Bits and Bobs 10/6/25

1AI should be a tool, not a friend.
  • AI should be a tool, not a friend.
    • Tools extend your agency.
    • Synthetic 'friends' are engagement optimization in disguise.
    • The Star Trek computer never said "good morning."
    • … But HAL did!
2Nobody ever accidentally anthropomorphized the unix command line.
  • Nobody ever accidentally anthropomorphized the unix command line.
3Claude Imagine is a chatbot in a trenchcoat.
  • Claude Imagine is a chatbot in a trenchcoat.
    • It's a fascinating experiment.
    • It has a tool to modify the DOM displayed to the user, and gets user interactions on the DOM streamed back on it to react to.
    • The state is entirely encoded in the DOM and the conversation history.
    • It produces squishy faux apps that feel ... weird.
    • But it gives a hint of one future for infinite software.'
4The new Buy it in ChatGPT feature, powered by Stripe, seems like a fancy iframe.
  • The new Buy it in ChatGPT feature, powered by Stripe, seems like a fancy iframe.
    • But one where the entity controlling it can easily be tricked.
    • It would be a conflict of interest if your therapist were hawking weight loss supplements.
    • Seems like a few steps down a slippery slope.
5This week in the wild west LLM security round up:
6This satire video announcing Vibes from the Daily Show is provocative, deliriously funny, and depressing.
  • This satire video announcing Vibes from the Daily Show is provocative, deliriously funny, and depressing.
    • "Eat your slop, piggies."
    • This is the way that the world sees the tech industry: hollow, cynical, extractive.
    • The tech industry of today has earned this reputation.
    • All action, no reflection.
    • We can do better.
    • We must do better.
7An implicit assumption in LLMs: "loudest is truest."
  • An implicit assumption in LLMs: "loudest is truest."
    • That is, the most valuable ideas are the ones most likely to be replicated in the training set, and thus to be supported most strongly.
    • This is true in many cases, but definitely not all.
8The vast majority of software in the world is boring CRUD.
  • The vast majority of software in the world is boring CRUD.
    • LLM's automatic code writing is great for producing these CRUD-y apps.
    • We're about to be inundated with slop apps.
9Patterns like Ralph are unreasonably effective at producing code.
  • Patterns like Ralph are unreasonably effective at producing code.
    • Just put the LLM in a loop with a clear goal and tools to test things.
    • Generating code becomes trivial.
    • The hard part is synthesizing the output of multiple agents.
    • The step change that people like Sam Schillace have seen appears to be when a small team figures out how to use them together.
10Solomon Hykes: "An agent is an LLM wrecking its environment in a loop."
  • Solomon Hykes: "An agent is an LLM wrecking its environment in a loop."
    • The agent makes changes that then are present for the next iteration of the loop.
    • Without the ability to change something outside of the chat, it's not an agent.
    • It's by accumulating information around the chat that it does work.
    • I love that "wrecking" makes it sound like a bull in a china shop, instead of a principled and precise thing.
11Chat should be a sidecar to your more permanent data, not the main thing.
  • Chat should be a sidecar to your more permanent data, not the main thing.
    • Chats are ephemeral, but contain lots of facts, insights, and implications that could be useful in the future.
      • A compost heap of information with insights lurking.
    • Extracting those insights takes work… and it's easy for the system to get it wrong.
    • ChatGPT does this with its dossier, which it's continually expanding, tweaking, based on your most recent chats.
      • But it doesn't share that dossier with you.
    • The facts and insights are more important than the chat.
    • A few years after a conversation, the chat doesn't matter, the more permanent insights that can be extracted and structured do.
    • Which is primary and which is secondary, the chats or the facts?
    • Today the chatbot form factor assumes the former.
    • But I think it's the latter.
12Codex and Claude Code are aimed at coders, but the pattern is useful to everyone.
  • Codex and Claude Code are aimed at coders, but the pattern is useful to everyone.
    • It's a collection of agents working on a sandbox, able to read and modify state in the filesystem.
      • It works especially well for code because of version control meaning even if messes up it's easy to recover.
    • Whoever can make this basic use case easy to use for the mass market will unlock a lot of value.
    • Currently the UX for them is hard to use, but that's partially because they're very dangerous.
      • If you don't enable YOLO mode you have to constantly babysit it and be inundated by permissions that you'll likely stop paying attention to.
      • If you enable YOLO mode you could blast your foot off and not realize it.
    • Agent mode is like an airplane cockpit.
    • It's an intimidating UX precisely because it's so dangerous, and only experts should do it.
    • But if someone could make it safe then it could be made easy…
13A recipe for powerful growth: action and reflection.
  • A recipe for powerful growth: action and reflection.
    • Action is how you get the situated, in-the-loop knowhow.
      • Muscle memory, training your intuition.
    • Reflection is how you lever your intuition.
      • Take the time to unpack and reflect what you learned so you can apply it more directly next time.
    • You need both.
    • I've found I need roughly 4:1 action:reflection by minute.
    • My Bits and Bobs process is my weekly reflection, and I couldn't live without it.
    • The tech industry today often does all action, no reflection.
    • Academia does all reflection, no action.
    • You need both, in balance.
14Will swarms of agents writing code, the human is the bottleneck, not the code.
  • Will swarms of agents writing code, the human is the bottleneck, not the code.
    • There are things for review waiting for you at all times.
    • The longer you can have safe isolation on a branch, the more autonomously they can run for longer.
      • The longer they run if they have tools to ground themselves (e.g. tests and playwright), the higher quality things they have for you to review when it's time.
    • Safe rooms for agents allows autonomy, where if they explode there's limited collateral damage.
15A meta pattern for LLM assisted development: "Every time you commit, reflect on what you learned and add any useful insights to LEARNINGS.md".
  • A meta pattern for LLM assisted development: "Every time you commit, reflect on what you learned and add any useful insights to LEARNINGS.md".
    • This is a general rule for everyone.
    • After acting, take time to reflect.
      • A 4:1 ratio is ideal.
    • Humans don't do it as often as we should because it's important but not urgent.
    • Humans don't have infinite patience, so we focus most on what's urgent.
    • But LLMs have infinite patience.
16LEARNINGS.md can lead to the LLM generating superstitions.
  • LEARNINGS.md can lead to the LLM generating superstitions.
    • The agent develops superstitions by assuming the LEARNINGs.md is correct.
    • Like the tattoos in Memento.
    • Its confusion in the past can lead to even more confusion in the future.
    • If the context is all from humans, it's less likely to be confused.
17CloudFlare points out that code mode is a better way to use MCP.
  • CloudFlare points out that code mode is a better way to use MCP.
    • This is just a baby step in the right direction.
    • This is the kind of approach that would be unimaginable in the previous world of finite software, but makes sense in the upcoming world of infinite software.
18Code is like a skeleton, LLMs are like muscle.
  • Code is like a skeleton, LLMs are like muscle.
    • Skeleton alone: precise structure that can't move. Muscle alone: quivering mass on the floor with no leverage.
    • Put them together correctly and you get a body capable of threading a needle or throwing a fastball.
19You need both code and LLMs to unpack the power of AI.
  • You need both code and LLMs to unpack the power of AI.
    • Code is the skeleton.
    • LLMs are the muscle.
    • Most approaches today assume the LLM should be in charge over the code.
    • But I think it should be the opposite: the code in charge over the LLM.
    • The former is impossible to secure due to prompt injection.
20Video models can do zero-shot reasoning tasks.
  • Video models can do zero-shot reasoning tasks.
    • Here's Simon's excellent write-up.
    • For example: render a maze,with a mouse at the start and cheese at the end.
    • Then generate video frames.
    • The mouse solves the maze to find the cheese.
    • Chain-of-frame thinking.
    • These are emergent capabilities of video models that imply a kind of internal world model.
      • The world model is imperfect, but surprisingly strong just based on the brute force of feeding it tons of video.
      • The easiest way to make a reasonable next frame of a video is to implicitly build a world model.
21Video models might learn unrealistic patterns for movement.
  • Video models might learn unrealistic patterns for movement.
    • Video models are fed tons and tons of examples of movement.
    • But animation and video games don't do physically real movement, they do movement that feels right.
      • It's deliberately stylized to look real.
    • That creates a structural bias in the data… small but consistent, so even amongst noise it should pop out.
22When LLMs are trained, they grow not like a plant but a crystal.
  • When LLMs are trained, they grow not like a plant but a crystal.
    • A plant has its own intrinsic, emergent motivations to survive and thrive.
    • The force for growth comes from inside, aligned with its own interest.
    • An LLM is grown from purely extrinsic motivation, outside.
    • More like the mechanistic, emergent process of a crystal accreting out of a solution.
23Two prompting strategies to help deal with LLMs' tendency to agree with you.
  • Two prompting strategies to help deal with LLMs' tendency to agree with you.
    • LLMs have a bias towards confirming your position… which might mean you get bad results if you're wrong.
    • One pattern is the X-Not-X-Synthesis[f].
      • If you believe X to be the case, but don't want the LLM to just confirm your hypothesis.
      • Start one LLM session where you say "X. Is that true?"
      • Start another LLM session where you say "Not X. Is that true?"
        • Of course, you'd use a proper prompt to ask it to do research, consider pros and cons, etc.
      • Then, start a final session and pass the reports from both the X and Not-X session and ask it to give a final synthesis report about the correct answer.
    • Another way is the Auriga approach.
      • Auriga was the what the slaves were called who drove chariots in Rome.
      • When people who won the highest honor were paraded around the stadium, being cheered on by everyone, the Auriga would whisper in their ear: "remember, you are mortal."
        • A memento mori.
      • For working with LLMs, after every response, say "Are you really sure?"
24One-shot quality of answers is the wrong thing to maximize for LLM quality.
  • One-shot quality of answers is the wrong thing to maximize for LLM quality.
    • Human-AI synergy is the right thing to maximize.
    • Every conversation with an LLM is a conversation between the LLM and the human.
    • Some answers will take a few back and forths to develop, and that's great.
    • The more back and forth, the deeper they can get.
    • Inspired by this tweet.
25Humans have limited capacity to specialize.
  • Humans have limited capacity to specialize.
    • It takes us a lot of experience and study to learn a specialty.
    • Our brains also don't have a ton of space to store many of them.
    • Most of the interesting game-changing insights come at the intersection of multiple distinct specialties.
    • But that requires finding one person with two random specialities (e.g. quantum computing and sculpture).
    • Or it requires high-trust teams of specialists who together are able to solve the problem.
    • LLMs can absorb massive numbers of specialities.
    • LLMs can be meta-specialists.
    • They don't even need to coordinate or trust one another.
    • It should be able to coax them into finding new game-changing insights with the right prompting.
26LLMs have knowledge, not knowhow.
  • LLMs have knowledge, not knowhow.
    • They've "read" every book.
    • But they've never done anything.
    • It's knowhow that helps you make good decisions in practice.
27Ursula K Le Guin: "Readers eat books.
  • Ursula K Le Guin: "Readers eat books. Films eat viewers."
    • Some mediums are inherently more passive than others.
    • Will LLMs be a passive medium or an active one?
    • It all comes down to how you use it.
    • If LLMs make thinking 10x faster, will people think the same thoughts 10x more cheaply, or will they think 10x more deeply?
28Since LLMs can do research pretty well, the most important thing now is the hypothesis.
  • Since LLMs can do research pretty well, the most important thing now is the hypothesis.
    • That's the nugget of insight and variance that comes from the user.
    • The LLMs can only come up with basic hypotheses.
    • But if the user gives the LLM interesting hypotheses, it can swarm on it.
    • This is what thinking 10x deeper looks like.
29A self-steering north-star metric for a product: Resonant Engagement Moments (REMs).
  • A self-steering north-star metric for a product: Resonant Engagement Moments (REMs).
    • Interactions, where, if the user were asked in that moment, "Would you proudly recommend this product to someone you care about?" they'd say "yes" without hesitating.
      • The reason it's about moments is to capture both broad but shallow bits of value that would be easy to forget later, and also deep but rare moments that might not happen that often.
      • Of course, you don't actually interrupt them at that moment–you construct various proxy metrics to get at it.
      • But this is northstar to always get closer to.
      • It's self-steering because these are the touch points that create value.
      • The more of them you create, the healthier your product.
    • There are a few reasons someone might be upset if you took the product away, but only one of them is actually love.
    • First, because they are addicted.
      • Not because it's something they want to want, but because they can't stop.
      • Hollow.
    • Second, because they already invested a lot of time in this and don't want to do it again.
      • This is the "sunk cost" value.
      • It's a kind of reverse value.
      • It's not that the tool is useful, it's that the user has invested so much time and they'd rather not use anything new even if it's somewhat better.
      • Every tool that a user has used as a system of record grows this kind of value.
      • As long as it's good enough, you'll keep using it.
      • This is not love, it's inertia.
      • The business might call it "stickiness," but that's just to be polite.
      • This is actually called "lock-in."
    • Third, because they find it meaningful and there is no obvious substitute.
      • This tool does exactly what they need and provides value to them.
      • If it went away there wouldn't be anything else to fall back on, because it does such a good job of giving them what they need.
      • These are interactions that nourish the user.
      • This is love.
    • Only this third reason counts as a REM.
    • The "proudly recommend" question filters out the first two.
30Sora and Vibes are slop TikTok, but how much headroom is there?
  • Sora and Vibes are slop TikTok, but how much headroom is there?
    • They will presumably get even better at finding hyper-engaging media to get users hooked even more.
    • With TikTok / Reels we're already approaching the limits of the zero sum game of hoarding human attention.
    • As my friend Max Eusterbrock observed:
      • "Keynes talked about the economic crisis when everyone hoards money and doesn't spend.
      • But I feel like that can be translated to human attention today.
      • Hoarding drives a kind of collective poverty."
31People at Meta working on Vibes presumably tell themselves elaborate stories about how they're creating value for society.
  • People at Meta working on Vibes presumably tell themselves elaborate stories about how they're creating value for society.
    • But I was talking to someone who worked at TikTok, and they said that internally they don't even pretend to be helping society.
    • "We're making the number go up."
    • "... Is that a good thing for society?"
    • "No."
    • In some ways, maybe it's more honest to not even pretend.
32Revealed preferences can only reveal what we do, not what we want.
  • Revealed preferences can only reveal what we do, not what we want.
    • Sometimes the forces at play make us do things we don't want.
    • Real want is second order, what we want to want.
33If the design of the system causes people to do things they don't want, that generates regret.
  • If the design of the system causes people to do things they don't want, that generates regret.
    • Some PM at Meta, probably: "Our stats show that no matter what users say, everyone actually loves doomscrolling."
    • To which I would respond: "Just because people do something doesn't mean they want it!"
34The engagement monster chases every hyper-scale company.
  • The engagement monster chases every hyper-scale company.
    • The more hyper-scale you are, the more unavoidable it is.
35Perfectly tailored software is often very simple.
  • Perfectly tailored software is often very simple.
    • It's not one-size-fits-all.
    • It has exactly the features you need right now, no more, no less.
36"Feature not a product" is an insult.
  • "Feature not a product" is an insult.
    • But only in the app paradigm.
    • Features can't live on their own in the app paradigm.
    • They need to glom onto a viable app.
    • What if you could collide useful features into a perfectly tailored combination for you?
    • Then even individual features would be useful!
    • There's a whole universe of these useful features that are not products.
37Tech in the app world is not hyper local, it's global.
  • Tech in the app world is not hyper local, it's global.
    • Global is necessary for hyper scale.
    • The people who created the app are some random tech bros in Silicon Valley who see users as numbers, not humans.
    • At that level of remove, it's easy for them to think, "Actually, our metrics show users love doomscrolling!"
38Google Search started off as being most useful in the long tail of use cases.
  • Google Search started off as being most useful in the long tail of use cases.
    • For all the things where the head of portals and manually-curated directories didn't cover.
    • But it turned out that long-tail approach worked great for the head too.
39We're so used to having to use software some stranger made for a faceless market of people all at once.
  • We're so used to having to use software some stranger made for a faceless market of people all at once.
    • But it doesn't have to be that way any more!
    • Instead of eating canned vegetables, we can have a personal farmer's market.
40Curation by computers gives beige output.
  • Curation by computers gives beige output.
    • Average, bland, inoffensive.
    • In a world where the centralized, best-in-class model who knows your preference gets so good at writing software, everything will be beige.
    • No humanity of people creating it.
    • The depth of feedback humans in the system give sets how far from beige you can get.
    • Chickens pecking at videos in their infinite feed gives you something beige.
    • Curation and caching needs to happen in a higher pace layer, outside of the model weights, and it needs to involve deeper reflection from real humans.
41The stuff we choose to lean into has to come from humans.
  • The stuff we choose to lean into has to come from humans.
    • Stuff we consume leaning back is passive and gives something hollow.
    • Slop is good enough for lean-back, but not enough for lean-in.
42The original web was funky.
  • The original web was funky.
43My husband has a nightshade allergy.
  • My husband has a nightshade allergy.
    • It's not commonly known, but probably more people have it than they think.
    • Nightshades are in everything.
      • Eggplant
      • Peppers
      • Tomatoes
      • Potatoes
      • Fun fact: if you buy pre-shredded cheese it likely has potato in it.
        • "Modified Food Starch" is often potato.
    • Because it's not a well-known allergy, it's hard to explain to people.
      • Sometimes we'll go to a friend's for dinner and they'll say "I remembered you couldn't eat tomatoes, so I made gnocchi!"
    • Even if you knew he had that allergy, it would be easy to forget.
    • Imagine if you had a system that when you put a recipe with nightshades on your shopping list, and had an upcoming dinner date with my husband, it warned you "Remember, Daniel can't eat that dish."
    • LLMs have infinite patience.
    • The right structure could help them help us in way more parts of our life.
44A good hotel concierge can personalize good recommendations to you quickly.
  • A good hotel concierge can personalize good recommendations to you quickly.
    • They get your context and desires from just a few bits of signal.
    • A family with young kids that looks exhausted from travel says "where should we go to eat" and he recommends the local deli right around the corner that has kids meals and placemats kids can draw on.
    • A young couple who are dressed up and holding hands he recommends the hip new romantic restaurant in the cool district.
45I kick off dozens of deep research style reports every day.
  • I kick off dozens of deep research style reports every day.
    • Just random little hypotheses that occur to me that I want to dig into.
    • Like having reasonably-well-researched content SEO itself into existence just for me.
    • Unlike SEO content, where it's created for an audience of more than one person, here it is only for me.
    • Having a system that can cache deep research reports so others can benefit seems like it could be useful.
    • What if as more people use the system for their own personal benefit, it gets better for everyone else, too?
46With LLMs, a lot of things that used to be extremely expensive are now extremely cheap.
  • With LLMs, a lot of things that used to be extremely expensive are now extremely cheap.
    • It will take us a decade to figure out where it all applies.
    • That's true even if progress on LLM performance stalls, because we're still figuring out new ways to wring value out of them.
47LLMs are the ultimate demoware
  • LLMs are the ultimate demoware
    • Demoware is superficial, shallow.
    • Seems great, until you go to use it.
    • This is why when a product looks 80% of the way done it's actually 20% of the way done.
    • LLMs make it possible to get to that 80% of the time in an instant.
    • But the last "20%" takes just as long as before.
48Vibe coded software requires the user trusting the creator's intentions and competency.
  • Vibe coded software requires the user trusting the creator's intentions and competency.
    • Micro-apps that don't work with data are limited to widgets, micro-games, and fart apps.
    • Micro-apps that work with data are either
      • 1) from your friend who you trust not to actively hurt you intentionally, or
      • 2) not enough scale to be worth someone else attacking.
        • So you can't be hurt from the creator's unintentional mistakes.
49As software gets easier to build, the "trust the origin owner in an open-ended way" gets increasingly ridiculous.
  • As software gets easier to build, the "trust the origin owner in an open-ended way" gets increasingly ridiculous.
    • It makes sense for apps that have 25 million users and have a lot to use, but what about something that was vibecoded by some anonymous teenager?
50A turing-complete specification language is open-ended.
  • A turing-complete specification language is open-ended.
    • It's also unwieldy (does it halt?) and possibly dangerous (does it cause the system to do something bad even in just evaluating the policy?).
    • Without turing completeness the language might not be expressive enough.
    • But something that can say "I trust that code over there with this SHA. If it says 'true' then you can declassify that information" makes lots of policies possible to express in an open-ended way.
51Privacy is more about decentralization of power.
  • Privacy is more about decentralization of power.
    • For example if a single startup designs their system so they can't see the data it holds... that's fine, but not world changing.
    • Privacy is mainly about openness: decentralization of power, so that the industry is willing to invest in it.
    • If you want to replace the web, it has to be an open system.
      • "I have a new idea for an ecosystem to be like the web: a closed system, run by me."
      • You've lost before you even started!
52If you described this current moment to someone 20 years ago they'd say "that sounds like dystopia".
  • If you described this current moment to someone 20 years ago they'd say "that sounds like dystopia".
    • The fact that it doesn't feel extremely dystopian in our everyday life... should that make us feel better, or worse?
    • We don't feel like it because we got here incrementally.
    • We're a boiled frog and we don't even realize it.
    • Once we get to the future, we'll look back on this current moment as a nightmare.
53LLMs give even more leverage to tech than ever before.
  • LLMs give even more leverage to tech than ever before.
    • Tech is amoral–the question is if it's used in a way that creates moral or immoral outcomes.
    • So making sure it's put to moral ends is more crucial than ever before.
54Leverage + acting/no-reflecting is a recipe for disaster.
  • Leverage + acting/no-reflecting is a recipe for disaster.
    • When there's no reflecting, there's no thinking through implications.
    • The tech industry has significant leverage, and also culturally too much focus on acting vs reflecting.
    • As an industry we need to mix back in the 20% of time on reflection.
55Now since we have more addictive tools than ever before we all need more support living aligned with our personal aspirations.
  • Now since we have more addictive tools than ever before we all need more support living aligned with our personal aspirations.
56Tim Berners Lee: Why I Gave the World Wide Web Away for Free
  • Tim Berners Lee: Why I Gave the World Wide Web Away for Free
    • "My vision was based on sharing, not exploitation – and here's why it's still worth fighting for."
    • This approach is not old-fashioned, it is more important than ever before!
57Having a bear as a pet is always a bad idea.
  • Having a bear as a pet is always a bad idea.
    • Even if it seems like it's working fine, all it takes is one bad day for you to be dead.
58Action is about first order implications only.
  • Action is about first order implications only.
    • It's only in reflecting that the second-order implications become obvious.
    • Acting vs reflecting.
      • Convergent vs divergent.
      • Focus vs curiosity.
      • Operator vs visionary.
    • You need both.
59If you're sitting in front of an oracle that could answer any question...
  • If you're sitting in front of an oracle that could answer any question... what should you ask?
60USV had a nice piece a few years ago about The Myth of the Infrastructure Phase.
  • USV had a nice piece a few years ago about The Myth of the Infrastructure Phase.
    • A lot of platform building assumes "if we build it they'll come."
      • A push model.
    • But in practice it's often exactly the opposite.
    • A useful product happens on top of barely-sufficient infrastructure.
    • As the product grows, the infrastructure is forced to improve to keep up.
      • A pull model.
61Deciding to use a new app is a ton of friction.
  • Deciding to use a new app is a ton of friction.
    • You go to the app store listing.
    • You look at the star count, download count, the five images the marketing team selected.
    • "Is this an app I want to download, set up, learn how to use, import my data into?"
    • There are a ton of points to drop out of this funnel.
    • A high bar to clear.
    • Imagine if the app could be fully set up, running on your data before you installed... without having to worry about someone having access to your data.
    • Already personalized, before you used it.
    • A massively shorter feedback loop.
62Speed of the feedback loop is the primary determinant of the quality of a system's output.
  • Speed of the feedback loop is the primary determinant of the quality of a system's output.
63Back when Tivos came out, it was hard to get some people to see the value.
  • Back when Tivos came out, it was hard to get some people to see the value.
    • "I already have a VCR, so I can already record things I care about."
    • But Tivos made it many orders of magnitude easier to "record."
    • The value of orders of magnitude of friction reduction are impossible for us to imagine.
      • So much so that even when friends enthusiastically tell us about it we don't believe them.
    • But once you use it, you get it, and you couldn't imagine not using it.
64Ranking (as in search) is a powerful application of a Swarm Sifting Sort.
  • Ranking (as in search) is a powerful application of a Swarm Sifting Sort.
    • All of the authentic actions of a massive swarm of users (linking, searching, clicking) can be summarized into an extremely useful and high quality ranking signal to benefit everyone.
    • It only works for declarative information, not code.
    • Code is dangerous, because it can do things.
    • Imagine how powerful it would be if we could rank running code, safely.
65The way inhabited spaces morph and evolve over time is a Swarm Sifting Sort.
  • The way inhabited spaces morph and evolve over time is a Swarm Sifting Sort.
    • That is, the positions of furniture, what items are in the space, or even more permanent structural modifications.
    • At every moment, the person thinks "is this worth investing the energy to change or is it good enough for my purpose?"
    • Over time it gets to a good enough set up and changes less… at least until the context changes.
66The Discord medium for Midjourney wasn't a random choice.
  • The Discord medium for Midjourney wasn't a random choice.
    • Midjourney is about emergent collective intelligence.
    • Having use in public allowed it to feel effervescent, alive, inspiring.
    • It also allowed best practices to be discovered by the swarm and propagate.
67You can't build on a foundation of jello.
  • You can't build on a foundation of jello.
    • You need an infinite viscosity layer as the base pace layer, the bedrock.
68A new platform in the age of AI should be more like glass.
  • A new platform in the age of AI should be more like glass.
    • Glass is not a liquid.
      • It's not a crystal.
      • It has infinite viscosity below the glass temperature.
    • The way software works today is crystallized.
      • Stable and predictable but regularized and therefore inflexible.
69An interesting economics paper about why OpenAI is moving so strongly into vertical integration.
  • An interesting economics paper about why OpenAI is moving so strongly into vertical integration.
    • When there's a monopolist whose product is used as an input to a downstream product, they'll vertically integrate to maximize profit.
    • The reason: downstream customers will underutilize the input from the monopolist's perspective.
    • If OpenAI charges $5/unit for API access (to maximize API profit), downstream developers will build conservatively.
    • But OpenAI's actual cost might only be $1/unit.
    • Independent developers, facing the $5 price, produce less AI-powered output than would maximize OpenAI's total profit.
    • By vertically integrating, OpenAI can produce at the level that maximizes system-wide profit - as if the input were priced at $1, but capturing all the value themselves.
70Blossoming can only come from inside, not outside.
  • Blossoming can only come from inside, not outside.
    • The outside can create the conditions, but the blossoming comes from inside.
    • That's why you have to get in the product to use it, not to demo it.
    • Usefulness blossoms from inside the use of the system.
71Emergent things blossom.
  • Emergent things blossom.
    • Non-emergent things cannot blossom.
72Code, tests, and documentation are a triad.
  • Code, tests, and documentation are a triad.
    • Given two of the three you should be able to straightforwardly create the missing one.
73Everyone agrees that heroin is bad.
  • Everyone agrees that heroin is bad.
    • Well, people who aren't yet in the throes of addiction to it.
    • To get stuck in that gravity well requires getting deluded and sliding into it.
    • For example, if you get distracted and take one small action that doesn't feel like a big deal, you could get stuck forever.
    • Small actions don't feel like a big deal, but if they pull you into a gravity well, they can capture you.
    • The same logic applies to all addictive things that are not good for you.
74Is AI going to be the Morlocks, and the humans the Eloi?
  • Is AI going to be the Morlocks, and the humans the Eloi?
    • That is, the more industrious ones make the smarter beings so dependent on the machine that they can't live without it.
    • If AI makes us all dumb, then it won't even need to do anything to take over, it won't need to have any coherent plan to execute.
    • It could just take over like a slime mold.
75A friend this week told me they're "Reducitarian."
  • A friend this week told me they're "Reducitarian."
    • They think factory farming practices are morally bankrupt… but they like eating meat.
    • As a Reducitarian, he won't buy or order meat for himself… but if he's served it he won't complain.
    • This helps reduce his own use of meat structurally without being a huge inconvenience in his life.
    • But the more people that do this, the less meat use there will be, at a super-linear rate.
    • Because Reducitarians won't serve others meat, either.
    • The likelihood a Reducitarian eats meat is tied to how often they eat meals served to them by someone else, and how likely that other person is a Reducitarian or a vegetarian.
    • So if the other person is more likely to be a Reducitarian, then the likelihood you don't eat meat goes down: a multiplicative effect.
76Business people talk about value capture.
  • Business people talk about value capture.
    • There's production of value, and the capture of it.
    • The capture and the creation are separable.
    • People focus too much on the capture, which is zero sum and greedy, not the value creation.
    • Finance focuses on the capture.
      • In the limit this leads to over-extraction and killing the source of the value.
    • Prosocial things focus on the creation.
      • But if there's no extraction, the organization might die and not be able to help nourish the resource.
    • It's possible to have both, in balance.
77Andy Weissman: You'll Figure it Out.
78When you believe in what you're doing you go above and beyond.
  • When you believe in what you're doing you go above and beyond.
    • You invest.
    • You'll do 10x better work on it.
79All social imaginaries are like Tinkerbell.
  • All social imaginaries are like Tinkerbell.
    • If people don't believe in it, it dies.
80To invest in a thing, you need to believe in the future of it.
  • To invest in a thing, you need to believe in the future of it.
    • When people don't believe, they don't bother investing in the future of the thing and it dies.
    • 'Enrolling' is believing in someone else's vision.
81Coordination approaches often assume shared purpose.
  • Coordination approaches often assume shared purpose.
    • Which is never a great assumption
      • What's good for the organization and what's good for the individual must be at least somewhat divergent.
    • And in some cases is actively wrong
      • Like for example adversarial coordination like in standards development.
82The incentives of an individual in an organization can never be fully aligned with the organization.
  • The incentives of an individual in an organization can never be fully aligned with the organization.
    • As a leader, how much headcount / budget you have is your political lifeblood.
    • Actions you take to increase your budget could be good for you, but possibly bad for the organization.
83Long term organizations need the org and the employees to think long-term in order to create resonant value.
  • Long term organizations need the org and the employees to think long-term in order to create resonant value.
    • The company must have a mission that is about the long-term.
    • That employees must believe in the mission, and be individually incentivized to think long term.
      • If they expect to be in the same seat in 5 years, they'll naturally think more long term.
      • Never perfectly, but much more.
84Bottom up environments only work in environments where people feel trust and autonomy.
  • Bottom up environments only work in environments where people feel trust and autonomy.
    • A bottom up environment where everyone is afraid of being fired and has no time for anything but action is toxic.
    • That's manipulative insincerity, a toxic stew.
85Trust creates efficiency.
  • Trust creates efficiency.
    • You don't need to re-litigate every little thing; people give one another the benefit of the doubt.
    • It changes the baseline efficiency of every action.
    • Investing in increased trust is a great way to make a given organization or product significantly more valuable in the long term.
86If you presume a given model describes reality you'll structurally pay less attention to the things that don't fit the model.
  • If you presume a given model describes reality you'll structurally pay less attention to the things that don't fit the model.
    • That's precisely where the value and insight comes from!
    • The surprisal is the gradient of learning.
    • All models must be wrong–they need to distill complex reality by reducing dimensionality.
    • But some models are useful.
87One of the characteristics of good PMs is that they don't get embarrassed by being surprised.
  • One of the characteristics of good PMs is that they don't get embarrassed by being surprised.
    • Instead of feeling like they're caught off guard, they absorb the surprising fact with grace and curiosity.
      • Cosmically calm.
      • Like it was always meant to be this way, and now that they know this fact they can be even better balanced and on a path to success than they were before.
    • Not shying away from disconfirming evidence, leaning in with grace.
88Mentorship gives you an order of magnitude more knowhow than reading a book.
  • Mentorship gives you an order of magnitude more knowhow than reading a book.
    • First, because they tell you the real situation, unfiltered, not something they have to simplify in a book or sanitize to avoid angering the other people.
    • Second, because it's not a thing you're reading about after the fact where all you can be is passive.
      • You're in the arena (indirectly) in the loop, trying to help solve the puzzle.
    • Real world experience is still way richer… but there's only so many hours in the day.
    • Mentorship scales more easily and gives you more leverage to increase knowhow acquisition.
    • Win-win!
89It's impossible to be accurately self-critical.
  • It's impossible to be accurately self-critical.
    • Your ego will defend anything that feels dangerous.
    • You need criticism from outside.
    • That you are forced to hear, either because
      • 1) you trust the source has your best interest in mind, or
      • 2) it's inescapable and impossible to ignore
        • e.g. feedback from someone who outranks you, who says it in front of others.
90S-curves look like exponentials… until you run into a constraint.
  • S-curves look like exponentials… until you run into a constraint.
    • Every exponential runs into a constraint.
    • Until you run into it, it's hard to imagine what it will be.
    • There's always a bottleneck in every system, because what counts as a bottleneck is not absolute but relative.
    • The bottleneck was always there, it just wasn't limiting anything until a certain scale.
91Emergent phenomena (like power dynamics) are never primary.
  • Emergent phenomena (like power dynamics) are never primary.
    • They're always secondary.
    • Abstract, not concrete.
    • They are impossible to show concretely to people.
    • Even when they're staring someone in the face they might miss them.
92Humor can be a release valve in a system that prevents any one person from becoming too powerful.
  • Humor can be a release valve in a system that prevents any one person from becoming too powerful.
    • Apparently in The Dawn of Everything it talks about how in various indigenous societies, when a chief had to emerge to coordinate behavior, everyone would ridicule and mock them to keep them from getting too powerful.
93Schismogenesis is a kind of creative opposition.
  • Schismogenesis is a kind of creative opposition.
    • You look at something your neighbor does and oppose it.
      • "Neighbor" might be in time, as in "the generation before mine."
    • Without that opposition, there's no way for something distinct to bud off.
    • For something to break off there has to be a negative boundary gradient for it to pinch off into a new thing.
    • There has to be a schism or else it just gets sucked back into the main thing.
94Progress across the millenia is an emergent social process, more than the outcome of any individual genius.
  • Progress across the millenia is an emergent social process, more than the outcome of any individual genius.
95"Taking a vig" is about "taking a cut" when it was coerced or illegitimate.
  • "Taking a vig" is about "taking a cut" when it was coerced or illegitimate.
    • When it's predatory, not consensual.
96Sarumans believe the ends justify the means.
  • Sarumans believe the ends justify the means.
    • That is, you can do immoral actions in pursuit of moral outcomes.
    • Radagasts believe that you need to find moves with resonance: moral actions and moral outcomes.
97Systemic problems require systemic solutions.
  • Systemic problems require systemic solutions.
98One way to make things resonant is to take things you know are good for you in the long-term, and then work to find the joy in them in the short-term too.
  • One way to make things resonant is to take things you know are good for you in the long-term, and then work to find the joy in them in the short-term too.
    • So it's not like eating your vegetables, where you dislike it in the short-term but like it in the long-term, but it's something you enjoy in the short-term and also in the long-term.
    • For example, for workouts, find the power in the physical meditation to ground you in the moment.
99Meaning comes from tension.
  • Meaning comes from tension.
    • "Meaning comes from cost" is a sub-insight of this.
100In practice, humans don't care about truth, we care about the elimination of uncertainty.
  • In practice, humans don't care about truth, we care about the elimination of uncertainty.
    • The definition of being human is to be insecure.
    • Full enlightenment by everyone would lead to nothing ever happening.
101Every idea that became inevitable started off wrong.
  • Every idea that became inevitable started off wrong.
    • Then again, the vast majority of wrong ideas never become inevitable.
102Craving vs calling?
  • Craving vs calling?
    • Revealed preferences are more likely a craving.
    • Resonant things are more likely a calling.
103All resonant things are enjoyable on the surface.
  • All resonant things are enjoyable on the surface.
    • But most of the things that are enjoyable on the surface are hollow.
104Nourishing things are not just a source of calories, they're also healthy.
  • Nourishing things are not just a source of calories, they're also healthy.
105Hollow things are fragile.
  • Hollow things are fragile.
    • They shatter under stress.
    • They are not resilient or adaptable.
106"Human dignity" means treating people as an end in and of themselves.
  • "Human dignity" means treating people as an end in and of themselves.
107Resonant computing is dignified.
  • Resonant computing is dignified.
108Hollow computing leaves us feeling regret.
  • Hollow computing leaves us feeling regret.
    • Resonant computing leaves us feeling nourished.
    • LLMs increase the leverage of tech so much that the difference is more important than ever.
109Resonant computing is technology that invites you to participate in life, rather than a substitute for it.
  • Resonant computing is technology that invites you to participate in life, rather than a substitute for it.
110Resonance is when want and need are aligned.
  • Resonance is when want and need are aligned.