Bits and Bobs 10/20/25
Alex Komoroske
This is one of the illusions that happens with logarithmic-value-for-exponential-cost curves.
The initial ramp of value is so extreme that it feels like it will ramp up to infinity.
But actually it approaches an asymptote.
Apparently ChatGPT’s subscriber counts are stalling.
I hadn’t realized how much of ChatGPT usage is free users.
Only 5% (!!) of users are paying subscribers.
It’s easy to get momentum in a business by selling dollar bills for 90 cents.
That can go on for as long as you can get other people to give you dollar bills for IOUs.
For their business success, it’s imperative for them to get higher engagement / stickiness.
A polite word for ‘addiction’.
A crucial feature: businesses that use it can subsidize inference for users.
This is a classic aggregator play to lock in an ecosystem based on an early advantage.
The subsidy makes it increasingly impossible for other providers to compete.
The bet is that by having predatory pricing (aka “dumping”) they can force other businesses out and then corner the market.
They’re already doing this by their intense subsidy of inference on ChatGPT.
OpenAI is betting that their overall story and momentum is strong enough that they’ll be able to continue raising capital when others can’t.
But a stable equilibrium to me seems to be Anthropic and Google staying in the race indefinitely.
They have the capital and backing to stay in the game no matter what.
And of course, there’s the possibility of open models, especially out of China, catching up.
That is a very different end state equilibrium.
In that world, there’s not a duopoly but a triopoly.
There’s always one shorter leg of the stool who is willing to do things that the stronger players wouldn’t, e.g. keep frontier model access available via API.
A duopoly has a very different game theoretic equilibrium than a triopoly.
A duopoly is effectively like a monopoly.
Both competitors position their policies right next to each other.
Like the famous Hotelling’s Law where businesses on a line place next to each other.
The appearance of competition without actual variance.
But a triopoly is different, there’s always an odd man out.
That odd man out is incentivized to do different plays that the stronger players wouldn’t.
That changes the dynamic so the bigger players have to compete, too.
Triopolies are radically better equilibrium for the market.
Chatbots are like filming stage plays.
We're still waiting for the "montage" moment for LLMs.
My vote for the montage capability is LLMs being able to create software.
LLMs turn out to be insanely good at writing and executing actions.
Code, or just english descriptions.
In the past few months, as an industry our minds have continually been blown by how powerful it is.
We keep finding new ways to get even more out of this ability, ever more easily.
Claude Code has only been out for 8 months.
That’s kind of crazy to think about how much has changed since then!
This revolution is just getting started.
LLMs ability to create software is catastrophically powerful.
We keep on discovering new orders of magnitude more powerful techniques for getting more value out of it.
For example, Anthropic Skills, which Simon thinks is a huge deal.
MCP felt cool, but it had a low ceiling and was easy to overwhelm the context window.
Really what matters it’s the ability of LLMs to do tool calling.
More generally: to create software, to do things, whether in code, or with tool calling.
LLM’s ability to create software is like nuclear fission.
Catastrophically powerful.
The default manifestation of it is a nuclear bomb.
But if you can figure out how to harness it in a nuclear reactor you could create near limitless energy.
If you could figure out how to get the “catastrophic” downside part capped so it was safe for the mass market, you could change the world.
Claude Code is not about code, it's about anything your computer can do.
The ability for LLMs to create software, to do things.
Anthropic Skills is powerful for the same reason my old Code Sprouts project felt unreasonably powerful to me a few years ago.
Giving just a teensy bit of structure to the LLM (in that case, a Typescript schema for state), and allowing a hierarchy of english language instructions that the LLM could peek inside if it wanted but not be bothered with by default.
An example of efficiently unlocking the LLM’s ability to create software.
Anthropic keeps doing something simple and elegant…
That also causes traffic jams.
Impeccable first-order thinking.
Non-existent second-order thinking.
The faster the twitch of components, the faster the clock speed of the overall system.
But also fast twitch can’t get more leverage.
A swarm of shallow actions.
In lots of tech companies today, they use Slack, never email (too slow).
That means they can theoretically go faster to respond quickly to things that happen in the market... but also they are now default at that speed all the time, and unable to think deeply.
Vibe coded software is fragile and shallow.
It looks great but you can break it easily.
Claude thinks everything it outputs is brilliant.
You need an external ground truthing process.
Your curation and judgment.
There needs to be a quality control entity in the middle that isn't eager for your approval.
A sphincter position in the quality pipeline.
A bottleneck with taste.
When you create something with AI, you make something you think is great.
Everyone else thinks it sucks and rolls their eyes.
Developers think everyone else's code is horrible.
Claude makes this effect an order of magnitude worse.
Producing content way faster gives leverage to your taste.
The calibrated judgment becomes extremely important, instead of blindly accepting and making towers of slop.
Are you happily making things... that others are actually choosing to use?
That last bit is the key question.
Often the technology isn't the limiting factor in a pipeline.
The limiting factor often isn't what we think it is.
For example, making a movie faster is not about making more images.
It's about getting things in front of the director for approval.
A way to get leverage: focus on the meta-thing, not the thing.
The meta-thing is the system that generates the thing.
If you can get it to work well, then you improve not only the thing, but a whole class of things.
For example, if you have a compiler, optimizations improve everything it compiles.
Another example: in search quality, don’t check in a configuration file of synonyms, check in the process to generate that file based on an analysis of the querystream.
Now it can improve itself automatically and keep itself up to date.
Another example: never touch the code directly, touch the specs and LEARNINGS.md that you give to the LLM to generate the code.
The way to decentralize AI is to decentralize the applications.
The models will likely centralize, due to capital requirements.
But that doesn’t really matter if there are three or more high-quality ones that are easy for an application layer to swap between.
That’s why OpenAI is desperately vertically integrating and storing state, to avoid being a swappable component lower in the stack.
LLMs for software development are a foot gun and a rocket pack.
Catastrophically powerful.
Big companies can't enable YOLO mode in Claude Code, it's far too dangerous.
Small startups can try dangerous things because they don't have much to lose.
LLMs are catastrophically powerful for building software.
The benefit of this ability accrues to the entities that have lower downside risk.
The asymmetry is now much stronger than before.
That means startups as a class have an advantage.
Many startups that use it will blow themselves up.
But some will get lucky.
We’ll see swarms of “fast-fashion” startups and apps.
Everyone who stumbles across leverage thinks they're a genius until they die.
Leverage gives you speed for risk.
The risk is hard to see, the speed is easy to see.
You borrow from the future to go fast today.
If it works, it works great.
If it doesn’t work, it works terribly.
And perhaps knocks you out of the game.
Lots of things are levered in ways that aren’t obvious.
With LLMs, you don’t need to use a dependency, you can distill the equivalent on demand.
A dependency brings in complexity and risk from the rest of the ecosystem.
It’s also not fit to your specific purpose, but a general one.
One bonus: you can get security fixes for free.
LLMs are great at writing code on demand.
If there are lots of examples of a given library, you can have it distill a custom one on demand, just for you, perfectly fit to your purpose.
No dependency risk!
With LLM’s ability to execute and build programs, Innovation coins just got cheaper.
You can have more of them than before.
The difficulty of a programming task now comes down almost entirely to novelty.
It used to be that there was a difference between integration hard and algorithmically hard engineering.
Integration hard is easy to do, just requires a long, detail-oriented slog.
Can be parallelized relatively easily.
Algorithmically hard is hard to understand, but then easy to execute once you do.
Requires carefully reading papers, brainstorming at the whiteboard, going on long quiet walks.
But once you write the code it’s often 1000 or so lines.
Very difficult to parallelize.
But that distinction was pre-LLMs.
LLMs are great at algorithmically hard problems… as long as there are a lot of examples of it in the training set.
No matter how arcane those examples are to discover or reason about.
So the difficulty of executing now comes down entirely to novelty.
Less novelty: more likely the LLM’s first guess works and it can iterate its way to the solution.
A surprising LLM pattern Jesse Vincent discovered: GraphViz.
Graphviz, when rendered, makes sense to humans to understand flows.
LLMs can understand it based just on the markup.
A boundary object for flow graphs between LLMs and humans.
Git worktrees are a huge unlock with agents.
They used to be hard to understand how to use, but LLMs are great at them.
Giving agents isolation, so if they blow up something, they don’t blow up everything else, allows them to move fast.
A powerful piece of infrastructure will be containerization++.
Not just containerizing the code, but also the data and resources.
Imagine if you had a copy-on-write filesystem and data layer.
You keep on hearing about companies who are hiring new hires, and they are massively more productive than senior engineers just by YOLOing it.
Irresponsibility is the unlock.
A form of leverage.
Take shortcuts now that might blow up later.
YOLO!
Claude Code in YOLO mode is a radically better experience.
Without YOLO mode you need to baby sit it, constantly responding to its inane swarm of permission prompts.
But with YOLO mode you can set it and come back hours or days later and see how far it got.
But that opens you up to massive downside risk.
Poor quality, insecure code.
Deleting code or files.
Deleting data from production.
Getting prompt injected and divulging secrets to adversaries.
The irresponsibility is the unlock!
Systems that make it possible to unlock that power without being irresponsible will unlock huge amounts of power.
The limiter for LLMs’ ability to create software is the safety model.
In development, it’s about having a sandbox and isolation.
Containers, etc.
In deployment: not leaking secrets or having data loss issues.
A substrate that’s self-extinguishing jelly makes even dangerous software safe to run.
A kind of ballistic gel for untrusted software.
It’s malleable and can use it to provide scaffolding and structure–and most importantly, containment–around the explosives.
Assume the software is poor quality.
Explosively powerful but maybe reckless.
If you can make it default safe to run, you unlock that explosive capability.
Will AI adoption be like “the year of the Linux desktop?”
Every year for decades is supposed to be the year of the Linux desktop.
Linux has never become mainstream for desktop use.
But it quietly spread literally everywhere else in computing.
Maybe the change is happening, just not in the place you’re looking for it.
The AI submarine.
Bruce Schneier’s Agentic AI's OODA loop problem is worth a read!
In a world where code couldn't scale, software written by some stranger had to be good enough.
But now code can scale, and we can see that software written by a stranger was never good enough.
Software should feel like a personal garden that grows for you, not something some stranger constructed.
Imagine: a new distribution medium for software, where software can distribute itself and tailor itself to you, private and aligned with your interests.
If you had to have a global data structure for everyone, what would it look like?
Probably a reactive JSON graph.
You’d need links to reference other parts of the graph.
Things that are in memory like Redux state objects get that for free, but if it has to be serialized you need a formal reference capability.
A common fabric for computing.
Innovative LLM offering from Stripe: an API where an app can charge their users a consistent markup on the underlying tokens.
Karan Sharma, PM at OpenAI, is dreaming about AI Home Cooked Software.
Home-cooked software has to happen in a personal kitchen, not some factory kitchen owned by a corporation.
It’s long been possible to offload memory to external brains.
The people who could do so effectively were able to 10x their capacity.
But now you can offload thinking.
“Go think about this question and tell me what the options are.”
That's a step change.
Seems like a combination of:
the Skills / Learnings.md compounding loop
Crowd-sourcing
driving AI browsers.
A catastrophically powerful combination.
This kind of looks like RL if you squint.
RL researchers might say this is an under-powered hack to get something like RL.
But it’s different, because it has a swarm of human curation and judgment in the loop.
Distributed caching.
LLMs are great at reverse engineering software.
Reverse engineering software requires incredible patience.
LLMs have infinite patience.
LLMs help diffuse knowledge of a system faster.
To open a restaurant requires navigating a bureaucratic maze.
Talking to people who have done it before, scrutinizing overwhelming, poorly documented, kafkaesque processes that use arcane jargon.
It requires a knowledge of that jargon and infinite patience.
Something that LLMs have!
LLMs can help you navigate these kinds of processes more easily.
They effectively help metabolize arcane knowledge and allow people to operationalize it more easily.
Ideas that fit in one brain are an order of magnitude easier to execute.
Serializing intuition across the brain barrier is an extremely lossy and expensive process.
Once crossing the single-mind threshold you also start having coordination costs, which can balloon massively.
When new technology makes what previously were big ideas into minutiae, it frees up your brain to think about new big ideas.
Alfred North Whitehead: "Civilization advances by extending the number of important operations which we can perform without thinking of them."
LLMs make orders of magnitude more ideas fit inside one mind.
AI has a last mile problem.
For integrating our sensitive data, and also integrating them into our workflows.
LLMs are massively great at writing code, that will transform the world, and also... where are the results?
What is the right substrate to unlock its potential?
Integrating with the real world is much more difficult than coding an app
Rhymes with complex vs complicated.
A version of the last-mile problem.
LLMs are great at starting things, not at finishing things.
The human needs to poke them and structure them to do the finishing steps.
"The first 90% is done, now it's time for the remaining 90%."
You get the experience of going really fast without necessarily getting that much closer to the end.
A version of the last-mile problem.
Are agents bad at finishing real-world tasks due to a lack of metis?
We aren't seeing a Cambrian explosion of apps because everyone has apps 90% of done.
One of the benefits of having software experience is instincts about how long things take.
Some things are way easier than before, some aren't.
A jagged frontier.
Getting from idea to demo is now super fast.
Going from demo to production is just as hard as it once was, if not worse.
An external analyst thinks that vibe-coding traffic is falling off a cliff.
For example, a 50% decline for Lovable from June to September.
Of course, external data is of very poor quality.
But it does track to me as being plausible.
It turns out that as a non-technical person you can’t ship a production product even if the LLM writes the code for you.
There’s more to shipping a production app than writing the code.
The clean up from “demoable” to “usable” (especially to make it not just usable but also safe) is a huge amount of work, that LLMs don’t do a great job at unless you tell them to.
You need to know to tell them to.
Many of the vibe coded apps that succeed in the market get taken down by security issues.
The Lovable founder responded with stats showing continued growth.
But he did it in the most eye-roll-y, least-convincing way ever.
No y-axis numbers or even describing what metric it’s charting.
That’s an extremely easy signal to make misleading.
Every PM worth their salt knows how to cherrypick data to give the appearance of momentum.
You make the strongest case you can with the data you have.
A weak case implies you don’t have data that tells the story you want to tell.
I didn’t give the rumor that much credence until seeing that weak-sauce retort.
A new pattern from prominent open source contributors: have a different GitHub account for stuff you’ve vibecoded.
Those prominent contributors have a brand of significant quality for code he’s hand-written.
Instead of muddying that brand, he has a separate one for things he vibecoded and are thus more “use at your own risk”.
Verification is the bottleneck for LLMs.
Verifiers of taste, or of quality, or correctness.
This is where humans typically still need to be in the loop.
With variation, some systems converge and some diverge.
If they diverge, you get a rat’s nest.
The variation compounds and becomes mutually inscrutable.
If they converge, you get a boring, over-saturated middle.
LLMs converge their output.
So now the middle of the distribution of output will be over-saturated.
That will push out the differentiation to the tails.
Hyper-niche.
Hyper-scale.
This was already happening before LLMs, due to the zero cost of content distribution.
But LLMs turbocharge it.
There is structurally less qualitative user research than there should be.
It’s extremely expensive and manual today.
But they are much higher signal than quantitative user research.
Quantitative research requires you to ask just the right questions in just the right way.
If you ask them wrong, you’ll generate faux insights.
Harder to find some classes of disconfirming evidence.
You can’t discover your unknown unknowns.
You can’t learn unexpected things as easily.
But LLMs can do qualitative nuance at quantitative scale.
"Slop → Aura
innovative → institutional
personalized → individual
generative → creative
dissociative → enigmatic
monetizable → valuable
platform distribution → dark social distribution
mechanical → artisanal
scalable → singular
disposable → canonical
dead internet theory → dark forest theory"
I attended a talk where the head of one of the major labs mostly talked about AGI and what society will be like.
I wonder: what if some of this AGI talk is in some ways a long con?
First and most obviously, the more that people believe it, the more they can attract insane amounts of capital.
Secondly, this massive, society-defining outcome overshadows the more mundane, everyday concerns of power concentration in a hyper-aggregator.
Everyone’s fretting about this infinite outcome that might never come, instead of the almost-certainly-will-happen hyper-aggregation and power centralized in one company.
It’s not that people should care more about privacy.
It’s that they shouldn’t have to care about it all because what happens by default is aligned with their expectations and interests.
GDPR closes the barn door after the horse has already bolted.
It’s a post-hoc fix to the medium of software distribution having nothing structurally to say about privacy.
It’s not even closing the barn door, it’s describing how you might close the barn door, and prescribing onerous and unrealistic ways of doing it.
The goal of GDPR is a good one, but it’s impossible to retrofit satisfactorily onto the same origin paradigm.
My summary of Helen Nissembaum’s notion of Contextual Integrity: data use aligned with the user's expectations and interests.
Nobody reads EULAs or even could.
Most are much worse.
A cool paper from a few years ago about the kinds of cool things you can do with confidential compute.
Remote attestation is a massively powerful capability that the industry is sleeping on.
Cory Doctorow asks if AI assistants can escape the enshittification trap.
The answer is no, in my opinion.
Going down the path of chatbots / LLM-as-friend / hyper-scale leads to a path to inevitable enshittification.
A paper: "Sycophantic AI Decreases Prosocial Intentions and Promotes Dependence"
Just in case there was any doubt that the sycosocial relationships inherent to chatbots are bad for you.
The Atlantic points out that ChatGPT is a fictional character.
But unlike many characters, it was not authored.
Instead, it emerged.
A very different dynamic.
Yet another way that the implicit frame of Chatbots as “LLMs as friends” is unsettling and potentially dangerous.
A few thoughts on ChatGPT allowing adult content.
It leans into hyper addiction and engagement even more.
At that scale of deployment (beyond some niche for things like Replika) it could destabilize society even more.
This seems like a cynical growth play... not something you do if you had runaway momentum.
Also, do you really want this company to have records of all of your intimate interactions?
A milestone in game engine development: rendering the teapot.
The first milestone is “first triangle.”
A single triangle rendered on screen.
The next milestone is “first teapot.”
Render a 3D object, lit and shaded.
When you render the teapot, people see the teapot.
But the demo is that you made the system to render the teapot, not the teapot!
They understand what they see, not how it works.
Making a computing system that tracks dataflows palatable to normal developers is a massive accomplishment.
If you can make it feel “basically like React,” that would be huge.
Under the covers you’d have to be able to chart specific data flows with precision and accuracy and have taint checking that mostly doesn’t bother you unless it’s important.
But if you accomplished it and showed it to people, for any given program people would say “yeah it looks like React, what’s the big deal?”
It’s only the emergent ecosystem of software that can now be safely composed where the value becomes obvious.
A catalyzing accomplishment that superficially looks like nothing.
A general rule in optimization: auto-tightening systems.
Invest effort to optimize it in proportion to how often it’s used.
For example, in V8, the first pass is a quick-and-dirty compile.
But hot code paths (for example, in a loop) get another pass of optimization to make them faster.
This insight comes from the original HotSpot Java compiler.
Optimize it from generalized/sloppy to specific/tight in proportion to how many times it runs.
The same is true for workflows that use LLMs.
If it’s going to run once, just have an LLM execute an english language prompt.
But if you’re going to run it thousands of times, have the LLM write deterministic code.
Right now we don’t have enough real-world last-mile use of AI workflows so we’re in the cheap/sloppy phase of deployment, not yet to the auto-tightening parts.
Workflows are hard to change because there's an intra and inter network effect.
Inter: You need to coordinate with others.
If you change the workflow but others participate in it and don’t change, then you can’t change it.
Intra: each step is dependent on the others and possibly needs to change.
Changes to a workflow that don’t change the shape of inputs and outputs are a limited subset of the changes you can make.
Stratchery dove into why Walmart decided to integrate with ChatGPT Commerce.
To me this implies that the current model of software is stagnant.
There’s the personal OODA loop and the situations’ OODA loop.
The personal OODA loop is the classic Douglas Adams quote:
“1. Anything that is in the world when you’re born is normal and ordinary and is just a natural part of the way the world works.
2. Anything that’s invented between when you’re fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it.
3. Anything invented after you’re thirty-five is against the natural order of things.”
But sometimes the situation really does have a faster OODA loop than before.
The ability fo LLMs to write software is truly a fundamental speed limit that has changed.
If it’s not you… is it the singularity?
What would the singularity feel like?
"I can absorb this new change once there's a breather" …but the breather never comes.
Ben Mathes: "My differentiated skill is I read everything Ben Thompson writes… and actually understand him."
The Geek Fallacy: anything that cannot be understood via STEM frameworks is unknowable or unimportant.
Geek machismo leads to a kind of Robert Moses effect with much more leverage.
Putting on blinders and steamrolling the world.
The Sarumans are all about the head, never the heart.
They think they're being hyper rational.
In reality they've hollowed themselves out.
The heart is where the meaning comes from.
The intuition for indirect effects.
To be in harmony with the world you need both.
Sarumans hate bureaucracy.
But so do Radagasts.
Although they’re less likely to use the word “hate.”
Bureaucracy is about the status quo and downside capping.
Innovation is about upside.
Both the Saruman and Radagast magic are about innovation and upside.
The lack of magic, the dull, dreary, mundane company man, is about not innovation.
These are the banal "organization kids".
"Turn the crank, don't think about what the crank is connected to, that would distract you from the only thing that matters: number go up!"
Sarumans don’t care what other people think.
A lack of conflict-aversion as a super power to dominate over others.
No shame.
Finance and titans of industry typically have this power.
Also, Karens.
It’s an inherently antisocial power.
Simply don’t care what other people think.
In the hyper era, we went from meaning to MOAR.
In a world with extremely low friction the energy flows to the winners on the top.
When there's more friction it can create different pockets of smaller winners.
More variation leads to more diversity; more adaptability and resilience in the ecosystem.
One empire can spread throughout a vast plain; mountainous terrain often has fractal chiefdoms.
Somewhat surprisingly, low friction leads to hyper optimization leads to centralization leads to hollowness and fragility of the system.
The system is overfit.
Overfitness feels great in the moment.
Whether it’s overfit is whether you have a model that doesn’t generalize in time to new scenarios since it fits to what is temporary noise.
Overfit structures are 'fit' but not in a robust way.
Even if you were exactly right about prediction for today you’d be wrong tomorrow.
The world changes in ways you can’t predict, fundamentally
This riff and the next 10 were based on a talk I attended this week by Emmet Shear.
The modern world is overfit.
It’s overly optimized.
Modernity is about the systematizing of accuracy.
You create a meta model of the world, and then every process relentlessly optimizes it.
If there’s no slack in the system you are definitionally overfit.
In Seeing like a State, it’s not that they’re doing a bad job at modeling the forest.
It’s that it’s impossible.
You can't take a photograph of Métis.
There’s a tension between complexity of your model and accuracy.
The more complexity you add, the more you overfit.
You explain what you see well, but also over-explain noise.
You’re now overfit, poorly fit to novel inputs.
If you update your model by n bits it had better give you more than n bits of accuracy or you’re falling behind.
Ambiguity is that which you don’t have a model of.
The unknown unknowns.
The best you can say is “I’m still alive so whatever I did in the past must have worked…”
In the modern hyper-connected world, connection is free, sparsity is precious now.
The more connected, the more overfit the system gets.
Overfitness is a form of mania.
By every metric things look great.
But everyone can tell something is off.
It’s hollow, hyper.
Manic trips end in one of two ways: you calm down or the universe calms you down.
Modern society assumes that global connection is as good as local value.
But that’s not true.
Local connections are healthier and more nourishing.
They allow more diversity, variance, and adaptive capacity of the system.
Don’t follow high frequency information sources.
Talk to your friends and stay local, it’s good for you and good for society.
Gradient clipping is a way of forcing regularization.
In ML, one way of regularizing is keeping track of, for each weight, how precise you believe it to be.
That is, how often it has changed in the past.
When you need to update it, you update precise values less than imprecise values.
But this requires significantly more complexity.
Another approach is simply gradient clipping.
Simply cut off extreme values.
It’s less precise individually, but on average, stochastically it is apparently equivalent.
A “modest proposal” Emmet offered as a thought experiment, that I found thought provoking: have a max speed limit of information of 90 mph.
If it goes faster, it gets taxed, at a rate that goes up quadratically.
The main knob to control is the tax rate.
It could be an infinitesimally small tax rate, if you wanted.
This would naturally tilt information to local sources.
You could use the Global TikTok for a fee, or the Bay Area TikTok content for free.
A consistent pressure towards local connection, which is healthier for the system.
Downward pressure on inequality.
Like regularization, it forces it into flatter distribution.
Fewer billionaires, but more centi-millionaires.
Fewer centi-millionaires, but more dece-millionaires.
Fewer deca-millionaires, but more millionaires.
A healthier, more balanced system.
You could argue that a system of tariffs (consistently and thoughtfully applied) is a form of gradient clipping in this context.
It gives you a rough version of this policy, in the trade domain.
Obviously, an unworkable proposal, but still a fascinating thought experiment.
When everyone is in the same information stream everything becomes boring.
No variance to select over!
Global connection makes the world boring and wildly unequal.
One way to not get stuck in fast-twitch information streams: get your news from newspapers.
A day after it happens, not as-it-happens.
That gives some time for synthesis and distillation.
It also gives some time for perspective: what was the stack rank of most important things that happened yesterday?
Online sources can have infinite content, updated infinitely quickly.
But newspapers have scarce space and have a built in time delay.
They naturally regularize the signal.
Even better would be once-a-week roll-ups of curated what matters most.
In the last couple of years, letters of recommendation broke as a quality signal.
Before they were a signal of quality because they took a long time to write.
That meant that professors only had time for a handful of them, so the students they agreed to do it for were at the top of their distribution.
But now professors can do them 100x faster, so they can do many more.
Some new signal will emerge as a quality signal.
Quality signals come from scarcity.
Meaning comes from tension.
One way to bring back scarcity: have professors publish the names of their top recommendations.
Perhaps in a stack rank.
If it’s publicly viewable in a single stack rank, there’s scarcity.
As a company, raising venture capital is a red queen race.
It locks you on a hyper-growth-or-bust trajectory.
Even if you don’t want to get on that trajectory, if your competitor does it, they will dominate you in the market.
That means that even if everyone would prefer not to, everyone must.
A prisoner’s dilemma.
Anywhere where there's no moat (bits, not atoms) and some kind of compounding effects, you have to play.
I asked Claude why heights feel higher looking down than looking up.
It gave an interesting report.
This is the kind of question I wouldn’t have even bothered asking before.
The first time you hear a word, it means what you first guess it means.
Unless the world quickly disabuses you of that notion.
This is what Simon calls an inferred definition.
Conversations are fundamentally generative processes.
The process of ping-ponging ideas back and forth unearths mutually interesting ideas.
Each volley is:
1) a vote to keep going.
2) a curation of which part to respond to.
An iterative “yes, and” that zeroes in on the most interesting thread.
The fact that two people find the thread interesting makes it an order of magnitude more likely another person would, too.
Complexity is a balance of integration and segregation.
Everything needs to be in the critical point to be able to surf.
In complexity, surfing is better than trying to control.
Control is impossible and attempting to have it will tire you out.
A great piece from Every that calls it Rugged Flexibility.
Everything is "trivial" if sufficiently abstracted.
But the abstraction misses some of the grit and texture of the real world.
The reason things are hard in practice is because that grit is load bearing.
Rocket science is mostly just the fundamental dynamics.
Problems that are complex are almost all grit and texture.
Could and should are distinct.
There are things that LLMs could do that you shouldn't do.
If the s-curve is so steep, just try to stay alive until it levels off.
But then you could end up like Blackberry.
In an infinite game, rule number one is to stay in the game.
Finite games are often embedded inside infinite games.
Finite game within (zero sum), infinite game for the collective (positive sum).
The energy of the inner finite game propels the momentum of the enclosing infinite game.
Finite games that are run too hot will steal all the oxygen.
They’ll hollow out the infinite game they’re embedded in.
In the limit, they kill the host.
For example: the idea of America is the infinite game.
The political parties are the finite game.
The hyper era has much more efficient competition.
Finite games hollow out infinite ones at faster and faster speeds.
Swarms are adaptable but have Goodhart's Law.
The antidote is trust in the collective and long-term goals.
When individuals trust each other to behave as a collective they believe in, they will take actions that don’t follow Goodhart’s Law and don't destroy the collective.
Instead of only optimizing for their local incentive gradient, they also balance what’s best for the broader collective.
That gives you the best of both worlds.
Similar to Francis Frei's observation that diverse teams that trust one another are the way to get reliably great results.
Swarms can't see or grapple with systemic problems.
No vantage point to look non-locally.
Trust between individuals pairwise only scales to Dunbar's number.
Past that you need something to give you leverage.
Typically that requires reducing nuance to numbers.
Once you do that, Goodhart’s law starts showing up.
Are you “giving up” or are you “accepting”?