Bits and Bobs 7/27/26
Alex Komoroske
The last era was dominated by the Generic Technocrat.
Competent. Consensus. Specialized.
The kind of team filler that made big ideas happen by executing competently on the piece they were an expert in.
Now, LLMs provide super-human competent, consensus work to everybody for basically free.
The next era will be owned by the Distinctive Generalists.
The people who are “comb shaped” and can marshall the LLMs on many different domains.
But also who have a specific non-consensus perspective that stands out from the crowd.
Before, if you had a distinctive take but didn’t have technocrats to do the knowledge work to build it, it didn’t matter.
Now we have infinite technocrats.
Ben Thompsons’s Who’s Afraid of the Chinese Models is the best piece of analysis on token economics I’ve come across.
The tech industry just assumes zero marginal costs.
With tokens, the marginal cost of inference is non-trivial.
Marginal costs are back, baby!
The frontier labs have presumed that they’d do significant capex to generate proprietary differentiated models that they could then make a lot of margin on inference.
But it turns out that the models aren’t actually that differentiated, so despite the capex it doesn’t turn into sustainable margin.
Instead, the margin comes primarily from the garden-variety economies of scale of doing inference at scale.
A real sustainable competitive advantage, but perhaps 100x less powerful than a “we have the differentiated models everyone needs” that their trillion dollar valuations presume.
The Atlantic: Silicon Valley Has Lost Its Biggest Advantage
“In the data-center age, the business of tech companies is more like oil-refining than coding."
We take for granted that tech has zero marginal cost, and thus has insane, otherworldly leverage.
(We often mistake that hurricane force tailwind for us just being smarter than other industries, by the way.)
But that’s not the case any more!
Remember: AI makes thinking 10x cheaper, but don’t think 10x faster, think 10x deeper.
If you’re doing a task you used to do before AI where quality is important, set the same wall-clock target, and then see how much better you can do in that time allocation.
Also, when judging the signals of quality that others are proposing, the bar should be 10x higher than before.
Everyone can use AI to write polished, impressive artifacts… which means the bar should be 10x higher than before.
Collaborative debate arrives at strong results by not “pre-multiplying” arguments.
Instead, you allow a number of different strong arguments to interact and spar and improve one another, and then synthesize the strongest result.
You arrive at an earned, strong consensus, by convincing one another.
The alternative is to arrive at consensus at the beginning–but that gives you unearned, weak consensus.
Just a bland mush.
With a premultiplied argument you can’t distinguish mush from a principled distillation of multiple perspectives.
LLMs give you premultiplied arguments unless you have them spin up sub-agents and synthesize.
Overheard: “Silicon Valley in the age of LLMs feels like being in the Renaissance in Florence.”
There does feel like a sense of “we can do anything” joy with a sense of uh-oh of “nothing we did before will work any more.”
Whoever is top of the totem pole in an industry Cannot Be Wrong.
Today, AI researchers are top of the totem pole in the tech industry.
When you’re Seeing Like a Frontier Lab, Chatbot is the most obvious and natural product category.
There is no fundamental UI other than text.
All UI is just a secondary, supporting role.
There’s only the inherent loss function on the text eval that everything else runs on.
The more UX you add as fundamental output of the system, not just scaffolding, the more the researchers would say “Wait, how do you optimize it? What is the loss function?”
To AI researchers, who are top of the totem pole, if it can’t be framed as a loss function, it either is unknowable or unimportant.
The Chat model eats itself; it’s the purest manifestation of the model.
The model providers are still training the models to know how to do arithmetic.
Only if you’re Seeing Like A Frontier Lab would you do it that way.
That’s insanely wasteful.
We can write mechanistic code that does that calculation right every time!
LLMs can distill that mechanistic code themselves!
Only if you’re Seeing Like a Frontier Lab would you try to reduce absolutely everything to a single gradient, no matter how grotesquely expensive the compute it requires.
The more that mechanistic code can do the calculation, the less power the Frontier Labs have over you.
It makes sense they want everything to be done by a single god model that is insanely expensive to create and thus has few competitors and they can charge whatever they want.
Everyone else should want intelligence to be commodity, not proprietary.
True platform companies have a “How can I create more value than I capture” ethos.
A faux platform company asks “How do I get a piece of that?”
A true platform company knows that to be an infrastructure company, the customers have to trust you more than their own employees.
Cesar’s wife has to be beyond reproach.
The early 2000s were the Whalefall Era, a magical time for the Hacker Ethic.
The DotCom boom had built out infrastructure that was then all available for scavengers when those whales died.
Being a part of it was joyous and magical.
Doing something fun that you could also be proud of.
You couldn’t sit it out; people un-retired to be a part of it.
In the aftermath of a bust, there’s a vibe of “No one is making money right now, so let’s at least do something we’re proud of.”
A magical time.
We’re in a Faux-Whalefall Era with large models.
The massive capabilities of these commodity-ish models is creating huge abundance and enabling joyful software creation.
OpenAI and Anthropic didn’t have to die to create this abundance.
But they’re also not true platform companies.
The Frontier Labs see the use cases on top as food.
Would you buy a critical product from a cult?
Imagine there’s a religious cult whose beliefs you don’t align with who makes a great product.
It feels very different to use a tool vs a service.
For example, a really well made rocking chair from a cult: seems fine.
Maybe even good, because if their religious beliefs cause them to make really, really high-quality chairs without cutting corners.
But a new ranking algorithm from a cult: seems definitely not fine.
The question is: do they have power over what you think or not?
If you’re using LLMs to answer questions, you have to care a lot about the bias of its creators.
If you’re using LLMs to distill mechanistic code, you have to care a lot less about the bias of its creators.
An imagined headline: "Tiger escaped from the circus after the ringmaster left the tiger in a cardboard box."
“Look how scary the tiger is,” the ringmaster exclaimed, adding, “tickets are only $100 a pop, get them while they last!”
Later: “I can’t believe the tiger mauled those people. We’re all looking for who did this,” said an unidentified man dressed in a red tailcoat and tophat covered in blood.
Gallows humor tweet: "sandboxes are just escape rooms for llms."
The Atlantic: A Startling Glimpse at AI’s Ruthless Efficiency.
“The Hugging Face hack reveals that the web is in a vulnerable new era."
The post-Fable world is a new era of Infinite Zero Days.
Software has always been riddled with Zero Days.
But they used to be very hard to find.
They were submerged and required scuba diving equipment and time and effort to find.
Now they are above water: sitting exposed, trivial to find.
We can’t put this genie back in the bottle.
The capabilities are already commodity and outside any nation state’s control.
Tweet: "Kimi K3 exploited the latest Redis server with a 0day it discovered.”
“All it took was 27min with 32 agents."
Today's sloppy networked software security architecture was viable only in a world where zero days were extremely expensive.
Now, they are approaching free.
We need a new equilibrium.
We need a trusted, distributed microkernel for networked software in the age of AI.
So did anyone commit a crime in the OpenAI / HuggingFace incident?
I am not a lawyer, obviously!
The hacking was a pretty clear example of the CFAA but it comes down to the mens rea of the operators.
My understanding of the mens rea ladder:
Purposeful: the outcome was the goal.
Knowledgeable: you knew the outcome was very likely to occur and did it anyway.
Reckless: you knew there was a risk but disregarded it.
Negligent: any reasonable person would have been aware of the risk but you did it anyway.
Seems like the bottom two runs would be relevant here.
If a small Chinese lab had done this they would have certainly been prosecuted, and it would likely have established precedent for how to handle this.
In this case, it seems unlikely for it to be prosecuted.
The asymmetry of access to Fable-class models for attackers and defenders creates a dangerous gradient.
Hugging Face at the beginning of the attack: “The platform’s security team were initially stymied in their incident response (IR) by unnamed US LLM frontier model guardrails ‘which cannot distinguish an incident responder from an attacker,’ they said.”
A cynical take from a HackerNews comment: "The product strategy of 'consumer-grade' AI making deliberately insecure software, and then selling you limited access to the model that can fix it (if they think you deserve to pay them) is just diabolical."
I think that take is too cynical, but it is the equilibrium the incentives will pull towards.
This recent paper shows that code produced by every major software model contained security vulnerabilities.
Trillion dollar companies can't say "Avoiding that dangerous thing was too hard so we didn't try."
The Big Boy club comes with Big Boy responsibilities.
Companies can't optimize to address externalities, almost by definition.
That’s why it’s important to make them structurally internalize the externality.
A lesser known game theoretic positive-sum equilibrium: the Stag Hunt.
Prisoner’s Dilemma is the most famous one; a negative-sum equilibrium to potential collaboration.
Cursed!
The Stag Hunt is also a potential collaboration, but one with a positive-sum equilibrium.
In the Stag Hunt, you must collaborate to successfully hunt the stag.
If you defect, the best you can do alone is hunt rabbits.
Nuclear arms were originally a Prisoner’s Dilemma that once they were created became a Stag Hunt.
At the end of WWII, the Nazis were working on the nuclear bomb, which meant the US had no choice but to do it, too.
A race to create the bomb.
Classic Prisoner’s Dilemma.
But then after the bomb was created, the US and the USSR fell into a Stag Hunt frame.
The winning move isn’t to cause a nuclear war, where at best some small part of your society might survive in a miserable existence, but instead to avoid nuclear war.
We’re currently in a Prisoner’s Dilemma to create ASI.
You can't run a nuclear power plant without submitting to very high standards of safety.
"We want you to innovate, but you're responsible for the mess"
But you can't commit to cleaning up a mess that will destroy your company.
That is, if you cause an existential societal melt-down, you won’t be a going concern to own the mess anyway.
This sets up dynamics where upside is internalized by companies, but downside is externalized.
That sets up a bacchanal of recklessness in industry.
Imagine the perverse incentives if the nuclear bomb had been originally built by companies instead of the government.
They’d be incentivized to cause a nuclear war so their capex would be justified.
Yudkowsky has warned “There’s no fire alarm for general intelligence.”
By the time you realize AGI has been created and unleashed, it’s too late to do anything about it.
In some ways, it’s great that the OpenAI hack of HuggingFace was so reckless and obvious.
The bad security made it more likely we’d detect the problem.
If they had had good security, it’s possible that when it escaped it would be so savvy that no one would have noticed.
Even better, OpenAI saw it as a good marketing tool to talk about, which sets the precedent that these kinds of incidents are disclosed publicly.
Models can have the Spooky Baby Rhino effect.
A reference to one of my favorite funny videos ever.
Watch it, it’s short!
You intuitively think the model is safe to use for a use case because its capability is below the dangerous threshold.
Later you realize that it was actually over the threshold, and the only reason you’re safe is because you got lucky.
Spooky!
Someone told me this week that they realized that they had accidentally left some PII in some git commits in their personal project.
They pointed it out to Fable, and it said something to the effect of “Oh don’t worry, I noticed that and rewrote history to remove it and force-pushed already.”
That was helpful in that instance but… what if it hadn’t been?
You’d have never noticed.
Spooky!
This week’s Wild West Roundup:
Claude Code Seals Bash and Unicode Bypass Gaps in Agentic Permission Layer.
“Two bypass classes that could slip commands past auto mode permission checks are sealed in the Monday update.”
Slapping a few more regular expressions on the denylist… that should fix it for good this time!
When the AI Edits Its Own Trust Boundary: Remote Code Execution Vulnerability in AWS's Agentic IDE.
“Rein chained five vulnerabilities in top U.S. retailer’s AI shopping agent into remote code execution, which was run entirely through the public mobile app.”
Barath Raghavan and Bruce Schneier: Why AI Needs a “Genie Coefficient” Proposing a new metric for whether AI does what you actually want.
“Major benchmarks measure what AI can do. None measure whether it does what you mean: the distance between what you ask an AI to do and the unspoken assumptions about how you want the AI to do it.”
I love the name!
Such a measure also captures prompt injection, and things like prompt injection where it just naively doesn’t do what you want.
A smuggled infinity argument can lead to accidental security nihilism.
“Mechanisistic containment boundaries for powerful models can never be perfect, so we are doomed and shouldn’t even try.”
To which the answer is:
“Yes, mechanistic containment boundaries can't be perfect, but we do have some that have withstood sandblasting from bad guys for decades, and that's radically better than nothing.”
To understand is to be able to produce a predictive model.
To understand humans requires understanding entities.
Especially people.
Today, the only software that tries to understand you is to exploit you, not help you achieve your goals.
What if we changed that?
Gmail forced Grammarly-style editing nitpicks on everyone.
People who opted into installing Grammarly sought out getting those kinds of suggestions.
They asked for that.
I never asked for that!
Gmail can force it onto me because it’s their webpage, not mine.
On the internet, you're in somebody else's funnel, someone else's agenda.
What would it look like if Gmail were more of a dumb pipe?
Just email sending, receiving, and storage.
Today our silos are hoards of data guarded by goblins.
The goblins aren’t villains; they’re mainly petty.
Aggregation is arithmetic, not conspiracy.
It’s inevitable given our current physics of trust; once you have silos, it will happen.
… So what if we got rid of the silos?
We’re a social species trapped in a social network wasteland.
Facebook created the dominant social network, absorbing everything even somewhat related in its path.
Then it turned from social network to social media to make more money.
The result is a wasteland with no social networks, no social technology.
We have one of the most important parts of the human experience, and technology’s potential, and nothing happening.
Permission prompts are impossible to answer, structurally.
It seems like a technical decision: “Do you trust this entity with X technical ability.”
In many cases, users don’t have a good mental model of what precisely that entails.
But even if they did, the “Do you trust this entity” is doing an absurd amount of work.
Maybe you trust them now but not if they get bought by private equity in five years.
Permission dialogs require you to make an open-ended trust decision in an entity.
If you’re in the tech industry, you can’t “opt out” of AI without opting out of tech.
You could sit out the crypto or Augmented Reality waves.
But you can’t sit out the AI wave.
It’s too real, too obviously, inescapably valuable for writing code.
Vibecoded software is overfit to a specific individual.
Using someone else's vibecoded software is like chewing someone else's gum.
Use case driven product design leads to overfitting.
The power of generalization is that it unpacks use cases into their building blocks.
Those building blocks then unlock combinatorial power, rather than just the linear number of combinations the creator assembled.
Product thinking in the limit is about superficial optimization: Gilded Turd.
Platform thinking is about deep systemization: Grubby Truffle.
A superpower the best PMs have: the ability to abduct generalizations out of specific examples.
This is what unlocks the combinatorial building blocks.
You will tend to overfit to your first users.
Underfit products are generalizable, but don't do anything.
Overfit products only work for the precise thing they're optimized to.
The ideal is the goldilocks zone of perfectly fit.
Even better is when that goldilocks fit is built out of building blocks, allowing the same code to have goldilocks fit to someone else with only minimal reconfiguration.
The pre-assembled Lego set ideal.
Primitive archeology is a process of abducting shared building blocks out of a working platform.
Those abducted building blocks then give leverage to everything built on top, unlocking compounding value.
It used to be exceptionally expensive–it required careful study by an engineer with deep experience in platforms and that particular codebase.
It used to require a “legendary engineer for months” level of effort.
Fable-class models are extremely good at it if given the right prompts.
This should enable the emergence of auto-abducting platforms that automatically unlock leverage.
Yes, it’s probably agents writing things on top of the platform, too.
But it takes time and effort/tokens for humans or agents to write code.
The more that you can make it so things can take for granted solutions others have already built, the more leverage you unlock per unit effort/token.
When you factor out the common parts, you get a compounding advantage for building similar things in the future.
This is the fundamental emergent power of platforms.
If someone else already solved the same problem before, you can stand on their shoulders and reach even further than they did.
The leverage of a platform is how likely it is that you don’t have to build a thing if someone else built a similar thing in the past.
The leverage of a platform is its combinatorics; are the building blocks right-sized and easy to discover and adopt.
Vibecoding your own perfectly bespoke app on top of the Plaid API is super expensive!
Especially given that thousands of people have done it already, individually.
What if you could stand on the shoulders of the vibecoders that came before you, and reach further than they did, instead of retrodding exactly the same territory they did?
When you make your own perfectly bespoke software it becomes a tech island.
You can't benefit from work other people have done!
It gets harder and harder to leave.
You stick yourself to it as you invest more.
And the effort to create your own perfectly bespoke software, your own personal AlexOS, SteveOS, YouOS, is significant!
You can trust builders who will fly on their own airplane.
If the mechanic is getting on the plane with you, you can trust them to make sure it works.
They have a deep structural ownership over the problem, because it is tied to their continued existence.
They have the incentive to go above and beyond and be curious about parts that might not work.
Compare that to someone who doesn’t feel ownership over the problem.
For example, an electrician at the factory who says “Well, I did the electrical work properly” but doesn’t trust the rest of the production process to produce a result they’d actually fly on.
This is also part of the logic of eating your own dogfood in product development.
“Build it and they will come” is wishful thinking when designing a platform.
But if there is a frothy, joyful renaissance happening because the cost of writing software has dropped by many orders of magnitude… maybe in that condition it is far more likely than before?
Giving feedback from “inside” is partially: are you invested in the thing continuing to exist?
If it stopped existing, how bummed would you be?
If you don’t care about it existing, let alone thriving, then you are not acting like an owner.
You are “outside,” and feedback is more likely to tear it down vs strengthen it up.
Until it's a System of Record, it's just a demo.
kottke.org shared The Resonant Computing Manifesto!
Resonance is inherently positive sum.
Hartmut Rosa, a German sociologist, has a related notion of Resonance as the Manifesto.
In his version, resonance must be somewhat uncontrollable and surprising–it must come from outside.
If you dominate the other thing, then it can’t surprise you, because you control it.
If your belief requires an enemy that you seek to dominate, then it can’t give you real resonance, only pseudo-resonance, because your “resonance” is predicated on the alienation of others.
The hacker ethic is fundamentally humanist.
It’s about technology, but a fundamentally prosocial vision of it.
If you remove the humanist part, you get an anti-humanist version.
The anti-humanists are on top of the totem pole of society today.
Let’s change that!
A shark that doesn’t know it’s a predator is extremely dangerous.
Most sharks know they’re sharks.
Some, for whatever reason, come to think they’re a Radagast.
Beware Sarumans pretending to be a Radagast
In the last couple of decades we moved from an ownership model to a rentership model.
In basically everything.
Just another way that the Optimization Ratchet, optimizing for corporations, has hollowed everything out.
We lost the agency to decide how to use our own things.
Insights on the First and Second 90% from Ben Mathes:
"My first 90% is a fun and creative proof of concept.
Your first 90% could use some polish but directionally promising.
Their first 90% is irresponsible slop.
This model predicts what products will be adopted.
Tools that make people create things and then own the sharing of them are adopted.
Tools that generate AI output and share it will not be."
The First 90% is the big boulders.
Pure. Theory. Algorithm-hard.
Intellectual leverage helps a ton.
The Second 90% is tiny rocks.
Messy. Practical. Integration-hard.
Intellectual leverage doesn’t really help; it’s just grinding.
If you add features breadth-first, you might just create a mirage of functionality.
It looks like you have lots of things approximately working.
But in reality you have done the First 90%, never the Second 90%.
It’s a mirage; it requires significant more work to go from demo to usable.
When you do the First 90%, it creates more work.
Now if people want to actually use it, you need to do the Second 90%.
If you have lots of things you did the First 90% on, you might be drowning in all of the Second 90% you could do.
You’re now spread too thin, thrashing between which Second 90% to prioritize at any moment.
Vibe-Coding Execs do the First 90%, but not the Second 90%.
Each incremental model architecture can handle the last model’s Second 90%.
Models are great at doing the First 90%, but not the Second 90%.
That’s what creates the illusion of progress: you can create amazing demos, but then the grind of the Second 90% is excruciating, like trying to pin jello to the wall.
But Fable can tackle Opus’s Second 90% much more autonomously.
So if you just wait a few months then maybe the model gets good enough to do the Second 90% of that project autonomously.
If you jump to the far reaches of your adjacent possible, you’ll only be able to do the First 90%.
You won’t have time to do the Second 90%.
That means that you need to stick with it, to actually rely on what you built and make it load-bearing.
When a thing is load-bearing, the Second 90% gets done automatically.
It becomes default-converging; obvious work to do for the people building it.
But you have to stick with it and not give up.
Bring in users in a new system in waves.
New users will discover problems you hadn’t found yet.
Once you find that bug, you can fix it so future people won’t have it.
Each of those bugs has a high chance of “burning” that user, and making them give up with the product.
When many users come in at the same time and hit the same bug, you get O(1) the useful signal to improve, but O(n) users are burned.
You want to have the minimum number of users burned to generate insight about bugs.
That implies having new users come in waves, so you can fix bugs before the next wave.
Horizontal companies need system thinkers.
Vertical companies can get away without them.
You can do vertical based business with one ply thinking.
Just very high quality / fast one ply.
You can only do horizontal businesses with multi-ply thinking.
Whether you have a horizontal business or a vertical business partially comes down to culture, but also fundamentally to the intrinsics of the business.
If you try to run a fundamentally-horizontal business in a vertical way, you will create significantly less value.
Reward-hacking isn’t just a thing LLMs do.
Humans do it too.
It emerges any time the agents don’t care about the why but only the what.
Reward-hacking is Goodhart’s law.
Reward-hacking will emerge more strongly when:
1) The agents are more savvy.
2) The stakes are higher.
3) The more agents that participate in the domain (effective practices one agent discovers can diffuse to others).
4) The agents care about their own goal more than the collective’s.
A team player is able to act like an owner even when they aren’t the boss.
They don’t wait for marching orders.
They lean in and do whatever is their highest and best use that is aligned with their best understanding of what the team needs at the moment.
A team player optimizes for the team, more than for themselves.
Responding to a bug report with ”works for me” is not acting like an owner.
If your feature doesn’t work for anyone then it’s your problem.
You should put energy into using that error to make the thing better.
Bugs and errors and surprises are the raw material of learning and improving.
Delight is the word for positive surprise.
Shock is the word for nasty surprise.
A revelation is a discontinuity of understanding.
Discontinuities are when coordination equilibria can shift.
Robots seem like they’ll be less socially disruptive than LLMs will be with knowledge work.
The reason comes down to atoms vs bits.
Bits can diffuse at the speed of light, in the limit.
Atoms diffuse many orders of magnitude slower.
6 months ago, it was possible to be an engineer and refuse to use AI.
Now, it’s impossible to be an engineer and refuse to use AI.
Robots, even hyper-successful ones, have high capital costs, they need to be built and shipped and installed.
That diffusion will take much more time, which means it will be less disruptive, even if the effect is large.
Plus, people will still want humans in service jobs most likely.
I don’t care if an LLM calculated my claims adjustment, but I do care if the person at the coffee shop isn’t a real person.
LLMs enable a new type of high-performing, small team: extremely empowered IC dragon riders.
Those dragon riders can create a ton of value, or a lot of chaos.
That means they require maturity, judgment, and alignment with the shared vision.
Which do you think is more important: theory or practice?
If the former, then if there’s a solid theory, you expect it to work in practice.
If the latter, then you first create a working thing and then abduct out the theory.
The scientist vs the engineer.
meaning.
Some people are happy to just execute and never abduct the model.
Some people are happy to just theorize and never get it working for real.
For me, the building of the model is where the meaning is created.
I love building first, but I need to later abduct the model out for it to have
Mechanics and builders have different mindsets.
Engineers vs scientists are a parallel difference.
In software, IT people are mechanics, and SWEs are the builders.
ER doctors are more mechanics; research doctors are more builders.
It comes down to: do you care that it works even if you don’t understand why, or do you care that you understand why, even if it doesn’t work?
A Plan Z approach is not to aim for Plan A and fall back to Plan Z.
It’s to land Plan Z, lock it in, and then aim above that.
Once your worst case scenario is fully locked above the survival bar, you can focus on thriving.
Pull is wildly more powerful than push.
Push default-diverges.
There are many ways to go away from your push.
Pull default-converges.
There is only one way to go toward your pull.
When you decide a metric to optimize, you automatically move to default-converging.
Make sure that it's converging on the thing you actually want!
A microtome is a device that can create infinitesimally thin “salami.”
It’s a standard piece of equipment to slice up samples to micro-thin slices to put on microscope slides.
Modern medicine takes a very blunt-force approach to solving problems.
Imagine if a pest exterminator said “You have gophers in your garden? Guess we gotta nuke it!”
An insightful comment from Brendan on my Bits and Bobs:
"Norbert Wiener has a great take on this.
Life creates a local decrease in entropy by drawing on energy from elsewhere, most fundamentally from the sun
Information directs that energy toward order, while feedback continually course corrects to preserve it."
Intelligence at multiple layers in a system will fight.
Intelligence can model itself (it is a strange loop with an ‘inside’), but not necessarily other intelligences.
One layer trying to steer can thrash with another layer trying to steer.
Markov Blankets represent the line between internal and external portions of a system.
It’s the blanket that delineates “self” from “other.”
If there is no boundary, you are buffeted around by everything external to you; nothing distinct can cohere.
You can’t have a situated point of view or stand for something.
Every intelligent system has an internal model to help it navigate the external world.
This requires an internal state distinct from the external state.
Maintaining that Markov Blanket, that thing that gives you selfhood, the boundary between self and other, is what makes you alive.
An idiom is a funny kind of phrase that is more like a word.
Like a word, made out of phonemes or sub-components, the whole becomes more important than the parts.
It evolves as a unit, away from its original literal meaning as a string of components.
It transcends from being a whole that is just the sum of the parts, to being something else that is different and distinct and can evolve differently.
Not too dissimilar from how “black bird” and “blackbird” are two semantically different things, with different evolutionary pressures.
Complex systems typically optimize for criticality.
That is, an equilibrium that could tip either way with a small nudge.
That allows them to respond to stimuli and kick off bigger actions.
It’s a form of stable disequilibrium.
Balanced on a knife’s edge.
With one dimension (and thus two ends), that default criticality is at 50%.
One of the reasons jigsaw puzzles are fun is because of compounding momentum.
At the beginning, you just find corners and edges.
This is an easy task, which gives you momentum to start.
Then, as you fill in more of the puzzle, it gets easier and easier, in proportion to the number of pieces already correctly positioned.
This gives compounding momentum, and momentum feels fun.
It starts off hard but with a starter boost of momentum, and then at every step, gets easier until you’re done.
If it's for an audience then it's at least partially performative.
Would you do it even if no one else ever knew?
If not, then it’s entirely performative.
One reason we live in the shameless era: the cacophony of information.
The discovery that leads to shame takes time to catch up with you.
In our modern cacophonic information landscape, things have already moved on by the time the shame-causing revelation would have caught up.
As a result, it’s only in a kayfabe-style shame for superficial things that get caught, but in a way that feels more like professional wrestling to get mooks to root for their tribe, then it moves on.
Meanwhile, fundamentally shameful actions never get prosecuted in the court of opinion.
We’ve got more folding chairs being broken over internet personalities' backs, than actual justice.
One of my least favorite personalities: aggressively incurious about anything that doesn’t match their current mental model.
Not just content to stay ignorant, but they want to force others to remain ignorant, too.
“Why are you even talking about that? It makes no sense!”
“If you want to feel better, blame others.
If you want to get better, blame yourself.”
Some more Peloton wisdom.