Bits and Bobs 2/23/26
Alex Komoroske
OpenClaw has captured significantly more imagination than usage.
And it’s captured a significant amount of usage!
But still, the audience of people who are intrigued but intimidated is 10x the size of people who are actually using it today.
Claude Code's PMF is orders of magnitude stronger than ChatGPT's.
Despite ChatGPT’s being much larger on an absolute basis.
Every bit of Claude Code usage is paid.
Perhaps with heavy subsidies for top users in Max plans.
Most of ChatGPT’s usage is given away for free.
When I start tasks now, I often create a new local repo, git init, and run Claude.
I find I do this for any task that even possibly might accumulate some code or data.
A Claude web app session is about the chat primarily.
A Claude Code session is about the durable files and code that accrete.
Which is more important, the chat or the data / code it creates?
I only need to pay Anthropic when I need to make changes to the local files.
That data and code in that project work even if I stopped paying Anthropic entirely.
They are mine.
Contrast that with Claude web chats, which are owned by Anthropic.
The codebase matters less than the process to create it.
Just like previously, the compiled binary mattered less than the source code to create it.
Now, LLMs can translate specs and ideas into working code.
Some of the most valuable information is what the developer told the LLM in the process of creation in Claude Code.
Yet Claude Code deletes those transcripts after 30 days today.
Those chat logs should be treated with more respect than the resulting code!
We’re going to see a ton of security veneer in the next few months.
Tons of companies trying to take OpenClaw style experiences and make them “safe.”
For most people coding is a means.
For some it is an end.
Joyful creation.
Now with Vibecoding that joyful creation audience has grown 100x .
Because the amount of broken glass has gone down by an order of magnitude.
But even so, most people will never get joy from coding.
Software for most people will always be a means to an end.
Means should be as invisible and low friction as possible.
The vast majority of people don’t want to think about the features of their software... they just want them to work.
Software is a means to an end.
Each successful layer of abstraction gives you an order of magnitude more leverage.
“Succesful” layer of abstraction means a layer that you could peek under… but never need to.
It successfully abstracts over the complicated internals so you don’t need to worry about them.
Coding agents getting good enough that you don’t ever have to look at the code is one of the big unlocks for this new order of magnitude of productivity.
People don’t want a TEMU of apps.
Seemingly infinite apps… all of which suck.
People don’t want more apps.
They want software to do what they need help with without requiring them to think about it.
It should be so well aligned with what they’re trying to do that it just fades away.
This week’s Wild West roundup:
A Cline AI tool had a prompt injection attack that… installed OpenClaw on the user’s system.
There’s a large-scale poisoning attack in OpenClaw skills.
A security researcher took a perfunctory look at OpenClaw and found oodles of issues.
Meta and other tech firms put restrictions on OpenClaw for their employees.
There’s a real-world infostealer attempt targeting OpenClaw users.
All of that data and keys in one place is a tempting target!
This week’s “through the looking glass” roundup:
The operator of the AI agent that published the hit piece came forward.
If their account is to be believed, the incredible thing is how anodyne the prompt was to get such toxic behavior.
LLMs allow you to do more.
Are you doing to do MOAR, and hollow yourself out?
Or are you going to go deeper, and create more resonance?
Similar to "If thinking is 10x cheaper will you think 10x faster or 10x deeper?"
If you aren't using claude code you're being left behind.
The Atlantic: The post-chatbot era has begun.
Even non-tech outlets get that we’ve already transcended mere chatbots.
With agentic engineering, features that were once nice-to-haves are now just haves.
It’s so easy to add features, there’s no reason not to.
Jenny Wen makes the case that our current best practice design process is obsoleted by LLMs.
The process that was the best practice fundamentally assumes that software is extremely expensive to create.
Two very different kinds of beauty: the Gilded Turd vs the Grubby Truffle.
The Gilded Turd is superficially beautiful, but fundamentally disgusting.
The Grubby Truffle is superficially disgusting but fundamentally beautiful.
Even an infinite amount of polishing can’t make a Gilded Turd fundamentally beautiful.
Closed systems tend to be Gilded Turds, and open systems tend to be Grubby Truffles.
In the modern world, superficial appearances are all that people have time for.
We’re surrounded by Gilded Turds.
Hollow. Shiny. Empty. Gross.
The word app implies silo.
That makes them clean and easy to reason about.
But it also makes them fundamentally non-composable.
Disconnected islands of functionality.
Each island is its own thing.
Nothing greater can emerge out of the collective.
Remember: LLMs are able to execute text, which makes all text basically code.
Even if you trust the creators of the skills you’re using, any untrusted text from any source you’re working on can screw you.
App stores get nearly infinite power because there is no way around them.
But an app-store like thing for an open ecosystem (like npm) can’t get the same power.
It’s always technically optional.
It’s more just a convenient schelling point for trust to accumulate.
App stores on mobile OSes are load-bearing, not conveniences.
The ecosystem currently assumes an npm-style approach will work for getting good-enough security for skills.
But skills can evolve and apps can’t.
Skills are open-ended in that they are fuzzy and non-deterministic.
Skills are also vulnerable to prompt injection, fundamentally.
Even if you trust the skill creator, the data you import can attack you.
Npm-style approaches work best for finite software, where there is non-trivial overlap for long-term trust to accrue.
It also requires knowledgeable users to look at trust signals.
There’s been a vibe shift about AI’s ability to create.
NYTimes Opinion: The A.I. Disruption We’ve Been Waiting for Has Arrived.
The same author, a few months ago: The Argument for Letting AI Burn It All Down.
NYTimes: Dinner Is Being Recorded, Whether You Know It or Not.
Your privacy is being invaded by strangers and their decisions in the public spaces you’re in.
Cory Doctorow, Tim Wu, Ezra Klein: The Internet Feels Miserable ‘By Design.’
Rob Dodson: How I Built My Mobile Second Brain.
Apple could never build OpenClaw.
It requires emergence.
It requires messiness to the point of recklessness.
Imagine if your OS’s kernel could be prompt injected.
That should terrify you!
The kernel is the laws of physics.
It must operate correctly.
This is something I think Karpathy gets fundamentally wrong in his sketch of an AI-era OS.
Anything with this architecture cannot be made into a safe mass-market product.
It can at best be a gilded turd.
Data has power.
It's not just bits.
The bits mean something.
Abundant cognitive labor is now possible.
So make sure it's working for you, not against you!
An important concept is cognitive debt.
Cognitive debt is your inability to understand the system and how it works.
Cognitive debt doesn’t matter if the abstraction is non-leaky and you don’t need to know how it works.
The more you automate creating systems, the less you understand how they work.
Each incremental bit of work it does for you, the more remote it becomes.
When you take on too much cognitive debt, at some point you hit a wall where you’ve delegated so much you can’t even reason about it.
If you delegate too much, at some point the human becomes a copy-and-paste peon.
Or even just a meat appendage.
Memorably captured in the short scifi story about the Whispering Earring.
You learn when your brain is engaged.
Tutorials work better if they don’t introduce cognitive debt.
The tutorials in Halo hold your hand through every single step… the user doesn’t need to turn on their brain.
Compare to Minecraft, that drops you right in.
The only entity who can decide what features work with your data is some random company.
Wait what?
This only made sense when software was precious and artisanal.
A fascinating paper by some friends: Reasoning Models Generate Societies of Thought.
Some of the techniques that human systems have evolved because they create resiliently great ideas (e.g. having diverse teams) show up automatically in LLM architectures, too.
LLMs present a coherent, “it’s a person you can talk to” mental model that is easy to grasp and also fundamentally wrong.
LLMs work much more like a blackboard system.
Bottom up, emergent, unknowable.
The comforting mental model gives us a miscalibrated intuition for what they can do and what their failure modes are.
When you’re talking to LLMs you’re talking to an emergent system.
More like pond scum than a human.
Then again, you could say the same about human minds and the easy, comforting, and incorrect mental model implied by consciousness.
Tech people will play with a new product based on its promise.
But for the mass market it needs to work and work reliably.
If a system is open but so different from everything else, it is effectively closed.
It must be possible to adopt a new system across a gradient.
Vertical SaaS is a spreadsheet in a trenchcoat.
LLMs are the universal universal computer.
Code is data in a very particular shape.
Now it doesn't have to be in that particular of a shape.
LLMs can make sense of just about any data.
Even data with an intent to execute.
Software eats the world because it can be replicated ad nauseum.
Bits are non-rivalrous.
The magic to create this software is no longer held by a cabal of magicians.
Top-down approaches have logarithmic value for exponential cost.
Bottom up / emergent approaches have exponential value for logarithmic cost.
The former create value quickly but then hit a fundamental ceiling.
The latter take time to get going but then are unstoppable.
The most powerful processes in the world are Convergent Emergent processes.
Emergence (self-driving) plus convergent (coherent).
These processes get stronger the more they scale.
Folksonomy is a good example.
So is Wikipedia.
They are rare but common if you know where to look and how to harness them.
They cannot be created in a lab, they have to be grown.
Wisdom of the crowds only works in default-convergent contexts.
If it’s a default-divergent context, the random noise doesn’t cohere.
In a cacophonous environment, the selection pressure is for superficial optics.
No one has time to do the deeper check on fundamentals.
Token furnaces are where you burn tokens to produce value.
Sometimes the value is so large that even if you have to burn insane numbers of tokens it’s still worth it.
A 2026 burn about bad writing: “This would have been better if an LLM wrote it.”
It’s an insult… but it’s also often true.
LLMs are distinctly better than the average adult at writing.
Why do we have open source software, but few other industries do?
Mostly because software is data, data is bits, and bits are non-rivalrous.
You can have the bits, and I can too.
Atoms are inherently rivalrous, but bits are inherently non-rivalrous.
Open source is easy in a world of bits, and hard in a world of atoms.
Things that are easier to measure get more attention.
This is a fundamental, inescapable, core asymmetry.
It’s why optimization always wins.
This asymmetry leads to a fundamental overweight of short-term, direct effects.
As a system gets increasingly cacophonous you get more superficial takes.
That creates more cacophony.
A compounding loop.
Most Rust engineers don’t need to know how borrow checking works.
They just know it’s protecting them in a deep way.
If a system constantly cries wolf with security questions, users will just turn on YOLO mode.
Either explicitly, or implicitly by just hitting accept blindly on any dialog.
Ads should not happen at the level of the ISP.
Ads in an ecosystem are fine… healthy, even.
But not at the ISP level.
The ISP level should be about serving the bits without interference.
The pipes should be dumb, not “smart.”
When you enter polish mode, you hunker down.
You don’t add functionality to the product, you just add robustness and polish.
When it feels like you’re 80% of the way done, you’re actually 20% of the way done.
If your goal is to have a perfectly polished thing, you’ll spend most of the effort on polish.
You’ll lock in whatever fundamentals you had quickly gotten in place.
If the product is complex, it will take time to polish all of it.
That means that you lock yourself in place, and it could be up to a year to get it out to market.
If the market is moving quickly, by the time you launch, you’ll have a year-old product.
In a fast-moving environment, big-feature sets with high polish aren’t viable.
You’ll be late by the time you ship.
Either have very small feature sets, or low polish.
Where data accumulates is the center of mass in a system.
Both in terms of where the strategic power is and for bootstrapping a system.
Nearly every useful feature of Twitter started first in userspace before being formalized in the platform.
Hashtags.
@ mentions.
Retweets.
In the era of agentic AI, we feel the onus of orchestration.
We’re always the hold up now.
The AI swarm is always ready for our next judgment call.
The worst kind of red-queen race is an addictive one.
You get trapped in it to start because it’s superficially enjoyable and addictive.
But now you’re in a red queen race that you can’t opt out of, lest you fall behind.
"I like this superficially but also if I stop I will die."
A maximum loop that is impossible to get out of.
This is the default state of modern society.
One red queen race possibility for the age of agentic AI: a race to get individual superpowers.
Everyone gets an edge over their peers if they apply AI more effectively.
But then their peers compete too and they have to push even harder to regain their edge.
This could end up in a hobbesian hellscape.
A piece of software that it’s hard to imagine a company making: Candy Crush, but it only works when you’re offline.
The end-user might want it: “only let me play this addictive game when I’m on a flight,” or “only allow me to play this if I turn off the other useful parts of my phone, which gives me a nudge not to use it.”
If you build the software yourself, you can align it with your interests easily.
But if a company built the software, adding a feature of “it only works offline” makes no sense; they might as well get the incremental use from online, too.
Now with infinite software we can make our own software more easily.
Is the system optimizing for your goals or a corporation's goals?
Corporations only care about the parts of you that align with their business interests.
The walled garden isn't evil, it's that it must by construction optimize for its interests above yours.
The same origin paradigm is fundamentally about a provider enticing you with nice software so they can hold your data hostage.
Before it was hard to do things with your own data at scale, so it wasn’t a big deal that someone else held it hostage.
The misalignment of incentives also comes down to the physics problem of the same origin problem.
Because you must give your data to someone else to get value out of it.
Now it works for them more than it works for you.
The optimization ratchet is often turned against you.
Companies are very good at optimizing.
Their interests will dominate your interests, naturally.
Where does the data accrete?
That's the most important strategic dimension of any system.
Data is state is momentum.
Protocols can't be taken away.
Services can be.
A service operates on someone else’s turf.
They call the shots, and can take it away from you.
StrongDM and OpenClaw are downstream of where LLMs hit a new scaling threshold of agentic ability.
They were inevitable; the time had come.
They were at the right place at the right time to surf the wave no one had seen yet.
Someone pointed out that “dragon rider” sounds like “dragon chaser.”
The latter is apparently slang or someone addicted to a particular drug.
The connection feels useful; the dragon rider is extremely capable… and also can’t stop themselves.
Some business models use float-based financing.
This is one of Warren Buffet’s favorite tricks.
Two businesses that look superficially similar but differ in this key dimension will have radically different long-term outlooks.
The insight is that growth can be self-funding if your cash flow timing works in your favor.
For example, you collect payment before having to pay for the goods.
If your business has this shape, the larger you grow, the more leverage you get.
The value is duration multiplied by volume.
It works for insurance (float of years) and it works for payments (float of hours).
In case there were any doubts about the political influence of X’s feed algorithm, this paper in Nature should put them to rest.
Remember, the entity that controls what you see controls what you think.
This week I learned about Community Memory.
It was the first local bulletin board system, in Berkeley, in 1973.
How do you get AI empowering real people?
Instead of only empowering the smarty pants tech oligarchs who already have so much power?
The .ai domain will be this era's .com
Just kind of "duh,” the unremarkable default.
Modern society is all about scale.
And thus transactionalism.
Finite has dominated infinite.
Axios: Integrity's moment of peril.
An article about prediction markets as the apotheosis of modern transactionalism and lack of shame.
Another example of our over-optimized society.
You know it’s bad if even Axios is calling it out!
Modern society optimizes the humanity out of interactions.
I ordered a couple of TVs from Costco.com.
Our primary office is unit 222, but we wanted them dropped off in unit 224.
The delivery contractor told us that he was personally liable if he delivered to the wrong address.
Due to that he felt like he was put in an awkward position.
What he wanted to do for us as a human was dangerous for him as an employee.
Some MBA somewhere thought that making the employee have personal liability would align incentives better.
But it does so by taking already marginalized people and pushing them even harder into the meat grinder.
The person who made the policy doesn’t have to deal with the awkward, inhuman interactions in person.
Kurtzgesagt points out that self-regulating around modern food is just too hard.
You have the whole weight of capitalism’s optimization machinery to compete against.
It’s not possible.
What is the human-centric enablement of AI, vs an AI-centric enablement of humanity?
Evolution is an emergent algorithm for innovation.
It runs as fast as the substrate it's operating in.
It needs variation and some selection pressure.
The variation doesn’t need to be “creative".
It can be random noise, or systematically introduced variation.
First, get the loop to close, then get it to be tight.
Getting the loop to close is going from default-divergent to default-convergent.
The shift is an infinite difference that in the moment feels mundane.
Once it’s default-convergent, tightening is a simple matter of hill-climbing.
It’s great when you find a clunky workflow that’s worth doing.
It’s worth doing when it’s clunky, so that means it will definitely be worth doing when it’s less clunky.
Now, you just need to make it less clunky.
This is a core asymmetry.
Once you clear the bar of the loop closing and it being worthwhile, you’re in default-convergent.
From there, you can ski down hill.
The hard part is finding the thing worth doing even when it’s clunky.
Immature systems only become mature through load-bearing use.
It doesn't have to be perfect but it does have to work.
It’s easier to get convergence with a team of line cooks than chefs.
Line cooks are hyper competent but don’t have their own vision.
Chefs have their own vision and need to do their own thing.
Pre-PMF needs convergence, otherwise the entire enterprise diffuses to nothing.
Convergence-oriented people have to be in an environment they’re aligned with.
If they aren’t aligned with the environment, they’ll either
1) burn themselves out in frustration, or
2) randomize and tear apart the system they’re in.
Which one it is comes down to how powerful the person is in that environment.
When someone doesn’t believe in the goal of the collective they try to stick out in a new direction.
They bet their direction will be more valuable.
But a) it’s more fun to do your thing than the existing thing and
b) misaligned energy erodes the strength of the collective's pull.
The smarter someone is, the more charismatic, the more they decohere it.
Which is like death to the existing collective.
Sometimes it really is radically better and everyone is better off.
But if it’s not it just grinds down the team.
"Yes, and" is not "yes to everything.”
It's "yes" to the core thing you think could be great.
Sometimes that's the whole thing.
Sometimes that's an itty bitty part, so small you can barely see it.
But “yes, and” must have curation and discernment.
It works better when you have well-calibrated taste.
How well it works is tied to:
1) how long your time horizon is,
2) how cheap seeds are to plant for you, and
3) how calibrated your taste is.
"Yes, and" on the interesting subset is novelty maximizing.
Big teams for successful products don't go slow because they get lazy, it's because they're at a lower pace layer!
Things depend on them.
They have leverage.
If you go fast, you break real stuff.
When you get more leverage, you go slower.
That’s the fundamental tradeoff of leverage on your product, you must drift to a lower pace layer.
A powerful combination: enough breadth to see, enough depth to feel.
Intellectually charismatic people can sometimes dazzle their audience.
Their audience is convinced, but not because they understand, but because they are overwhelmed into intellectual acquiescence.
Understanding is fundamental; acquiescence is superficial.
That means that they have to be repeatedly reconvinced in a way they wouldn’t have to if they understood.
The technical term for when a tool merges with our mental model is mechanical sympathy.
Humans are really really good at tool use.
When a tool is predictable, it starts feeling like an extension of our body.
It melds into our mental model, seamlessly.
Only tools that are predictable can develop this.
Investors would rather fund a taco stand than a Mexican restaurant.
The taco stand can demonstrate traction more cheaply and then make follow-on investment much less risky.
Investors will always pull the entrepreneur back to the taco stand framing.
A taco stand can sometimes be a zombie, impossible to grow beyond some small ceiling.
But they’d rather box in a potentially great idea into a taco stand cul de sac and miss its major payoff, than to go all in on an expensive Mexican restaurant that fails.
People who know how to do abundance can maximize upside.
People who know how to manage scarcity can cap downside.
They’re different skillsets and orientations.
You could live in a library, or study in a dive bar.
But you’d have to fight the structure to do so.
The Saruman mindset sees itself as anti-authoritarian.
But actually it’s against others being authoritarian.
The Saruman worldview is powered by the complete and total absence of self doubt.
So when they themselves are the authoritarian, everyone should just recognize how great that is.
Sarumans have a slight edge over Radagasts.
We all love narratives with clear stories.
The Radagast worldview is indirect; harder to distill into a story.
The Saruman worldview is easier to distill into a story.
So everyone is always looking for "who was the person who made this whole thing happen this way?"
N personality types get exhausted by tedium and details.
The routine details don't have any joy in them.
Some details are routine.
Some are distinctive.
N personality types may care about the distinctive ones but cannot bring themselves to care about the routine ones.
Public recognition is an order of magnitude more motivating than private recognition.
Knowing that everyone knows that you're high-status is really important to us social monkeys.
Zero based thinking helps you not get stuck in your confirmation bias.
Otherwise, that bias is an asymmetry that keeps doing what you were doing, all else equal.
Zero-based thinking helps you navigate your underlying assumptions much more effectively.
When you use a thing that resonates with you, you don't just like it, you feel compelled to share it, because it feels like the world will benefit.
Someone who can move in additional dimensions will be inscrutable to the people confined to fewer dimensions.
The lower-dimensional entity will see the higher-dimensionality entity seem to appear and disappear at will, almost teleporting.
Each ply of thinking you can effectively do is another dimension.
The meta is never urgent, but it's where the leverage comes from.
Not just incrementally better, orders of magnitude better.
The meta is a form of being able to pop up an additional dimension.
Thinking multi-ply takes time.
If you do it constantly you’ll always be on the back foot, late to the game.
You’ll be beat by
1) the swarm of one ply thinkers, one of which lucks into the right move, or
2) the seasoned high judgment operator who has a honed one ply intuition to make the right move quickly.
Better is to be ready for when the right opportunity presents itself and then pounce faster than anyone else can realize the opportunity.
Most of the time you will look like a surfer dude lounging around, but in that moment you will strike like a viper.
Bruce Lee: “Water can flow or it can crash. Be water, my friend.”
When the house is on fire don't plan rose gardens.
The tension between past you and future you is directly demonstrated in the insight that "It's never urgent to plant a tree,"
The world is always short-term thinking all the time.
Everyone is trying to get an edge tactically on everyone.
Run their OODA loop just a bit faster.
A red queen race situation that we all lose.
Find the seeds of greatness and then grow them.
Greatness cannot be imposed.
It can only be grown.
Greatness doesn’t happen later.
If you aren’t great by now you won’t be great.
So that implies, find the seeds that are great, and focus on those.
If you look carefully there are seeds of greatness everywhere.
There’s a country song: Bless The Broken Road.
All of the previous hardships you’ve faced put you on the road you’re on.
All of the lucky breaks in this specific road can’t be separated from the earlier hardships you’ve experienced.
Would you rather not have had a given hardship if it meant not having this daughter of yours?
The specific things you cherish in your life are inseparable from the hardships on the path you’re on.