Bits and Bobs 6/3/24

Lorem ipsum is placeholder text used when mocking up print layouts.

It used to just always be the same text.

But now LLMs can generate plausible placeholder text for any situation.

And not just text, it can give you placeholder structured data.

"JSON representing a TODO list for a college student"

"CSV of data for home listings in Berkeley"

A technique for innovation is "lateral thinking with weathered technology"

You use only battle-tested, cheap, resilient components.

The combination of components is where the differentiation comes from.

Imagine that you're trying to create a new thing using this playbook.

One of the components you're relying on becomes more popular or gets significantly more investment.

Naively, this seems bad: your idea is now less differentiated, because more people know about the components.

But actually it's good: your differentiation was not the components but the combination.

Your combination is now more viable because the components have been derisked.

As long as your combination is novel / surprising / requires specialized knowhow to assemble, you're in a better situation.

This happens when there's an insight: a thing that, once pointed out, is self-evident and cannot be unseen.

This can happen for individuals… or for whole swarms of people all at once.

For example, imagine if there's a slow, plodding, inevitable secular change happening.

At each time step, nothing obvious is happening, so no one really notices that anything has changed.

But then some prominent event happens.

Maybe there's some miracle that draws people's attention.

Maybe an entity that is prominent and widely respected makes a big stink about it.

In that instant, a whole swarm of people go from "huh?" to "duh!" and the landscape of things that everyone thinks is possible changes.

Before, if you wanted to make use of that secular trend, you'd have to convince everyone that it was happening to get them to invest.

You'd have an uphill battle.

But after, you can take it for granted that most people already get it.

There's more "weathered" components you can take for granted.

Apple is apparently going to talk about their use of confidential computing for AI, which could raise the prominence of confidential computing significantly.

You push it.

Instantly you hear a boom a mile away... and then a bunch of sirens.

Would you ever push that button again?

What about other buttons like it?

Coherence is the ability to do, to stand out from the noise, to thrive.

Adaptability is the ability to react, to change, to adapt, to survive.

You want both coherence and adaptability.

Coherence is easy to see even quickly at a distance.

Adaptability is subtle, indirect, harder to notice unless you look carefully.

Too often people only go after coherence, because it is more legible and obvious.

Coherence without adaptability makes a thing brittle and prone to death when the context changes.

Lo-fi things are superficially messy

The mess requires the viewer to mentally choose what to focus on, to unpack the meaning.

But to transmit it outside your mind requires distilling it.

Crunching it down to something more specific.

Transmuting it into a small package.

A seed of a thought.

That process requires you to hold the idea much more concretely, which helps you find its flaws better.

This is one reason dialogue is necessary for rigorous thinking.

If it's fully in your mind you never need to distill it deeply enough to interrogate it.

Dialogue is useful not just because you get the benefit of adversarial feedback and sources of disconfirming evidence.

It's also useful because it forces you to distill your fuzzy idea vibes into concrete arguments.

Dialogue and conflict from a peer reviewer makes it stronger and more likely to survive the real world.

Another example of where dialogue is necessary for rigorous thinking.

Because what's the one answer everyone can assume everyone will consider the default?

Whatever we did most recently that was good enough.

You can go with the flow and do that automatically.

No need to think or stick your neck out... to others, or to yourself.

Every other option has to show that it's sufficiently better than the default to be worth the risk of a new thing not working.

"If it ain't broke, don't fix it."

That is, you'll lose all of the investment.

There's another approach that you can do if you don't have to worry about running out of runway.

Put a small portion of investment into a diversity of very cheap acorns.

Then, wait to see which ones start to sprout, and put future investment into those.

This allows you to waste less of your investment in non-viable ideas.

Planting acorns is a thrive technique, not a survive technique.

If you don't have the runway to survive very long, the acorns can't help you even if they do start sprouting, because it takes time for the acorns to grow into something valuable.

If you have taken the time to discover and become in tune with hidden dimensions, you can find waves that no one else can see and exploit useful gradients in the other dimensions.

However, if your boss sees magic they don't understand, they are likely to conclude "they just got lucky".

Successfully deploying magic in an organization is unlikely to earn you credit.

It may even make you a threat.

"They keep on having unnatural amounts of luck… they must be cheating!"

Like you're saying there is no magic solution.

But actually there often is!

Have a swarm explore the solution, don't try to do it yourself.

The silver swarm framing is hopeful.

The problem was not the "silver", the problem was the "bullet".

Swarms are hard to reason about, hard to control, so we ignore them.

But swarms are common and they are powerful.

Attracting and nurturing a swarm is very different from building.

Swarms give you a hidden wave to surf.

Allowing you to do what looks like magic.

To find the swarm, to surf it, it just requires giving up control.

But that control was an illusion anyway!

You can do real magic if you let go of the illusion of control.

The definitions of each word are made up of other words in the dictionary, which have all kinds of loops back on each other.

A tangled web of tautologies!

But clearly out of this tangled web real meaning is bootstrapped. How?

Because it's not a single set of tautologies, it's a web.

The relationships in that web, which words are more related to others, creates differential meaning.

It's the differential structure and density of the web.

This is, incidentally, not too unlike how embeddings are calculated and formed.

Apps are hard. LLMs are soft.

LLMs aren't going anywhere, they're a new fundamental primitive.

Everything hard will need to melt to interact with the softness of LLMs.

A mini app is still an app-like mental model.

Too hard for this new era.

Sand is hard but only at the micro level.

At the macro level it's soft.

The hardness is on such a small scale that it's not hard at all.

We need "apps" that can flow like sand.

The game is the emergent kayfabe taking on a life of its own; terraforming ground truth to its own internal logic.

No one wants to play the game.

If you don't play the game, then you are knocked out of the game.

The game produces nothing of value except friction, heat, and the appearance of work.

Each additional ply is an order of magnitude harder to convince people of, and even if you do it quickly becomes stale or evaporates from their understanding.

So if you're scurrying around trying to get everyone in the swarm to hold the same multi-ply idea, you'll never catch up.

Like filling a leaky bucket with a small ladle.

Consensus swarms only ever really think coherently in a single ply.

If a swarm has a multi-ply insight it's because a rebel within the swarm had the clarity themselves and took the action and it worked well enough (or they had enough credibility) that others joined in.

The cathedral is a massive, beautiful object with a coherent vision.

It requires massive amounts of time, capital, and coordination.

Cathedral is top down.

It is a high-fidelity manifestation of a particular author's differentiated vision.

The beauty and innovation comes from a highly constrained and opinionated plan.

The constraints are a cage on possibility.

Selecting out of the infinite things that are possible a single, extremely high fidelity option.

The bazaar is a sprawling, organic, emergent swarm of innovation.

The structure is loose or even non-existent.

"At midday in the central square people congregate to trade"

Bazaar is bottom up.

All of the beauty and innovation comes not from the intricate web of constraints, but from a very small set of enabling constraints that everything else emerges from.

The constraint is a platform.

No internal selection pressure for a particular outcome among others.

Both the cathedral and bazaar approaches can create massive amounts of value.

A true cathedral approach is extremely rare and hard to execute.

If any point doesn't execute properly, or turns out to be non-viable, the whole plan is at risk.

If multiple people are executing on different, non-coherent cathedral plans, you will achieve nothing.

A cathedral plan is default-dead.

A bazaar plan is self-healing and self-improving; as long as there is a viable ecosystem, the swarm will automatically create antifragile outcomes… just ones that don't necessarily fit any grand plan.

But the trappings of a cathedral approach are easy to execute and very common.

The performative cathedral playbook.

You get the worst of both worlds:

The constraints and overhead of the cathedral approach

None of the sublime beauty of a high-quality and valuable vision made real.

With the cathedral approach, you'll likely fail, but at least you can show to others that you tried.

With the bazaar, if you fail, observers might say "what did you even do?"

You have to defend your approach to someone else, and if they don't get it, you could be knocked out of the game.

The bazaar requires multi-ply thinking to even grok in the first place.

Organizations are chaotic and busy; few people have time for multi-ply thinking.

This gets even more true as you go up the hierarchy and leaders' attention gets more and more distracted.

With the cathedral approach, you'll have heaps and heaps of paperwork, blueprints, coordination materials.

It will be very clear that you worked hard.

"We don't know if it's working, but you can see clearly that we did something."

With the bazaar, the value comes from the swarm; any individual component of the swarm, or unit of work you did, will look inconsequential on its own.

The bazaar technique is likely to work, but unlikely to give individuals credit for their effort.

The cathedral technique is unlikely to work, but likely to give individuals credit for their effort.

Each individual in an organization, in order to survive, needs to get credit for their work.

This is part of "the game".

So each individual has a default pull towards the cathedral approach.

Small asymmetries in incentives, if they are consistent across the swarm, create massive changes at the level of the swarm.

This means that organizations are way, way more likely to apply a cathedral playbook than its actual efficacy would imply is warranted.

Each individual is less likely to die within the system by playing the cathedral approach.

But the system itself will be more likely to die if everyone within it uses the cathedral approach.

No one individual will get fired for doing the cathedral approach, but if the organization dies because of lack of innovation, then everyone will be fired anyway.

Innovative + resilient = antifragile.

If they're so powerful why don't we use them for more things?

Because of the incoherence.

Hard to hold onto.

Hard to control.

Trying to control a swarm forces us to confront that we're less in control than we think we are.

More investment typically takes more time.

Time to completing a real-world feedback loop is critical.

More investment means more time means longer loops to feedback means less likely to be viable when you finally interact with the ground truth of the real world.

But also every bit of investment you do before viability might lock in non-viable constraints.

When you invest incrementally in a thing post-viability, the incremental addition is more likely to be viable (you can quickly see if it's not and unwind).

But if the incremental investment comes before the thing is viable in the real-world, you can't tell if the incremental investment was viable or not… and by the time you do, it will have ossified and become load-bearing and will be more difficult to unwind.

We act like they're aligned, but why should that be?

Our focus on superficial cathedral style tactics are more a comforting illusion of control than an effective and serious tactic to maximize value.

"This is what we can measure so this must be what matters".

This is backwards!

Wrinkly things are easier to describe not in their final state directly but by describing the system that generates them.

Most wrinkly things look chaotic and incoherent at first glance… but have hidden generative patterns.

If there didn't have some kind of highly efficient generator pattern they'd just be random noise and would dissipate and erode over time.

Stable variance that persists has to have some kind of generating system behind it.

Indirect benefit is very hard to capture formally, the space and effort to describe it (the surface area) expands faster than the value itself.

Because it is wispy and fractal, not smooth.

Not smooth.

But it's even worse than that, because it's constantly in motion.

It would be hard to do if it were frozen in time, but it's impossible with motion.

The desire to do precisely what you want and the desire to coordinate with each other are in tension.

Here coordinate means do a thing compatible with others.

The power of coordination has network effects that dominate as it gets momentum.

The more people you could plausibly collaborate with, the more the pull towards coordination will dominate.

You can do anything you want… if you live on an island and never interact with anyone.

We can do far more collectively than each of us could do individually.

Agency makes you stand athwart the entropy pulling towards background noise.

It goes uphill.

It separates you from the things around you, you stick your neck out.

But if others also go in a similar direction as you, if you stand out of the background noise together, then you merge together, by standing out from the rest together.

When people see independent things moving coherently, they assume they are actually one thing, and treat them accordingly.

The system allows experimentation, agency, discovery, not monoculture and centralization.

Systems have different stages

In the early stage it will be more about hill-finding and novelty generating.

In the late stage it will be more about hill-climbing, centralization, optimization.

Hill-climbing takes grit, dependability, and efficiency.

Mass-produced.

Hill-finding takes taste, craft, and cleverness.

Artisanal.

At the beginning of an era everyone focuses on hill-climbing.

At the beginning of the era you need hill-finding.

You need hills to climb.

It's only once you're on a hill where you haven't yet reached the top that it makes sense to climb.

Someone who has focused on hill-climbing will forget the deep magic of hill finding.

This manifests directly in the role of the PM today no longer being artisanal and instead being more about program management.

We're in the hill climbing phase of PMing.

PMs don't decide anymore, the metrics do.

The humans are subservient to metrics.

The hill -finding artisanal PMs didn't go away.

They're hanging out within bureaucracies they hate, biding their time.

Thinking they're alone.

Ready for the paradigm to change, not even aware they're missing anything.

"I guess this is just the way the world works, no use fighting it"

Don't try to create alignment, find pre-existing alignment and then surf it and accentuate it.

People love momentum, so when you help them go further in the direction they're already roughly going in they love it.

Differentially invest in the things that are already going in the direction you want to go in.

If you're like Maxwell's Demon with a swarm, you can do magic.

You'll harness a wave of asymmetric energy that is invisible to others.

You'll apply your judgment to select the sub-movements that you want to surf.

You'll create a coherent wave to surf out of the directionless background noise.

Solve their object level problem to get the credibility or take multi-ply openings as they present themselves, but never push on it.

Otherwise they'll think "He always seems to make things more complex than they need to be, or solve problems that are not the ones that are important."

This puts you in significant danger of being knocked out of the game in the future.

When you have the right multi-ply sponsor, you don't have to waste as much time on the object level tap dance.

Being able to do multi-ply thinking in a context too early in your career is a curse.

You'll be able to see multi-ply approaches your boss can't, but you won't be able to convince your boss of them, and to them it will look like you aren't doing anything useful and (maybe) just getting lucky.

You'll be surfing a wave they can't see, so they'll assume you're just lucky.

Every single assemblage of humans is dysfunctional in some way.

Some more than others, but there is no stably non-dysfunctional assemblage of humans.

Don't rail against the dysfunction, figure out how to survive it (or, ideally, surf it) to make things happen that you're proud of.

When you acknowledge that a situation is dysfunctional and you're going to try anyway, you switch from a victim mentality, making excuses, to an open mindset, a growth mindset, curious and scrappy and on a path to self improvement

Of course all of this is modulo some situations are so dysfunctional that the only move is to leave.

A supercritical state is one where it appears stable, but is actually on the cusp of a catastrophic cascade.

A canonical example is the Emperor Has No Clothes.

Every adult is afraid to embarrass the emperor (or themselves).

But everyone can see with their own eyes the emperor is naked.

All it takes is one spark (a child's laughter) and the whole equilibrium shatters in an instant.

When you can tell a system is in this state, plant some nucleation sites.

These nucleation sites can be cheap, a diversity of acorns of ideas.

Nucleation sites are schelling points, little seeds that ideas can congregate around.

For example, in an organization maybe share a meme-able document that is subtly subversive and suggests a different approach.

Once the kayfabe shatters and people can acknowledge the value of the previously subversive idea, everyone can by default congregate around the idea you previously seeded.

Execute without thinking, there's no time!

There's barely time for a single ply of thinking, let alone multiple.

This is the hypergrowth playbook.

It's one we take for granted in silicon valley, unquestioned orthodoxy.

When you're doing something deeply differentiated you have a bit more time to think an additional ply or two.

But bigger can make you more likely to die:

More sclerotic.

Harder to coordinate and adapt, since coordination costs rise superlinearly with size.

Trapped by a Lilliputian web.

More leverage.

Leverage allows efficiency… but also more risk.

If the underlying source of leverage changes, you lose not just it but all of the things that are levered out from it… which can be a lot!

Higher metabolic rate.

Just staying alive is more and more expensive.

If conditions change and resources are scarce, it might no longer be viable.

More and more of the energy goes into maintaining than creating.

A thing can grow so large that it extinguishes itself.

Every living thing is downstream of billions of fortunate mistakes that just so happened to have been useful and be conserved.

For example, a minor genetic mutation, or an accidental experiment.

Most mistakes go away quickly:

If they are actively harmful, they are immediately selected out.

If they aren't actively useful, they erode and fade back into the background noise, averaging away into nothing.

Useful things are conserved

This is true no matter how they originally came to be.

Useful just means they help out the entity that is making the selection decisions.

In artificial selection, it's for example the farmer doing the animal husbandry.

In natural selection, it's the ground truth of natural selection.

Natural selection is fully emergent; it must be so.

Things that are more likely to survive or replicate are more likely to be more prevalent in the next time step, on a fundamental level.

When you stare into a mirror, even if it's wrong and you can tell that it's wrong, it causes you to self-reflect.

Even if you're self-reflecting to say "wait, the reflection I'm seeing is wrong, what's wrong about it?"

The more self-reflection we do, the more insights we have about ourselves.

The imperfect mirror is just another excuse to self-reflect.

All anyone else will give you (the ones who have a transactional relationship, at least) is constraints.

The constraints will consume all of the space you give them and box you in.

Tied down by a Lilliputian web.

You being energized is valuable to the company even if no one asked you to do it.

No one will tell you to lean into your spark, but it's the best way to maximize the value you create… both for you and for the company you work for.

But it's also possible to get rigor only from saying yes.

You say yes all the time, but you decide where to apply it by applying your judgment.

You don't smoosh your yeses around like peanut butter, you invest differentially.

Applying a curatorial judgment to the garden of possibility around you.

You can't optimize it beyond a certain point without killing the generative function at the core of value creation.

The generative function that produces the fruit also needs to keep itself alive.

Value-creation is like agriculture.

You want the fruit (the meaning, the value).

But to make the fruit you have to create a plant that can survive and live in order to be able to create the fruit in the first place.

Thousands of years of coevolution got us lush strawberries. But you still can't say " I want the strawberries, not the plant".

The fruit can only exist when produced by a plant, and the plant needs to be able to keep itself alive.

This discretionary effort is a gift to the world.

Done of their own volition.

When you are compelled to do something, you need to clear the bar of the person who has compelled you to do it.

Satisficing mindset.

When you do something for discretionary effort, you're doing it largely for its own sake.

You're already going above and beyond to do the thing in the first place.

A maximizing mindset.

Volunteer effort is hard to control but very valuable.

When you ask a question, it reads like a demand.

"How does this thing you brought up fit in [frame that I'm imposing on you]."

In an environment where people are freely sharing ideas of their own volition, asking a hard-to-answer or finicky question of them chills them wanting to share in the future.

Humans don't like falling through on commitments.

This is especially true when the commitments happen in view of an audience.

Someone asking you a question is an implied commitment to answer.

People feel an implicit obligation to answer a question, even a hard-to-answer, not-particularly-useful one.

However, when you prepend "I wonder…" to your question, you make it about you, not them.

The receiver can choose to answer it, if they find it valuable or generative.

That means if they do answer it, the answer will be of their own discretionary effort.

So it will be higher quality than if it were compelled.

You think you're making it better by giving them a strong signal to improve it.

But you're actually snuffing out the generative function that would generate not just it but other things in the future.

And it's just, like, your opinion, man.

What if your opinion is wrong?

They did it of their own volition.

What right of yours is it to tell them no?

Part of the value of note-taking is transmitting the information into the future.

But another part of the value of note-taking is deciding what information to keep.

If you keep all of the information, you get a cacophonous mess.

Note-taking by hand is expensive, so we implicitly apply our judgment on what is valuable to keep.

You curate useful information to pay attention to later; synthesize insights into a more distilled version.

When note-taking is cheap, that judgment is applied less, and it all becomes cacophonous background noise you have to sift through later to find value.

Cut corners on the unimportant things, while spending your energy on the important things.

Applying your taste and judgment to create rough and ready experimental value.

Getting to viability quickly and then choosing what to extend and invest more in.

Slop is algorithmically generated.

Kino is high art.

Over time we'll see more concentration in the two extremes, less in the middle.

Similar to "Content or Cinema?"

You can't be in the intermediate nuanced view.

You're either in the stable black or stable white phase, like a light switch.

A distinct "click" as you swap from to the other, discontinuously.

Unless you're pushed in a significant way all at once you'll be pushed out of any nuance back to your current position.

The "tip of the spear" playbook works to get a maximally scoped down toehold to then incrementally invest in.

Key to it is the tip of the spear is sharp and a small scope.

A smaller scope is super-linearly easier to coordinate to build than a larger scope.

The tip of the sphere playbook is when you accidentally allow the scope to creep up significantly.

It's easy to add new requirements, hard to take them away.

The spear turns into a sphere slowly without people noticing, and before you know it you have a massive investment of energy before the thing gets to viability… which makes it more likely it never achieves viability.

A consumer app with an ecosystem (a swarm) captured inside of it.

The swarm never knew freedom.

And it's also not necessarily viable outside of it.

A mutual codependency.

It creates something that couldn't have existed before.

But then things that could have existed independently are also pulled into it as it gains momentum.

The aggregator swallows up the free energy in the ecosystem into a less-open environment.

The novelty is captured in a closed-system, not free to be maximally generative.

In this paper they put agents in a simulated hierarchy.

The same LLM can do translation better with a hierarchy of agents made out of the LLM than using the LLM itself.

If you organize agents into corporate hierarchies and give them personalities you get better results

What this implies to me is partially that dialogue really does create insights, that it is a catalytic process.

What if the way we run organizations of people today is actually useful and the thing you go to even in organizations of agents?

Hierarchies show up partially because boundaries show up naturally to contain the superlinear information cacophony growth, and hierarchies are nested boundaries.

You'll find a good enough person to fill the gap off the street.

The last-bad fit for the constraints of what gap needs filling.

But sometimes you have an obvious candidate already (maybe someone already adjacent to the role).

In that situation, if you instead focus on maximizing the value of the person you have in front of you, you can create something very special.

A maximizing mindset, not a satisficing one.

Every company falls into that gravity well, it can only pull away from it for a bounded amount of time before succumbing.

Everything succumbs to entropy in the end.

We create a handle to the idea when we want to be able to retrieve it later.

A handhold to thread ourselves to.

Deciding to name something is a vote that it's important.

Feedback requires loops.

Interlocked within a scale like chain mail.

Then embedded in loops at larger and larger abstraction scales.

An antifragile field of adaptation.

Every loop is constantly spinning, which allows small perturbations to amplify to significant changes.

The field is already rotating, it just needs to offset its angle by a bit, and doesn't need to get itself going from a standstill.

Dynamic equilibrium, not static equilibrium.

Metastable, but constantly whirring right on the edge of a phase transition when required.

The predictive processing model of human cognition implies this multi-scale chainmail of mutually reinforcing and influencing feedback loops.

They assume that the next "frame" of input will be the same as the last one.

When it is, it's unsurprising.

The less surprising the signal has been, the less likely we are to attend to it.

It fades into the background.

It comes down to whether competition or cooperation dominates among the constituent parts.

This is a smooth spectrum, not a binary.

Swarms form as individuals, and then sometimes they decide to cohere and become subsidiaries of a collective.

The collective emerges out of whole cloth, emergently, when the benefits of cooperation outweigh the benefits of competition.

A horse has a world outside of you.

You have to train it to do things.

A horse you have to think about even when you aren't using it.

Is it fed and watered?

Is it confined in a safe location?

A bicycle doesn't go on its own.

If you stop pedaling it stops.

Even electric bikes.

They don't have a mind of their own.

Bikes are about extending your agency.

No one gets upset about "killing" a bike.

Bicycles are dumber than horses.

But they're also easier to predict, easier to control.

They amplify your agency and allow you to go wherever you want to individually go.

Living things have a lossy API layer.

Sometimes they don't do what you tell them to do.

A bike always does roughly what you think it will unless it breaks.

Living things are also ends in and of themselves, you should feel bad if they die (especially due to actions you took).

LLMs are kind of like horses in that they have a mind of their own.

But they're unlike horses in that you don't have to worry about them "dying" through neglect or mistreatment.

Applying the "for the mind" frame to a few other technologies

Apps are a bus for the mind that only services fixed routes.

"I know your house is over there but the bus doesn't stop there. You'll have to overshoot and walk back"

The vision of a centralized cloud-hosted assistant is a self-driving car for the mind.

Removes agency from users.

But full self-driving is a hard standard to reach.

What happens in the cases where its quality isn't good enough to drive by itself?