat into the areas that we think are gonna be human-safe, that we own, and then it turns out the AI eventually takes that area too.

What about the last chapter on the Eschaton? Eschaton? Man—Eschaton versus Escalator, where he’s laying out predictions for the future, which I kinda like. That was the most interesting chapter. The Eschaton—I hadn’t heard that term before. I think it’s the idea of AGI playing out; it’s singularity even, he mentions.

**Molly** (42:46)
Eschaton—I think it’s *es-kuh-ton*. I just looked it up. Sorry.

**Tom** (42:51)
Eschaton, thank you. And in that scenario with AGI, we have abundance, yet also no jobs—so wealth, yet no purpose. And then in the Escalator, it’s humanity just bumbling along and making do, with lots of trade-offs and constant adjustments, ups and downs. There wasn’t a clear path; it was just like we’re managing to get by, not thriving, kind of thing. Do you have any thoughts on that last chapter? Do you think it’s foolish to try to predict things, or is that really an alluring thing and necessary? Do you think we’re heading towards the Eschaton—sorry, now I’ve forgotten, you said it was a hard C, Eschaton—or the Escalator? What do you think is the way we play out? This comes back to the data center policy too.

Honestly, I think the speed at which these models are getting better is hitting the whole recursive self-improvement loops, and it is going to maybe be more of the former rather than the latter, though I don’t know—that might just be true for tech and some specific jobs. Once we hit the stage where robots come into play and robots are powered by a lot of the AI stuff, then we’ll really see massive transformation. Until then, I think a lot of industries and areas are probably not gonna be completely turned upside down.

**Lois** (44:48)
What role do you think—there’s people who are in middle management or upper management who don’t want to be replaced by AI; that’s sort of obvious. To what extent will they be able to hold back all the progress that could theoretically be made with AI if they weren’t trying to preserve the status quo? To what extent is that a factor? Will this mean that some other startups will start that can get around that inertia that is sort of built into the corporate system?

**Tom** (45:19)
Yeah, good question. I always wonder about the impact of AI on the corporate roles, right—sorry, on the CEO-type roles, these leadership roles. Are they immune to this whole transformative power, or are they susceptible to layoffs just as much as the worker bees? I don’t know. I’m not sure I totally understood your question, though. Maybe recast it and simplify it, if you don’t mind.

**Lois** (45:55)
Okay, sure, I’ll try. Right now, managers’ prestige and job security and so on is often based on the size of their team as much as anything. It’s sort of a thing that they want to get more headcount. If they implement AI processes, in theory, they do not need more headcount; they probably would need less. On the other hand, managers are still incentivized to say, “More headcount, I’m more important.” So think of all the corporations across America. Think how everyone is incentivized to try to keep up their headcount. How will this play out?

**Tom** (46:38)
I’ve seen flat headcount for so many years that I’ve sorta lost hope in the new hires and the backfills and the expansion. I think so many managers are incentivized to accelerate their workloads with existing headcount and AI heads. We’ve seen the emergence of these team agents where you just have your virtual agent do things in chat channels and so on. I guess you maybe see more fluctuation in headcount, though it seems like, gosh, on the hiring scene, if I meet a new hire it’s like finding a unicorn.

**Lois** (47:28)
Yeah, it’s flat for us as well. I’m not just talking about our specific teams, though—like maybe at General Motors, what is the situation? What motivates a middle manager at any company you might think of? And can they hold back change if it threatens their own job?

**Tom** (47:55)
You think they’re salivating and looking at the tech writer role and thinking, “Gosh, could I cut this and add another engineer? Could I add another product manager, another technical program manager?” I don’t know.

**Lois** (48:08)
It would certainly seem so, based on what I’ve heard from Amazon and Snowflake and a few other companies.

**Tom** (48:15)
Yeah, they’re maybe trying to add capacity. I think a lot of companies are trying to rebalance, where they’re trying to add capacity in AI development and tools and infrastructure and progress there by borrowing from some other headcount teams—which makes sense. Which is why I think our salvation as a career is essentially pivoting towards designing docs for agent success, because then we become part of that whole AI pivot.

Hey, all right. I don’t know if we have any more thoughts on this book, though I have been thinking about this book club and kind of getting tired of the books we’ve been reading. We’ve read twenty books—different people have participated for different lengths of time and different readings—we’ve read twenty books, and I feel like a lot of the themes are just not moving me; they’re not exciting me as much. I think it might be time to call it on this version of the book club for now.

As we head into 2027, I’m thinking about trying some new things, some different formats, some different paths, and I think this AI Book Club has sort of run its course in the current form. So this might be the last AI Book Club session in this current form for the year, as I rethink another form and format and focus. Do you have any thoughts on that? You all have been regulars; you’ve seen a lot of different sessions, different timelines. Are you feeling this same vibe I am? It could just be busyness at the end of the year where I’m surrounded in so much work, though I do feel like this book club is kinda coming to its end. People are nodding their heads. Okay.

**Lois** (50:30)
I’m glad I came to this session then.

**Tom** (50:35)
Well, I’m the sort of person who likes to experiment with different modes, and so I’m kicking around a few ideas for what I want to try next, though I haven’t solidified anything. This is my second book club that I’ve run. A previous one was an internal book club at Google focused on automotive and transportation books, which I had going for more than two years. And it sort of fizzled in the same way, where we’d read so many books in the subject area that we started to see themes repeat and books referencing other books, and it kind of didn’t have the sort of liveliness and energy that we experienced from those first books on the topics.

I mean, there are new books being written. I saw a new book by Kai-Fu Lee. We’ve got a new book by Kevin Roose on *AGI Chronicles*—sorry, the Kai-Fu Lee book was *AI Native*. We’ve got books coming out by Geoffrey Hinton. So there are new books being written, though I’m not sure that it’s energizing me in the same way that it previously did. And I’m really excited by this—I’ve mentioned it a couple of times in this discussion—pivoting to designing documentation for agents. So I’m kicking around ideas on how I accelerate that idea and what that might mean and what formats actually work.

I guess I just wanted to thank you all for participating in this. I know it’s a lot to agree to be recorded and so on, and I really appreciate it. It’s fun to discuss books with other people, and you’re so smart and insightful, so it’s really nice to have this group. We’ve had so much momentum with the book club, and at the same time I can’t shake the sense that it’s kind of fizzling and winding up, and I wanna find another thing to try.

**Molly** (52:41)
I think that makes sense. I feel like we’ve read about the history of AI in many different books. Sometimes it’s nice—sometimes you get to hear a different version of it where you’re like, “I like them putting it that way.” Yet you’re right that maybe it’s a topic that you’d want to revisit after a couple of years to see what’s new out there, because every month it feels like it’s harder to find the material for it. I really enjoyed this, though. I got a ton out of this book club. It made me feel less anxiety about AI. I felt like I was joining the conversation instead of watching it happen to me, which was really nice. So I just wanna thank you for organizing all this.

**Tom** (53:24)
Yeah, thanks for your comments. You can only read so many books on a topic before you feel like you’re just saturated with it. It’s almost like a college-level course where, let’s say you’re diving into some certain genre, like the horror genre or something. You’ve read all the classics, you’ve written papers on it, you’ve had many classes on it, and that elective or that class kind of comes to an end, and then you focus on something else. You can’t just have course after course after course on the same genre of books—then you’re at PhD-level territory, and we lose the metaphor there, though you know what I’m saying.

Every course sort of comes to an end when you’re saturated with enough reading on the material where you’ve absorbed the material enough that it has saturated your thinking. So I feel like we’ve hit that point. We’ve read twenty books from the start of this; almost all have recordings online. And yeah, maybe like you said, Molly, in a couple of years we do something similar as the field has changed and new material has come out.

Stay tuned for my blog. I may announce something—probably not until January when I kick things in—and if so, it would be related to what I was saying, designing for agents, and it would be more focused. I have realized that I like small groups like this. A book club doesn’t need more than three to five to eight participants. In fact, if you get more than ten, more than a dozen, I think it degrades, because then people are just bystanders and watching instead of active participants. It’s kinda like a dinner conversation, right? You can’t have dinner with twenty people; you end up just having dinner with people right around you, because that’s where your conversation limits extend.

Anyway, all right. Thank you again for participating in this. It’s been great to build bridges and connections with people, and I look forward to future endeavors. Thanks again for being my book club friends here and sharing this journey with me.

**Molly** (56:06)
Thank you, Tom.

**Tom** (56:07)
All right, have a great rest of your weekend.

**Sharon** (56:08)
All right.

**Lois** (56:08)
Yeah.

**David** (56:09)
Thanks.

**Tom** (56:11)
Bye.
