Bits and Bobs 6/8/26

1Herding cats is an inconvenience.
  • Herding cheetahs is an impossibility.
  • Coding agents make engineers into cheetahs.
2Code is a liability.
  • It's always been a liability, but now it's obviously a liability.
  • Before, code was so precious and expensive to create that it felt intrinsically valuable, which obscured it being a liability.
  • Now that code is cheap, the true nature is revealed.
  • Code is debt.
  • It will bite you.
3Human judgment will always need to be marbled into agent swarms.
  • The human judgment grounds the swarm, making sure the output has meaning.
  • The leverage factor (agent to human ratio) might go up and change, but the human component will always be there.
4Anyone can add value if you throw a profligate, unsustainable number of tokens at the problem.
  • The question is what you can do with a reasonable number of tokens.
5A tweet making an interesting observation: "No one really claims that alphafold is conscious, or that sora or midjourney or dall-e are conscious."
  • LLMs feel more conscious because they can talk to us in our language.
6AI allows making surprising, delightful moments of craft.
  • Allowing craft and personalization that previously would have been insanely expensive and not viable.
    • For example, subtle cinemagraph animation to enchant the visual assets.
  • LLMs make it orders of magnitude cheaper, but it still has a marginal cost.
  • One way to do it is to have it feel open-ended and custom to the user… but subtly direct users to the same paths.
  • Then, you can generate one hifi, delightful experience or video per path and make sure to cache it and not regenerate for future users.
  • Users still feel the wow of an open-ended, hifi experience… but the cost of generating that content is logarithmic.
  • The users are helping pave a cowpath of hifi, magical content.
  • A resonant illusion that doesn't need to break the bank.
7"Craft" doesn't mean "didn't use AI."
  • It means "what did the creator bring to this creation process that's better than what a generic creator could have done."
8LLMs can make a superficially compelling argument even on absurd premises.
  • If you ask it to "Give me a mathematical proof that dragons exist," it will generate something that looks superficially compelling… but is total nonsense.
  • The trappings of quality have decohered from the true underlying quality.
9"Proof of craft" used to correlate with the trappings of production value.
  • Something that looked high-quality probably was high quality.
    • Producing that high quality is expensive.
    • The only creators who would get the capital necessary would have had to have a proven track record for the financiers to take a bet on them.
  • Things that look high-budget might now be low-quality.
  • Creators can look like they've paid their dues without actually paying their dues.
  • Similarly, vocabulary and writing skill used to be a marker of quality (a proof of intelligence).
10A review of Gemini Spark from The Verge:
  • "On the other hand, I can't shake the deeply creepy feeling I get from the whole thing.
  • What Spark did feels sort of magical, and very invasive.
  • It's weird that Spark is so casually telling me the names and ages of my children, reminding me that it knows where I live, and finding information I know for a fact I've never volunteered to Google.
  • Intellectually, I know that Google knows an incredible amount about me — add up my emails, my calendar, my photos, and my search history, and you've pretty much got me pegged.
  • But seeing Spark treat all that data not as something to be protected, but as something to be mined, just feels bad.
  • This is the trade we're all being asked to make right now.
  • There is a direct correlation between how much of yourself you're willing to share with an AI system and how useful that system can be."
11Maximizing AI is about controlling it.
  • Making sure it works for you.
  • AI isn't going away.
  • Other people and companies are going to use it, possibly in ways that affect you that you won't like.
  • You should use AI to extend your agency.
12Today the default policies of software are set by the company who made the software.
  • They have thought through this way more than you.
  • Their incentives also aren't aligned with yours.
    • They want to get as much data from you as they can get away with.
  • Also, even if you did think about the policies, there are no levers to tweak, it's take it or leave it.
  • As a user, you know you're doing it wrong by leaving all the doors and windows open and you should carefully set policies by picking which apps to use.
  • But ain't nobody got time for that, especially in a world so hostile to user control.
13A confidence man is able to make people feel a deep connection that is fundamentally hollow.
  • They take small facts about their target and weave them into a story that seems insightful, personal, and, most importantly, flattering.
  • The combination is irresistible.
  • LLMs can do this much better than the average person can.
14A tweet from the official OpenClaw account:
  • '"You can run OpenClaw inside your company now." Annoucing our work with @Microsoft to bring OpenClaw to the Microsoft and Windows ecosystems. Claws now work securly in the enterprise.'
  • There's something so perfect about the misspelling of "securely."
  • Reminds me of the quote from the Simpsons episode about the robots at Itchy and Scratchy land: "Nothing can possi-blye go wrong... possibly go wrong. Heh, that's the first thing that's ever gone wrong."
  • A self-undermining claim.
  • A casual assertion of something fundamentally and obviously untrue, where the act of asserting it is exactly what reveals the speaker isn't entitled to make it.
15Gemini Spark and OpenClaw is like having agents running around with razor blades.
  • Someone is going to get hurt.
  • A friend was using Gemini Spark, and it tried to move his birth certificate to a different folder in Drive.
  • The only reason he caught that that was happening was because it tried to move it to a folder with different permissions, so he got a notification.
16You can't call it "safe" if you hand users an armed grenade with the pin out!
17Headline: Company Blew $500M On Claude AI In One Month Due To No Usage Limit On Licenses For Employees.
18This week's Wild West Roundup:
19A signpost open source project: Chipotlai-max.
  • "The AI coding agent that steals Chipotle's support bot. Free inference paid for by burritos."
20Strava declares war on scrapers ahead of IPO.
  • Another early shot in the Silo Wars.
21A great new piece from Mike Masnick: Enshittification, Despotification, and the Open Internet.
  • Centralized systems get taken over not just by profit motives but also government control.
22When building a platform, how many degrees of freedom you give to userland content is a critical strategic question.
  • Everything you make customizable in userland is configuration space you've boxed yourself out of.
  • In the future, it's way harder to do platform-level features in those dimensions.
  • You either can't do it, or have to do really complex compatibility modes.
23A new piece from Ben Hunt: Contact: AI and the Semantic Dimension.
  • Makes the case that there is a semantic dimension that LLMs are able to surf through.
  • A dimension that is in a separate plane from the spatial dimensions.
  • Before, we could sense it, but only indirectly.
  • Now we have an alien intelligence that fundamentally lives in that dimension that we can experiment with.
24Gresham's law: bad money drives good money out of circulation.
  • Good ideas are held onto and carefully doled out since they confer an edge.
  • Bad ideas are circulated freely.
  • That means that all else equal, you'll encounter more bad ideas than good in open channels.
  • If you get something via email it's likely to be crap.
  • The world becomes more cacophonous.
  • "When everyone has a substack, no one has a substack."
  • This means that trusted recommendations are more important than ever before.
    • The "meat filter": if an actual person I know and trust recommended it, then I'll read it.
25To get more quality in a monolithic model, there's no move except "shovel more GPUs and ML engineers into it."
  • There's one monolithic thing at a low pace layer that's insanely capital hungry.
  • Other systems allow more incremental investment to unlock value.
26A two loop optimizer.
  • The inner loop optimizes the goal the best it can, climbing the hill it's on.
  • The outer loop judges the result and if it's good enough or not.
  • If not, it jumps the inner loop to a new hill.
  • This process doesn't work with a single loop.
27Your Personal Knowledge Graph is more important than any silo.
  • It's where your data canonically lives, where it can mix with all other data.
  • You need to be able to inspect your PKG, to be able to see what it models, where it's wrong, etc.
  • It should not be hidden.
28The web at the beginning of its mainstream era felt more like print than software.
  • Design felt more like fixed layouts you might see in magazines or newspapers.
  • The people who were responsible for sites were graphic designers or librarians.
  • The web page was software, but only incidentally.
  • The idea that your website needs a Product Manager felt almost absurd.
29A default-converging asymmetry for a codebase: "Every time you touch an area of code, leave it a bit better than you found it."
  • The "leave it better than you found it" is in some ways a cheat because the definition of "better" is left undefined.
  • But as long as there's a consistent asymmetry in what people think 'better' means then there will be coherent movement in a good direction.
  • Even if people disagree on some aspects, there are others that everyone can agree on: tidy, robust, elegant, well-documented.
  • In many situations, people mostly agree on the general direction of better.
    • It's only in a small number of decisions where people disagree… though those get all of the mental attention in political environments.
  • Rob Sand's slogan: "Not redder or bluer, but better and truer."
30The first 90% of software is technical, the second 90% is human.
  • That's why the second 90% takes so much longer.
  • The human element is the ground truthing of "making this something that users in the real world will accept."
  • Not just a good concept or demo, but something useful.
31A gate of a compelling demo only works if the output doesn't change for different inputs.
  • If everyone gets the same experience, then a compelling demo tells you it will work for most users.
  • But if the outputs vary widely for different inputs, and there are many different inputs, then the demo approach doesn't work.
  • The demo might just show a happy path, but what matters is how it responds to actual inputs.
  • Instead, you need a hill-climbing approach.
32Magic jello is still fundamentally jello.
  • It's not software engineering, it just looks like it.
  • If you put small boxes around the magic jello, you close off the potential of future models improving quality.
    • Locking in a how locks in the current model quality.
    • You've capped downside, but also capped upside of better models.
  • In some cases, more quality in output doesn't matter, and capping downside, and doing it cheaply, is more important than incremental improvements in quality.
33A gauntlet is a process where motivated users self-select.
  • Getting through the gauntlet is hard, so only the most motivated users do it.
  • This can be used as a product roll-out strategy.
  • It can also be used to minimize harm from dangerous features.
    • For example requiring someone to run a command from the terminal to turn on a dangerous developer feature.
  • The fact that Bits and Bobs is a Google Doc is itself a form of gauntlet.
  • The gauntlet tests two things:
    • 1) Does the user have the capability to make it through the gauntlet?
    • 2) Does the user care about what's on the other side of the gauntlet?
  • As the world gets more and more cacophonous, gauntlets as filtering mechanisms get more and more important.
  • The shorter the gauntlet gets, the lower the average quality of things on the other side of the gauntlet.
34If a group is a closed set it will become an echo chamber.
  • Perhaps slowly, perhaps quickly, but over sufficient time horizons it will become one.
  • Once it's an echo chamber, it's a zombie.
    • Dead, but doesn't know it yet.
  • The only antidote is to trickle in new perspectives regularly.
35Nerd clubs require an active moderator.
  • If you let a nerd-club run normally they overheat.
  • The very thing that makes them attractive leads to them breaking the engine that made them work.
  • The moderator rides the tradeoff between the inflow of novel perspectives, and the maintenance of productive norms and trust.
  • That tradeoff has to be actively surfed by someone with judgment and long-term orientation for the group.
36Every successful thing grows to the point where the property that made it successful breaks.
  • What is valuable is contextual; the mode can shift and as your thing gets successful it shifts modes around it.
37In scarcity, groups make sense because individuals can't deal with scarcity or risk.
  • In abundance, you get assholes because the group's protection is taken for granted.
  • When you succeed you get to abundance and now the group is the problem, not the wolves.
38As the world gets cacophonous, people retreat to cozy communities.
  • It requires moderation, otherwise it retreats to the mean and falls over.
  • You need active engagement to make it work and not have it just be overwhelming and cacophonous.
  • The more successful your community, the more that people who just want to troll on it come into it.
  • Another category is people who love the community too much, and become overbearing and insufferable.
    • They police it, and need to be in every conversation.
    • For example, the people who are focused on a particular lens and try to make everything about it, and don't have a platform elsewhere, but loving that community makes them want to share it there.
    • A hedgehog, in a secret group of "people who finally get me."
39Standardization is about scale vs arbitrage.
40Nuggets of wisdom from Warren Buffet.
  • "Time is the friend of the wonderful business, the enemy of the mediocre."
  • "It's far better to buy a wonderful company at a fair price than a fair company at a wonderful price."
  • The fundamental curve (is it super-linear? linear? sub-linear?) matters far more than the multiplier.
41Discontinuous value is created when something transitions from non-consensus to consensus.
  • If it stays non-consensus then no one agrees it's valuable.
  • If it's already consensus, there's no edge.
42Working on something non-consensus is a blessing and a curse.
  • It's a blessing because it has the potential to create discontinuous value.
  • But investors and participants have to believe that it will one day be consensus.
  • One way to do it: make a consensus-style product that adds obvious value.
    • "This thing everyone wants, but better."
  • Then, add a non-consensus bonus.
    • "Here's a bonus feature that's different in a way that changes the game."
  • You get people to see why it's plausible because it's not too out there, but then they see the benefit of the part that's different.
43Focusing on strategy is a tactic that only works if you can execute it.
  • That is, if you have the capability, the resources, and the time.
  • If you're a pre-PMF company the answer is "almost certainly not."
  • If you're a big company the answer is "perhaps yes."
44Compelling arguments feel inevitable.
  • Once you see the fundamental truth you can't unsee it.
45A product with PMF is naturally compounding.
  • Users like it well enough to recommend it.
  • A constant likelihood of recommending, multiplied by the current user base, gives compounding growth.
46When you are open to ideas from any direction, you must have judgment.
  • Otherwise you get buffeted around in ways that don't cohere.
  • Not even default-diverging, but worse: it reduces to just noise, brownian motion.
  • Especially important in cacophony.
47"Act like an owner" means flipping from asking for permission to asking forgiveness.
  • For change you think will improve the thing, flipping from a default-no to a default-yes.
48One reason my open office-hours work is because I'm not a VC.
  • If I had capital to allocate directly, it would presumably be swamped with people.
  • Because the "capital" I can deploy is all social and indirect, there's not a reason to meet with me unless you just want to have a conversation with me.
49Curiosity in a collaborative environment comes in two forms.
  • "Yes, and…" and "No, but…"
  • Curiosity is a self-sustaining energy that never runs out.
  • "Yes, and" affirms and lifts the collaborator up.
  • "No, but" drags the collaborator down.
  • The "No, but" curiosity personality is a form of Proactive Energy Vampire.
  • Not even boring, just infuriating and demotivating.
50The closest thing I know to a magic spell: "Yes, and…"
  • It accepts what the collaborator is saying and then redirects it.
  • If you didn't accept, then the collaborator would feel stymied or slowed down.
  • By affirming, you lift your collaborator up.
  • The "and" is extremely important: you chose which parts to build on, applying judgment and steering.
51Apparently property crime on BART is down 95% after they installed the tall fare gates that make it impossible to jump them.
  • Tiny amounts of friction can deter a long tail of low-intent troublemakers.
52Two generative questions when designing a system:
  • 1) Should these two things be one?
  • 2) Should this one thing be two?
53A composable system can't be infinitely flexible or the pieces won't compose.
  • The format of composition is critical to standardize.
54People don't have values, values have people.
55An important invariant to maintain: never lie to your kid.
  • That includes making sure you follow through on threats!
56You can't protect a boundary that you can't sense.
57I heard about someone who accidentally took a 5x dose of Ozempic.
  • For weeks, they didn't have any desire at all.
  • The desire to even drink water went away, making them get dangerously dehydrated.
58A tone for getting subversive arguments read: safe-subversive.
  • Give the writing charismatic, puck-ish jokes so it's fun to read.
  • Then, in the text connect nine of ten dots to a fundamentally heretical take.
  • Readers might connect that last dot themselves, but since you didn't connect it for them, you're less at risk.
  • But because it's a charismatic, enjoyable piece of writing, lots of people read it, so the likelihood that a critical mass of people connect the dots is much higher.
59Politics start as soon as there are two people and one blanket.
  • That is: multiple people and a scarce resource.
  • There's always scarcity in something.
  • Power, for example, is inherently relative, and thus inherently zero-sum, and thus scarce.
60Why do you have to make a hypothesis before experimenting?
  • Because the hypothesis requires a model of the system.
  • The point of the experiment is to confirm or disconfirm the model.
  • Seeing what happens is not the main payload of the experiment.
  • Seeing if your world model is supported or refuted is.
61Sherlock Holmes always derived huge conclusions out of tiny observations that could have been wrong.
  • Completely unrealistic in any real world situation.
62The process of a new thing differentiating itself from the old thing is called schismogenesis.
  • As the new thing buds off from the old, it goes out of its way to differentiate itself, to reject characteristics of what it came from and put distance between them, to individuate.
  • This process is how new art forms emerge, but also how teenagers rebel from their parents.
63In English "this is actually good" means "I thought it wouldn't be good but it is."
  • In German it means "this is really good!"
  • Very different valence for the receiver!
64Steve Job's "put a dent in the universe" is evocative and powerful...
  • Steve Job's "put a dent in the universe" is evocative and powerful... but also there's no valence to the dent.
    • Do other people think the dent is good?
      • Does it help society or hurt it?
    • There are lots more ways to put a negative dent in society than to put a positive one.
      • That's just entropy.
65Looking backward is mainly useful to the extent it helps you change what you do going forward.
66When you have momentum towards a goal you believe in, it gives you limitless energy.
  • If you're missing either, it takes energy to keep going.
  • When you have both, you're buffeted along, and it feels like flying.