Bits and Bobs 2/3/25
Alex Komoroske
Both The Algorithm and LLMs are ultimately powered by human decisions.
The Algorithm here meaning any ranking function that relies on human interaction to rank an infinite feed.
The Algorithm, pre LLMs, creates an emergently intelligent ranking better than any individual human could do.
It does that by extracting the consistent bias in massive amounts of noisy signal flowing through it; the cacophonous actions and decisions of a wide swath of humanity.
But it requires a constant flow of these human decisions to continue working.
LLMs also work by capturing the power of decisions of wide swathes of humanity.
But it works retroactively by looking at the things that humans cared enough to write down and share in the past.
It creates a hologram of that background human motivation that can then be queried forever into the future without humans in the loop.
LLMs allow a good default strategy when tackling a problem to be “just start.”
If you don't know how to even start, how do you start?
Most ideas stop in this cold start problem.
LLMs reduce the cold start problem, the empty search box problem.
LLMs warm you up for any problem.
Just talk to it with whatever raw thoughts you have and it will help you pull on that thread.
In Getting Things Done, you're supposed to do the task if it takes less than two minutes.
But now the class of things you can do within 2 minutes is much larger.
For example, if you have an idea for a little widget, why even write it down as a note to do later, just have Claude prototype it right at that moment.
LLMs could allow humans to think 10,000 times more.
But whether that’s a good thing depends on what they bother thinking about.
Thinking about creating additional meaning, collaboration, or value? That’s good.
Thinking about how to get an edge in a zero-sum meme war? That’s a waste.
The zero-sum social media hellscape is a black hole that absorbs all of the mental energy humanity is willing to give it, without limit.
People are acting like chatbots is the UX for the AI era.
But it will just be a UX.
Chatbots are like UNIX terminals.
They're eternal but not universal or mainstream.
Chatbots were the first breakthrough UX where users could look at it and instantly “get it”.
The only reason chatbots look like the UX is it's the most obvious demo and the models have so much excess quality that it can be compelling on its own.
But they also cement the wrong mental model: “LLMs are just like a person, but virtual.”
That vastly misunderstands what LLMs are, what they can do, how they could be used.
LLMs are an alien brain.
They can do things no human could do; they fail to do some things that any human could do.
The chatbox mental model is the “original sin” and puts us in the catchment basin of lines of thought like "think of chatbots like employees to do work cheaper than humans."
The skills to build at different pace layers are just different.
For example, the skill to make a good browser is very different from the skill to make a great web page.
So being great (culturally, skillset wise, experience-wise, etc) at one layer all else equal makes you worse at other layers.
An individual entity can't be great at multiple layers.
If one entity somehow does well at multiple pace layers, it's like winning the lottery independently twice.
If it happens, the entity likely has two sub-entities internally that are quite distinct so they can optimize their culture for the different pace layers.
Centralization is a late stage phenomena within a technical paradigm.
When a new disruptive paradigm blossoms, everyone applies the late stage playbook of the last era to it.
That's how you get new entrants AOL-ing themselves.
Putting themselves on a dead end tech island.
But then once the internal emergent disruptive logic gets activated and swarming, it overwhelms the centralized approach.
Three phases of a new disruptive paradigm:
1) Vertical integration at the very beginning for the breakthrough proof-of-existence.
One particular centralized / verticalized approach wins, but only for the beginning.
This is the AOL phase.
2) Cambrian explosion of open-ended ecosystem exploration.
Openness beats any vertical entrant.
3) Consolidation as the best practices are all discovered and now become increasingly efficient.
A few dozen highly centralized players emerge that act like gravity wells; everything else fades away or is gobbled up.
Then, repeat with the next disruptive paradigm.
This cycle takes a couple of decades to play out.
When analyzing the market dynamics of AI, it’s important to separate the model layer from the application layer.
The model layer is the creator and operator of LLM models.
The application layer is the creator of the UX that actual end users use.
These are two extremely different layers.
They are different pace layers.
The competitive dynamics are different.
They require very different skills to succeed in each.
Lots of people are mistakenly analyzing them as though they are joined.
For example, Ben Thompson’s analysis of whether aggregation theory will apply to AI or not is only actually only about the model layer.
This leads to lots of confusing takes on whether we’ll see centralization or decentralization in this new era.
They’re confusing the engine for the car: https://glazkov.com/2023/09/17/the-engine-and-the-car/
One of the reasons people are implicitly combining them in their heads is because OpenAI, Anthropic, and Google all have entrants in both levels.
But this is more an artifact of the “vertical integration for proof of existence” phase of the new paradigm.
It’s not at all clear to me how success at one layer gives an advantage at the other layer.
Arguably it gives a disadvantage since they require such different things to succeed.
The best arguments I see for vertical integration:
Cost structures: the application layer can pay less because they’re under the same roof and there doesn’t need to be a margin, or they can even be subsidized.
But costs for models are decreasing substantially.
Proprietary quality edge: the application layer can get access to the latest models not available via the API.
But quality is quickly commoditizing too, and it’s more likely to create a tech island phenomena.
I don’t find either of these arguments for integration compelling.
It seems pretty clear now that at the model layer we’ll see commodification–we already see it to a striking degree.
OpenAI is clearly gunning for an aggregator advantage at the application layer, but I’m skeptical they have strong enough fundamentals to pull that off.
Seems more likely to me they’ll AOL themselves.
It’s not yet clear what will happen at the application layer, though I personally hope we’ll see an open decentralized system take off there too.
I think people who talk about “the AI” in the next era are getting confused.
They are making two incorrect assumptions, I think.
1) That AI will be a centralizing force at the model layer due to capital costs.
2) That the model layer and app model will be vertically integrated.
Both seem likely to be incorrect to me.
It seems like playing with reasoning will be enormously useful in improving the capability of models.
In the earlier micro-era of LLMs it was all about the scale of how much world knowledge you could cram in.
Extremely capital intensive.
It feels like we’ve topped out on that, and now the new micro-era of competition is extracting as much power out of reasoning.
This is a distillation exercise, but also a UX and tinkering challenge.
Reasoning currently is not great; it often comes back to the wrong point multiple times before breaking through.
That implies there’s tons of low hanging fruit.
OpenAI tried to keep the reasoning tokens as proprietary advantage.
But it turns out it was extremely easy to copy.
Now they’re on a tech island.
OpenAI has only a few hundred employees to tinker and come up with ideas of how to get models to reason better.
DeepSeek and other open-ish alternatives can benefit from the exploratory capability of the entire ecosystem.
We seem like we’re just now entering the Cambrian explosion micro-era for the model layer.
Two very different approaches for LLMs now:
1) The ‘Cloud Provider’ model: commodity hosting of models.
Compete on cost.
The hosting is the point.
The model is commodity.
2) The ‘LLM Provider’ model: proprietary hosting of models.
Compete on quality.
The model is the point.
The hosting is commodity.
The latter works best when the model is an order of magnitude better than alternatives.
But if the proprietary model is only 10% better than commodity alternatives, the benefit goes to competing on cost.
Especially because even the current generation of models have so much excess quality.
An embarrassment of riches, if you use them in a way that is resilient to errors.
The AI native software ecosystem must not be a walled garden.
People assume it will be because we’re in the era of "oops, all walled gardens."
But that’s more an artifact of being at the late stage of this last technical paradigm, the hyper efficient hyper competitive part.
If all of humanity's potential thriving is in this new software ecosystem it simply cannot be a walled garden.
It would be a travesty to make it a closed ecosystem and not an open one.
Especially to give up in the first inning.
In uncertainty (like in the early stages of a disruptive paradigm) people cling to legibility, not importance.
An example of the streetlight fallacy–the drunks looking for keys not where they are likely to be found but where they can see best.
It seems clear to me the alpha is not in making the LLM but using it in a differentiated way.
That is, not the model layer, but the application layer.
Making the LLM has legible benchmarks that people cling to in uncertainty.
They can coordinate around them.
But they are not actually the most important thing.
The companies that laid fiber optic cables were never going to own the app layer.
This is self-evident now but it should have been self-evident even from the earliest innings of the internet.
The current class of preeminent LLM experiences are all fundamentally chat-first.
Imagine that you create an Artifact that you find very valuable in a Claude chat.
How do you find it a month or two later?
There’s no directory of all the artifacts you created.
The artifacts are a secondary thing that just so happen to have emerged from a chat.
You have to go find the right chat to then find the artifact.
What if you wanted the artifact to be the primary thing; the chat is just the thing that happened to create it.
You can’t keep a number secret; the asymmetry is too strong to break.
In the early 2000’s the encryption key for DVDs leaked.
The powers that be tried to sue anyone who tried to publish the number.
But that was an impossibility.
My favorite example of the era was a t-shirt that enumerated all of the integers immediately before and after the encryption key, with the actual key’s slot conspicuously empty.
The shirt underlined the absurdity of takedown notices for an integer.
Information wants to be free–because it is non-rivalrous and naturally viral.
OpenAI's trying to own the application model now looks much more like an AOL approach to the internet.
The idea that OpenAI could corner this whole market looks quaint.
You can't control a number, the asymmetry is too strong to leak.
LLM model weights are just really really really big numbers.
The downside of creating a reality distortion field is that you can get caught up in it too.
You can come to deeply believe it and don't even see how it could be wrong.
You will completely miss disconfirming evidence, because it so deeply goes against your fundamental belief.
The only thing more dangerous than putting on blinders is putting on blinders that hide details that you are intrinsically motivated to not want to see.
The person who contributed the stone to stone soup doesn’t get to control the soup.
Imagine trying to tell people trying to sample the soup that it’s yours, and they can’t copy it.
The soup’s quality came from everyone else’s contributions, not yours!
In the aggregator era, you get to choose which overlord you want.
But what if you don't want any overlords?
Mary Poppins meme: “We have both kinds of overlords: ads-surveillance and hardware-lock-in.”
Aggregation theory is true partially due to the laws of physics of the same origin model.
Aggregation theory is what gave us the overlord problem.
The overlord problem: the entity that is supposed to work for the user actually has the power over them.
Because each user is a drop in the ocean to the overlord, and the drop can't survive on their own, they'll evaporate.
The user has to hope the overlord deigns to solve their particular problem or use case.
The larger the ocean, the less likely the user’s bespoke problem is to be solved, and the harder it is for a new lake to be created for the user to escape to, because all of the water flows to the ocean.
Imagine: software that works for you.
That is, it does precisely what you want; it works for your use case.
In addition, its incentives are aligned with you, not the entity who created it.
Neither is true in today’s centralized world.
Things that are truly game-changing change how people think, feel, or act.
If they don’t change those things, then it’s just window dressing, a flash in the pan, superficial.
Game-changing things change fundamentals for people.
A nice distillation of a long Karpathy tweet by Israel Shalom:
“play = reinforcement learning
instruction = supervised learning”
Aggregators can't allow turing complete things within themselves.
One of Gordon’s classics: Aggregators aren’t open ended.
If they allowed turing-complete things, then they’d incubate the seeds of their own destruction.
It’s also very technically challenging to allow turing-complete things that interact with their surrounding context safely.
But this means aggregators can’t grow beyond a certain size.
When they start growing in the late stage, they look unstoppable, but in a new disruptive epoch that size is a liability.
Disruptive things disrupt everything.
"This one thing that was pinned down is now free floating, I can do a new business not possible before."
"Actually all of the stuff around you, including the table you're building on, is now free floating, so that thing you want to do isn’t viable for a totally new reason."
Disruptive contexts require luck or multi-ply thinking to succeed in.
At the beginning of disruptive eras, people only explore ideas adjacent to how it used to work, just with a twist.
It takes a long time for the swarm of exploration to diffuse beyond the immediate adjacencies.
Don’t add AI to a thing that used to work before, imagine the things that are only possible in a world of AI.
This is much harder to imagine; it requires multi-ply thinking, and the things will seem weird at first glance!
Because they will not fit in the last paradigm we’re familiar with, but only the new one that we aren’t yet familiar with.
A paradigm shift is like the Matrix.
You can’t be told; you have to experience it.
A paradigm shift can’t be explained in terms of individual use cases.
A paradigm shift changes the fabric of reality, reconfiguring what is possible.
No enumeration of individual use cases can express that totality.
In disruptive epochs, aiming for anything less than a paradigm shift is too small.
Paradigm shift or bust, baby!
If you want to make true believers, bend the spoon.
Imagine you realize we’re living in the Matrix and you want to convince other people that that’s true: to make true believers of them.
What do you do?
You perform something that is easy if we live in the Matrix, and impossible if we don’t.
Like the kid in the oracle’s apartment in the first Matrix, bending the spoon.
It looks like he’s controlling it with his mind–a miracle.
But really he just understands he’s in the Matrix and has hacked some bit of code for the spoon.
Less a miracle worker, more a script kiddie.
If you see a new paradigm and want to make others believe, bend the spoon in a way that’s easy in your new paradigm and impossible in the old paradigm.
Why is there not a market for spreadsheet templates?
One reason, spreadsheets intermix data and code, making them hard to separate into templates.
Another part of the challenge is that spreadsheets have a lot of complex state that is almost entirely hidden.
To understand someone else's spreadsheet you have to overturn every rock to see what logic's hiding there.
You can't see it, it’s hard to have the information scent of where to look first.
But also if all of the details were visible you'd be overwhelmed by it.
Humans and LLMs would both struggle.
A tangled mess of wiring between cells.
Finally, when the medium to express your ideas has no opinion, it just ends up being "just however you thought about it".
Which is good for getting it out of your head.
But inscrutable to everyone else.
It's really hard for another person to orient what's going on.
More degrees of freedom, create more things for others to have to orient in novelty.
Templates are great to inspire you about what's possible, but people want to keep things as close to the way they think about them as possible.
Software today is a static one-size-fits-none UX wrapper around someone else’s database.
You have to hope the operator of the database gets around to adding your use case, which, if it's a bespoke use case, or in any way against their business interest, they definitely won't.
The fundamental point of software is to help humans accomplish something they find meaningful.
Humanistic technology will be like organic tech.
Healthy for you. Aligned with your interests.
Today's software is unhealthy for us.
Junk food. Optimized to make us salivate and gorge.
"Have you seen this terrifying but captivating meme? What about this one? … "
Organic software would be healthy for us.
Help us do things we find meaningful.
Max Bittker: “social software needs to be grown”
Aldous Huxley: “By these means we may hope to achieve not indeed a brave new world, no sort of perfectionist Utopia, but the more modest, and much more desirable objective — a genuinely human society.”
What is liquid software?
Composed of disposable cheap small components.
Even if each individual bit of software is inflexible, they can still be flexible in combination.
The trick is combining them on demand in a way that makes sense.
If you can conjure up any small shitty bit of software you want on demand and coordinate with other bits, the combinatorial outcomes are limitless.
LIquid software will present itself as micro-apps.
Not because that’s what is its natural or highest form.
But because that’s the form that our minds can most readily accept, having been trained on shrink wrapped software from the world of today.
It looks like an app just for your convenience.
Like the aliens in Contact.
Liquid software: you pour it into a container, it takes the shape of the container.
It's just liquid, it's just bytes.
Who cares about its intrinsic shape? It doesn’t have one.
It takes on the shape of whatever container you put it into.
LLMs are a natural ingredient for liquid software.
We need better containers for liquid software.
The chatbot UX is fuzzy but not liquid.
A classic question for new platforms: “what is the killer use case.”
I’m not a huge fan of that question: the killer use case of an ecosystem is not any one use case, it’s the fact the ecosystem can swarm on innumerable use cases emergently.
A frame that gets at that ability: "I'm looking forward to being surprised by the system to do a thing I didn't expect"
The “killer use case” question is more often a “what is a concrete example of the edge of the wedge that will lead to adoption of this ecosystem,” which is at least a reasonable question.
Use case thinking comes from a world of expensive software.
"What's the killer use case?"
The question evaporates in a world of liquid software.
In liquid software, everyone can have their own personal killer use case.
Their personal killer use case won’t look that compelling to anyone else, because it’s hyper specific to them.
“Here's my killer use case.”
If there were lots of people with that exact same use case, then it would already exist as an app in the platforms of today.
The complexity of the browser coevolved with the web.
Current browsers are massive, complex beasts.
But the very first browsers were quite simple.
No scripting, no CSS.
They didn’t really need an explicit security model because they were 100% declarative.
As long as the bones that allowed the new distribution model were there, they could be simple.
If you were trying to create a new “browser” for a new web, distributed within the old web, it could start much simpler than you might think.
The web you can think of as disruptively shitty software that runs in a runtime (a browser) that users pre-installed.
You distribute one special program, the browser, that has totally different laws of physics within it that allow distribution within it radically unlike the surrounding universe of software.
Shitty software in the small is now free, so instead of thinking "how can I make the LLMs coding output more like how we write software today", think "what can we now create given that this whole class of stuff is now free"
You need a new architecture and distribution physics for a new class of shitty, newly-viable software.
The friction of a distribution model is inversely proportional to the strictness of the security model.
That is: how many capabilities the untrusted turing-complete code gets.
The ability to cause side effects in the world is the power... but also the downside it could cause.
Ideally you create a system to cap the downside and leave the upside open.
But that's hard, and by default the downside and upside are symmetrical.
The ceiling of usage of an ecosystem is tied to the expected value of the worst case downside.
A small problem that is very likely could have a high expected value, as could a game-over but rare problem.
Even if 99% of people in the ecosystem aren’t savvy enough to understand the downside risk, it doesn’t matter–the system will limit itself naturally to avoid that worst case downside risk.
Technically the expected value of the worst case downside is balanced out by the expected value of the normal case upside, so something that allows something radically better than before can get usage before hitting that ceiling.
Greasemonkey back in the day had this problem.
The reason security models are hard to retrofit is that everything in the system is downstream of it.
The security model is at the lowest pace layer.
Hard to change, but extraordinarily, mind-blowingly high leverage.
The leverage is why it’s powerful and also hard to change.
You can't retcon one onto an existing system, because they set the laws of gravity.
If you haven't done it before you won't even realize what to do. It's inherently multi-ply thinking.
The most helpful people are the ones most liable to being tricked via social engineering.
If LLMs weren't guidable, then they wouldn't be useful.
So they have to be at least a little gullible.
LLMs are fundamentally exposed to the prompt injection problem.
There’s no containment boundary between the data and control planes.
Unlike in SQL, there’s no structural way to escape possibly dangerous input and remove it from the control plane.
It’s all just squishy text.
Normal programs are robots that can't be tricked in novel ways, because they can't think.
They can only do precisely what they were programmed to do.
They can only have a structural pre-existing weakness exploited.
But LLMs are like helpful humans: super gullible.
When you mix in LLMs with the chonky same origin / POSIX model it’s fundamentally insecure.
LLMs are too flexible, too gullible.
The same origin model is too coarse.
It's easy to be tricked by an adversary with your gut reaction.
It's harder to be tricked if you get a chance to reflect on it.
That's why adversaries often create faux urgency.
This is a reason to be optimistic that reasoning LLM models are less likely to be tricked by adversaries, but the fundamental problem is a lack of separation between control and data planes.
The same origin model makes it easy to try a new thing, but hard to deeply engage with it.
Every new thing you try starts off knowing nothing about you.
That makes it safe to try a new thing.
But everything the new thing learns about you it can do whatever it wants with, so you need to trust it.
That makes it harder to come to incrementally use it for more private things.
Is evolution optimistic?
No.
But it is an impossible-to-stop force.
So we might as well try to direct it towards optimistic ends.
One challenge of creating a ubiquitous, open ecosystem: a schelling point everyone trusts to coordinate around.
At some point you need to reach out across the network to some other entity, and if you want to coordinate with others, you might all have to agree to use one entity.
Any entity across the network could do whatever they want on a technical basis (it’s their turf) so you just have to trust them to do what they say they will do.
This is one superpower of blockchains.
It creates a schelling point, a protocol that everyone can trust by construction to do the thing it purports to do.
No individual has to be trusted to do what they said they’d do, because everyone is trusting a visible mechanistic structure to do what it clearly does, not relying on a specific entity to keep their word on behavior and never change it.
Blockchains are one way to solve this, but they are an expensive way to solve it.
Another approach is to use confidential compute’s remote attestation capability.
Create an open-source runtime and then in the protocol require verifying the remote attestation of the remote node.
This allows a thing everyone can know is faithfully executing what the operator said they would.
Before you had to trust the operator of the server to not get greedy, evil, incompetent, etc.
But if all you need to know is that they are definitely hosting the runtime they say they are, you're done.
Then make sure there are multiple peers so if the main one goes evil you can fail over quickly to a new one, and only lose a few minutes of traffic.
There’s also no incentive for a node to go evil because they’d be found out and routed around in minutes anyway.
But critically the remote runtime has to be bit for bit the open source thing, no proprietary extensions.
As soon as there are proprietary closed source extensions in some nodes, you no longer get the ubiquity and it becomes possible to fork the ecosystem.
A decentralizable system has the incentive to not mess it up.
Because people can fork if they do.
This helps keep the current operators of the decentralizable system honest.
The operators chain themselves with constraints which makes it hard for them to do the thing the community fears they will do.
Financialization makes things more transactional.
Efficient transactionalism allows fluid markets to form, but in a way that removes the underlying thing’s soul.
Transactional creates efficiency while removing meaning.
Finite vs infinite.
Focused on means, not end.
A thing that is not legible can be infinite.
Once it's made transactional and legible it is automatically finite.
In the last decade humanity has created environments that are hyper-financialized, the manic pixie dream of financialization, the fullest, platonic form of financialization.
This hyper-financialized universe creates an auto-enshittifying vortex.
As it rotates faster the vortex pulls more energy into it, and hollows out the internal meaning even further.
The tech barons say, "tech is great, so give me all the power."
The crypto guys say, "tech is great, I don't want your power, I just want your money."
Someone should say, "tech is great. I don't want your power or your money, let’s work together to make tech work for all of us."
Knowledge is easier to commoditize than wisdom.
We're now in the hyper commoditization of knowledge.
Wisdom will get even more valuable.
Taste is a form of wisdom.
Last week I learned about “small giants”
"A ‘small giant’ is a company that chooses to optimize for mojo instead of growth"
Another frame for last week’s "tech island": an early proprietary advantage that catches you in a dead end where the rest of the world passes you by.
They’re more likely to happen in immature environments, where a proprietary lead helps at the beginning but where most of the energy will come later, in the ecosystem, not you.
They’re less likely to happen in very mature environments, where all of the good ideas have been found, and the environment is approaching heat-death.
Just because you move the cost external to your model does not mean the cost went away.
You just changed how you account for it in a misleading way.
A large portion of “cost reduction” approaches actually just move the cost to a less legible region external to the model.
Sweeping it under the rug.
Monsters swept under the rug can still bite you.
Programming is pure; the real world is profane.
It’s possible to have a pure, elegant model in programming.
The question is if you can bridge the messy reality of the real world into it, to make it actually useful, and not some blueprint for a castle in the sky.
The ultimate measure of the usefulness and power of a system is not “how hard it is to accomplish things in the pure virtual world of programming” but “how hard it is to accomplish real things in the real world with it.”
The former can conveniently ignore the work and effort it takes to try to cram the real world into the pure representation.
The latter captures the actual full cost and benefit.
A thing that I see lots of very “pure” programming systems missing is how hard they are to represent the mess of the real world in.
You can’t replace an ecosystem with a slightly better version.
For example, “Exactly the same as the one you currently use, but a bit more private.”
In practice, of course, to get that additional benefit on one dimension like privacy requires compromising on other dimensions, so the thing is actually arguably worse than the incumbent.
The network effects and internal momentum of an established ecosystem have to be actively overcome with something not just a bit better, but wildly better in some salient dimension.
Once you start having a narrative for an evolving phenomena, every additional thing you sense you'll interpret into the narrative.
Things that don't fit, you'll be structurally more likely to ignore as noise.
But the noise you ignore might collectively be structured enough to suggest that the model / narrative is wrong.
Theory coevolves with raw experience.
Be wary of putting on blinders too early in the process of model construction.
We’re never done constructing models; the world is not static and continues to change.
Just because the model has been a good one for the past ten years does not mean it will continue to work just as well in the next ten years.
The longer it’s worked, the more likely it will continue to work, but it can never be perfect; there’s an invisible asymptote it’s approaching.
Be careful about blindly following your early adopters.
The early adopters can pull you in weird, random directions.
A mental model to get the underlying dynamic:
Imagine an underlying distribution of points with 0.7 correlation.
Imagine adding one point from the distribution at a time randomly and drawing the line of best fit.
At the beginning the line will jump randomly with every new point.
When you get to the 1000th point, the line will barely move.
Disruption theory makes it seem like disruption is default.
It’s actually rare, we just talk about it a lot because it’s interesting.
We talk about man bites dog, not dog bites man.
That structurally leads us to believe man bites dog is way more common than it actually is, precisely because it’s so uncommon and the common thing is so unremarkable.
Startups mostly don't die of starvation, they die of indigestion.
That is, from trying to do too much.
When you’re getting close to finding the self-sustaining flame, you can feel the crackle of possibility.
The possibility voltage is strong enough for a crackle of electricity as things get close to touching.
Sparks are showering out.
None of the sparks have caught yet but at some point, with enough tinder nearby, one of them certainly will.
The best magic tricks get even more impressive when you know how they work.
Things with exponential growth start small.
If it has exponential growth and a high ceiling, it will grow to eat everything else, no matter their growth rate.
Compounding beats linear over time, no matter the multipliers.
But everything starts small, and exponential things are no exception.
In fact, at the beginning, exponential things often have less growth than linear things.
When you see that, you might think “it’s too early.”
When the soil has been overturned–tilled–by a disruptive technology, that’s the time when the majestic oaks of the next generation will take root.
That’s when they will start growing as seedlings, even if they don’t yet look impressive.
Copy and pasting strategies from one context you're familiar with to another context you’re not familiar with often doesn’t work.
Often there's water you’re swimming in in the last environment that you didn’t even sense, because it was everywhere.
But sometimes that water is what makes a given strategy work.
You didn’t see the water before, so you won’t see its absence in the new environment.
Constraints create coherence.
The constraints set the shape of the jello mold; the swarming exploratory competitive behavior is the slime mold that grows into the mold’s shape, filling every niche.
Sometimes the answer to find new breakthroughs is to operate within different constraints.
If you were in the US trying to compete with frontier models, the answer was always to try to get H100's.
In China, they knew they couldn't get them, so they took it as a hard constraint... and that led to finding a radically more efficient model.
A couple of useful frames for game-changing thinking:
How can I best make use of the constraints I was dealt that others weren’t?
How can I change the constraints in ways that others can’t?
Emperors can’t get good sparring partners.
True sparring is a great way to improve.
Sparring is disconfirming evidence that can’t kill you.
If you don’t spar the disconfirming evidence will come from the external swarm battering you, which might kill you.
Someone who is obedient to you cannot spar with you except performatively.
They all will tell you what you want to hear.
Loyalty is a two way street when it’s healthy.
What Sarumans mean when they say they want loyalty is actually something more like “obedience.”
Sarumans believe in heroes.
Radagasts believe in systems.
Sarumans only consider the direct effects.
If the direct effects are good enough, they do it.
Radagasts consider the indirect effects.
If the indirect effects are negative enough then they don’t do it, even if the direct personal effects are extremely positive.
Both Sarumans and Radagasts are suspicious of techne.
That is, the status quo, bureaucrat energy, the cog in the machine, the corporate zombie.
Saruman energy is “don't bother thinking about the implications... let’s be legends!”
The corporate zombie version is: “don’t bother thinking about the implications… let’s just do what the machine told us to do!”
Looks like Saruman energy but is entirely craven.
Both can be extremely dangerous.
But the Saruman energy is more likely to put a dent in the universe, and sometimes a positive one.
Lots of people think they could be good PMs.
“How hard could PMing be? Clearly the hardest part is [my speciality]. PMs are just a bunch of meeting monkeys, hyper-organized blowhards. They’re just faking it and couldn’t do the details anyway.”
Anyone who thinks being a good PM is easy doesn't know how to be a good PM.
It's hard, and it's excruciating to do it well.
You can't be a good PM if you aren't comfortable with ambiguity.
Trying to ‘grow’ another person is somewhat paternalistic.
Growth needs to come from within for it to be authentic.
If you don’t grow yourself, you don’t learn the skill of growing.
That is, the skill of discovering and creating who you are.
If you try to tell them how to grow, or what to grow into instead of letting them figure it out, you’re stunting their self discovery.
It feels like the compassionate thing to do is to cap their downside and give them only upside.
That seems like the way to help them grow most fearlessly.
But growth requires experiencing the downside, the error, to grow.
Protecting them from that is nice, not kind.
When building a team, don’t look for the best people, look for the right people.
“Best” here meaning a context-independent measure of quality and skill.
“Right” here meaning the context-dependent measure of applicable quality and skill.
Different contexts require different skills.
You could find the best person in a dimension that is the wrong dimension for this context, which would be the wrong person.
If you’re an individual, you want to be the best–good in a lot of different dimensions, more likely to be the right person for a lot of different scenarios.
But if you’re a team looking to add a team member, you want to find the right person.
A system cannot be both efficient and resilient. They are in tension.
However it is possible for a system to be both resilient and innovative. Those are not in tension.
It requires being adaptable, having internal diversity, and internal coherent belief.
Two types of sense-making that are most effective: building and reflecting on past experiences.
Both those types rely on experiential knowledge, tacit knowledge, knowhow.
Sense-making in a vacuum–that is, based on theory–is likely to be illusory insight.
Conscientious Feelers (in the Myers Brigg sense) are like tofu.
They absorb the vibes of the surrounding context and people in order to fit in with it.
They sometimes keep that absorbed context even when they go to new ones.
An Intuitive (N) person can struggle to make things linear for Sensing (S) people.
But if they're also Feeling (F), they can read the room to adapt it to land in the moment.
I think in the same way a lock picker picks a lock.
I feel my way to the correct answer.
In every interaction I sense and incorporate hidden constraints that my conversation partner understands but I didn’t yet feel.
I then ride the tangle of interlocking constraints I’ve collected until I can feel the lock click.
When it clicks for my audience, I can feel the possibility crackling.
I can do it fastest when I can read the faces of my conversation partners to know if I’m getting closer.
Thinking by surfing vibes.
It takes me quite a bit of work for me to uncover parts where I fundamentally disagree with someone.
That’s because I can typically find points of agreement with just about everyone, and do it naturally and without thinking.
The logic of streaks is tied to prominence.
“The last n times I didn’t do x. Why is this time different from all previous n times so that I would do x?”
Also works in the reverse: “did do x / would not do x”
The larger n, the bigger the difference needed to break the streak.
The bigger the difference from all other examples is best caught with the notion of prominence.
The logic of what you’ve done in the streak so far sets the default, and that default gets stronger the longer the streak is.
To do something different than the default, the status quo, requires activation energy.
That activation energy scales with how different it is.
The activation energy is effectively “why now?”
Lots of real world phenomena emerge from this dynamic.
Slippery slopes.
If you receive a box containing things you don’t acutely need (e.g. objects that are memories) then if you don’t unpack it immediately when you receive it it’s liable to stay unpacked for years.
The first step to win the game is to realize it's a game.
It feels like game over is death.
But actually the game is embedded in an even-larger game.
It's turtles all the way down–at least to your literal human death.
“Death” in the inner game doesn't kill you in the outer games.
Winning moves often risk “death” within the inner game.
Advertising works partially due to the handicap principle from evolutionary biology.
Because advertising does have cost, which means it carries some information signal.
"If the creator didn't believe in it they wouldn't invest in selling it so hard."
A mental model for retconning your thing: what is a plausible story that is compatible with you being a master of the universe.
If you were a master of the universe, everything–including things that superficially looked like mistakes or lucky breaks–would actually be things going according to your plan.
The story doesn’t need to explicitly say you are master of the universe, it just needs to be plausible that you are in everything that happened.
Infinities are large catchment basins.
They’re impossible to avoid because they're so large.
Pick the one that you believe in most, because you will get stuck in one.