Bits and Bobs 3/18/24

Hard things are efficient but brittle.

They're hard to change, hard to adapt.

One-size-fits-all.

Software is supposed to be soft.

Malleable, adaptable.

Software should be like water; fluid, possible to conform to any shape.

Apps are hard, like blocks.

LLMs are fluid, like sand.

Users shouldn't conform to software.

Software should conform to users.

Different types of nouns in English: mass nouns and count nouns.

One test: you use "fewer" for count nouns, "less" for mass nouns.

Count nouns are distinct, individual objects.

Mass nouns are a fluid mass.

Count nouns: Apples. Diamonds. Apps.

Mass nouns: Sand. Water. Happiness…

… and software.

Software should flow.

In the immortal words of Bruce Lee: "Be water, my friend."

Almost without question the vision in your head is an app.

Perhaps DuoLingo.

Maybe Facebook.

I think this is a shame!

Apps are little non-composable monoliths that are only allowed to exist if Apple says they may exist.

Software is so much bigger than that.

Software is alchemy.

It's what allows us to take human agency and extend it beyond ourselves.

Software is a language of possibility.

It allows coordination and cooperation even between creators who have never met, who never even imagined one another.

Software is combinatorial possibility, a fabric of cooperation, emergent leverage for our collective agency as a society.

It's a tragedy that we've locked it inside the box of an app.

AI supercharges everything.

In the future we'll either be locked inside of a box with a god AI…

…or we'll finally escape the box.

People seem to be implicitly assuming the former.

With some urgency we must strive for the latter.

Don't imagine software as it exists.

Imagine software as it could be.

As it should be.

Composability is what allows building blocks to nest within each other to create a whole larger than the sum of its parts.

Composability can be a dangerous operation when the components are untrusted.

The lower the friction, the more dangerous the composition can be in a naive system.

You cannot simply retcon safe, low-friction composability onto a system after the fact.

I've talked to a number of very smart folks building frameworks for 3P agents to cooperate.

I asked them how their security model worked to allow untrusted 3Ps agents to participate.

They told me "oh we'll figure that out later".

To which I replied, "...no you won't!"

The only way to have a system with safe, low-friction composition is for it to have that property from the beginning and then never lose it as you grow and extend the system.

The web has maintained a security model of "clicking a link should never be directly dangerous" and "same origin isolation" since the beginning, allowing safe composition.

App frameworks don't have any similar general purpose primitive.

We'll need to flip the current privacy paradigm on its head.

By doing so we'll redefine what software means.

This will unlock a whole new galaxy of software that was previously inconceivable.

We'll shift the power balance back to users.

We'll put the alchemy of software back in the hands of humans.

The browser was unlike other applications because it was fundamentally open-ended.

For the first decade of the web, "What's the use case of a browser" had no clear answer.

It was like the Matrix: you couldn't be told what the web was, you could only be shown.

Anything that is open-ended doesn't have a killer use case.

This is why platforms don't have killer use cases.

The web and browsers allowed safe, zero-friction teleportation between, and composition of, untrusted content you'd never seen before.

The web enabled whole new classes of software that were previously unimaginable.

Decentralization is expensive; it massively increases the coordination cost of adding coherent improvements to the system.

Even if the underlying components are decentralized, if the top layer (the way that all users access the system) is heavily centralized, then the system as a whole is effectively centralized.

Aggregators aggregate the consumer experience, creating a very small number of extremely empowered linchpins in ecosystems, no matter what the lower layers do.

Don't bother decentralizing the layers that don't matter to your problem.

Crypto decentralized the "database".

But what we need to decentralize is the notion of the app itself.

"Either you use this, or you don't" becomes the only choice.

AI's cost structure threatens to exacerbate that dynamic.

The web reduced the floor size and cost of the minimum viable experience, which allowed even small experiences to be viable.

The cost structure of AI messes with what we expect.

We've come to assume that consumer experiences will be free.

But LLMs have a non-trivial marginal cost.

Orders of magnitude below human labor.

But orders of magnitude above generic compute.

By default, this will lead to faster aggregation, as only aggregators can eat the costs by cross-subsidizing from their money makers.

The free model leads to one-size-fits-all, hard software.

Sometimes there are use cases where you want 30 cents of computation to be put into the task.

But sometimes there are important, challenging tasks where you want 30 dollars of computational work to go into the task.

With the "everything is free" model you get one-size-fits-all software and you can't choose.

These models might be rules, policies, predictions, ontologies, etc.

But the real world is fractally complicated.

The closer you look, the more nuance you will see, another level of intricateness.

This nests fractally and for practical purposes never ends.

Each rule costs energy to construct, test, and maintain.

The value of a rule is defined by the volume enclosed.

The cost is defined by the surface area.

But the fractal nesting leads to many orders of magnitude more surface area the closer you look for the same volume.

This is effectively the shoreline paradox.

This is one of the reasons that anyone who has ever tried to fully capture some subset of the real world to any degree of fidelity in a symbolic ontology within some domain invariably gives up roughly 80% of the way through.

The costs scale significantly faster than the value.

However, at a certain point, you can get away with a fuzzy approximation judged by some other system.

In systems made up of humans, that's often just what a generic employee could reasonably make a call on.

But that kind of "reasonableness" approximation used to be hard for computers.

LLMs do a great job with straightforward reasonableness approximations.

The result is that before you get too many layers of fractal complication deep, you can simply reduce it to asking the LLM.

"This cake recipe calls for 5 tablespoons of tabasco sauce. Is that reasonable?" would be the kind of edge case that is very expensive to exhaustively capture in a formal rule system but is easy if you can just ask an LLM.

It's not some black box built by someone else.

It's software that you've evolved and grown yourself.

You've been playing with it, developing your own "theory of mind" to predict what it will do, changed it to behave closer to what you want.

Situated software is active software, not passive software.

You're in the loop with the software.

This solves a confusing riddle:

Why is ChatGPT so compelling, but most LLM-powered UIs are underwhelming or hard to trust?

In a ChatGPT conversation, you're in the loop with the LLM; you can see how it's responding and tweak your interaction.

In traditional UIs you're on the receiving end of a final prediction with the LLM somewhere deep inside, impossible to access.

First, they have to have a spark at their center.

This spark is a new use case that is minimally viable.

This means its expected value is greater than the expected cost for some critical mass of users.

Second, they grow quickly based on the speed of their network effects.

Some network effects have a weak gradient, and some have an incredibly steep one that accelerates with momentum.

For example, open systems typically have a much steeper network effect, all else equal.

An open system with multiple interacting network effects might be described as having "network effects out the wazoo"

Even if a proto-system has a very steep network effect, it doesn't matter if there isn't some minimum-viable spark to get it going.

This is the "way in" to the ecosystem.

A spark has to stand out from the alternatives to get adoption to start.

If the alternatives are all very good, then that spark has to be very cleverly selected and invested in.

This is because there are a very small number of small sparks that might plausibly stand out, so you must carefully search for and build one.

But imagine that some new technology has shown up that has thrown the normal value landscape into a bit of chaos, with everyone trying to figure out what good looks like.

This means that finding a spark that stands out is more likely. There are many more plausible "ways in," any of which might work.

The longer your runway, the more open the system, and the more chaotic the alternatives, the more likely that some way in will activate, and the less important it is to have the way in.

Folksonomy was a term in the early 2000's for systems like Flickr's community tagging system.

I imagine I'm one of the only people on earth who have uttered the term in the last few years.

The general idea is that the system allows any participant to create a tag; there's no top-down structure for the ontology.

This gives a fully open ended system.

But open-ended systems create a chaotic, diffuse baseline.

Folksonomies typically work because the UI adds optional preferential attachment.

That is, users can optionally choose to adopt someone else's idea.

Concretely, this means that when you create a new tag, it first shows you a search result of related tags that already exist, along with how popular they are.

The more popular a tag is, the more you'll be willing to adopt it even if it's not exactly what you had in mind.

Adopting a popular tag will help other people find your content.

This setup combines the best of open and closed systems.

It's a kind of collaborative sifting sort of good or useful ideas in that context.

I imagine embeddings could turbocharge folksonomies if applied well.

Many on the core development team of the iPhone had worked at General Magic in the 1990's.

At General Magic they'd explored a lot of futuristic personal mobile assistance technologies, but the company failed to successfully productionize it because the tech was premature.

But the exploration and discovery they still accumulated as knowhow; they could feel the shape of the problem space (and, importantly, the dead ends to avoid) in their bones.

Many years later, the technology had caught up, and the ideas they'd been simmering on for more than a decade finally hit the real world situation that could activate them.

The right spark ignited the latent knowhow they had accumulated.

If it has its own agency to do things behind your back, you don't control it and it's not a direct extension of your agency.

That means that to rely on it you must trust it.

If the AI has a memory and agency and access to your tools, then you have to trust it.

If it's just an amnesic, neutered endpoint you don't need to worry about it; you can treat it like a tool.

Apps are harder than websites to distribute.

This accentuates aggregation.

We're in the app/agg era.

It's possible to get the upside without the downside.

WYWIWYG - What You Want is What You Get

There's a huge category of software that can be clearly imagined but cannot be built because it is not viable.

Vertical Saas is our modern poor man's equivalent of situated software for businesses.

It's one-size-fits-all, but for a smaller niche.

Wizards have to study very hard to figure out the precise magical incantations to make the machines do their bidding.

Now everyone can do that magic!

That's got to change something!

All computer interfaces have to contend with humans having extremely limited swap space in working memory.

In close-ended ones (e.g. messaging protocols) the overhead of decentralization just isn't worth the cost.

Apple in the 90's: Think different. Push the human race forward.

The juxtaposition of their 1984 ad with their current behavior is striking.

Apple today is acting like the RIAA in the 90's.

Similar to if you are supply-constrained, you don't know the full extent of demand of your product.

In The Incredibles, at the beginning Mr. Incredible advises Dash to fit in and hide his abilities.

Not rocking the boat or standing out is of the utmost importance to survive in that context, to be integrated into the broader machine of society.

Later, when the family is assembled on the island and fighting for their lives, he tells Dash: "I need you to run as fast as you can."

Despite the danger of the situation, Dash clearly feels joy in that moment.

Being able to push the limits of your special ability, to see what you're able to do without any constraints, is exhilarating.

This will happen even for urgent-not-important things.

The mundane toil will take every inch you give it.

You need to proactively make space for the important.

Because in the shower the urgent thoughts of the next action stop for just for a moment.

This allows the important thoughts to have room to breathe and blossom.

Another thing that has a similar effect: going on long walks with a dog.

Urgent thoughts will always steal the energy from important thoughts.

So you need to give the important thoughts a leg up and make some space for them, proactively.

Shower thoughts: not just in the shower!.

It gets lost, permanently, without continuing active effort to hold onto it.

An idea in writing is an acorn: a durable seed that can stay dormant for extended periods of time, but can sprout in the future when the conditions are right.

Writing down an idea requires work, to transmute it from a fragile thread into a durable acorn.

But when you do it, you transmit the idea into the future with significantly less carrying cost.

Instead of the idea evaporating away when you get distracted, it stays as a viable self-contained package of information, with the possibility of sprouting in the future.

The primary use case of writing is for you to sharpen your own thinking.

To think better.

To transmute fragile threads of thoughts into stable acorns.

This transmutation requires wrestling with the idea at a deeper level than when it's a thread.

If others find your writing valuable, that's a bonus.

Active threads of thought tether you to the waking world.

If you're trying to keep track of a thing in your working memory (because it's not stored in some external system), that's a thread that will tether you to the waking world.

If you do, you'll likely oversteer.

Just let it drift in the direction you want to go.

For example, an intricate copper sculpture perched at the top of a cliff.

Strength arises intrinsically but is also contextual.

When you have strategic clarity and conviction that is earned, you can set a goal, and employees throughout the org can never doubt that goal is possible or desirable, and they can have fun figuring out how to achieve it.

The conviction gives you wiggle room to experiment.

You have to actually have earned the conviction from a rigorous analysis that sought disconfirming evidence.

To practice, you want to first never give up.

If you give up, you can't grow on that dimension.

A bar to clear, to satisfice.

Secondarily, you want to excel, to grow.

If you don't excel, there's no stretch to grow from.

A bar to maximize.

When you excel, you're in your flow state, your zone of proximal development.

All of your mental energy goes straight into the laminar flow of your development, not the swirling chaos of everyday distraction.

What does 'excel' mean?

Excel means to push beyond the baseline.

The baseline is not what your peers do.

Perhaps they've been training for a long time and you just started.

It might not be possible to excel compared to that baseline, and you'll give up.

Game over.

The baseline is what you have previously been able to do.

By excelling beyond your own personal baseline, you are growing.

And you're growing in a way that is achievable and unlikely to be game over.

Compare yourself not to your peers, but to former versions of yourself.

At the beginning you might be far behind your peers. But if you keep growing, before you know it you'll accumulate enough progress to be in the running.

Error is neither good nor bad.

Error is the variance that gives you feedback signal to improve.

Without error, you cannot grow or change.

The messiness is where the aliveness comes from.

But it's just an illusion.

And in uncertainty, an extraordinarily expensive one, with super-linear costs.

"I made it to look like this, because I was in control." vs "I dunno, this is what it ended up looking like".

Design is about a credible demonstration of control.

Like any living thing, they escape the control of their creators.

Looking weak and being weak are often orthogonal.

Someone who looks macho could be very insecure.

Instead, it changes which direction you grow in from then on..

That's what makes it a subtle but profound shift.

As a reminder, an example figure/ground inversion: "United States" going from a plural noun before the Civil War (emphasizing the states) to a singular noun after the Civil War (emphasizing the union).

Because you'll be dead soon anyway by default, so you might as well have some upside!

Multi-ply thinking is dangerous and hard, because if you get any of the plys wrong, all of the later plys are wrong, too.

Multi-ply thinking is leverage.

Like any levered thing, if anywhere along the lever breaks, the whole thing topples.

1-ply thinking.

Goes fast in the short-term, slowly in the long-term.

People do things that they've seen others do, and that have seemed to have worked for others.

Social systems of norms are alive, and shockingly resilient.

An example: Google's peer bonus program.

It allowed any employee to give a $200 bonus to any other employee for just about any reason.

There were a few limitations:

No more than 12 a month.

The recipient's manager got an opportunity to veto anything inappropriate.

You couldn't peer bonus someone who recently peer bonused you.

Every year, some enterprising new APM would figure out a "triangle peer bonus" to create rings of peer bonuses that skirted the rules.

But whenever an elder APM heard about it, they'd say "that's not cool man," and the new APM would cut it out, no formal rules necessary.

With that little bit of social energy, the thing went from a "I'm cleverly exploiting a loophole" to "I'm undermining a system everyone values."

In biology, it's called basal metabolic rate (basic body processes) vs field metabolic rate (doing useful things).

Large systems grow complex internal processes that require a lot of energy expenditure.

In rockets, every pound of rocket fuel to lift a payload needs additional rocket fuel to lift the net new fuel (luckily, just a smidge less), ad infinitum.

In organizations of people, every additional person requires more coordination, which takes more people power to handle.

The largest organizations spend the vast majority of their overall effort just on dealing with internal organizational dynamics, not with producing externally-visible value.

You can do an end run around their expectations.

This is only true if the person who misunderstands you can't knock you out of the game.

If your manager underestimates you, that's more likely to be a problem!

Especially the more people who are aware of the commitment or see it not be met.

Because for people to trust your word you need credibility.

Credibility is the credit you've earned from the precedent of always meeting your commitments.

Your word, your integrity, is your power.

Shame is the powerful emotion baked into our firmware to make us follow through on commitments.

Human arrangements implicitly take advantage of this fundamental brain circuit:

Accountability partners in workouts.

A large guest list at weddings.

Someone twisting your arm to take an action item in front of the boss.

When people follow through on commitments, society is significantly more effective.

People who are low conscientiousness will sometimes not perceive a commitment where others do.

People who are high conscientiousness will sometimes perceive a commitment even where others don't.

This can lead to a shame spiral.

That loop can propel you forward to be constantly improving and collaborative… while being toxic for you as a human.

Only a primary use case sells a product.

A bonus doesn't sell a product.

It's a secondary use case.

It only makes a thing that people already want, want more robustly.

A secondary use case is an accelerator, a primary use case is an enabler.

A secondary use case has to not increase the cost of the primary use case.

AI as a bonus.

If the experience is entirely powered by AI, then if it fails in a given situation then there's no recourse.

If it's an experience that allows users to wire together things, and the AI helps them do it, and if the AI fails, a savvy user can open up the side panel and wire it together themselves.

That is, it's worth exerting energy to adopt.

When some meaningful subset of people think your product is special then you have PMF.

If someone who has a vested interest in the thing (e.g. investor, boss, someone whose day job is working on it) says it's special you can't tell if it's actually special.

The mark of it being special is entirely people without a vested interest taking actions that demonstrate they think it's special.

Everyone thinks their thing is more special than it actually is.

(This is only true for post-PMF companies)

A 1-year strategy needs to wrestle with quite a lot of rich detail about specifics.

A 5-year strategy is allowed to (and should!) elide a lot of details to focus on the core narrative arcs.

If you do a 1-year before a 5-year, you need to somehow retcon all of the messy real stuff into something that sounds like it's intentional, and it has to stand on its own.

If you do a 5-year strategy first, and you do it rigorously, then you can focus on the most important throughlines which is much easier to rationalize.

The future obscures lots of details, so you shouldn't have unnecessary details or precision in your very long-term strategy.

A lot of day-to-day trade offs simply evaporate over sufficiently long time horizons.

Then, when you do the 1-year strategy, you don't have to make it make sense on its own.

You can simply point at the 5-year strategy, which does make sense, and show "this is our incremental step from where we are today towards where we're going".

Many orders of magnitude easier!

Whenever there is a big power differential in a discussion, the person with the most power by default wins.

This can quickly mean that bad decisions get made because the local relevant context is only known by the significantly less powerful person.

Knowhow is difficult to transmit, which means someone with relevant intuition on why something is more costly than it looks might not be able to succinctly communicate it.

To the very powerful leader, it will look like the underling doesn't have a good reason, but actually it's just that the reason is hard to communicate.

This is one of the reasons that senior leaders shouldn't get involved in very small details.

Because when they do it will be very hard for the actual right answer to be decided due to the power differential.

And everything the very senior leader might look at has to be handled defensively, in case they do in the future.

Low Process Autonomy / Low Strategy Autonomy

A command-and control style; great when quality is of the utmost importance and the conditions to navigate don't change much.

Everything feels like a sub-component of the larger thing; teams can coordinate easily because everyone has the same process.

High Process Autonomy / High Strategy Autonomy

A chaotic but resilient soup of different orgs that behave like mini-companies, not part of a large whole.

Low Process Autonomy / High Strategy Autonomy

A swarm of sports cars.

Teams have considerable autonomy to make decisions themselves, but can coordinate with other cars easily because of the shared processes.

High Process Autonomy / Low Strategy Autonomy

The worst option.

Teams can't make high-quality situated decisions for themselves, but also teams work in totally different ways so they can't coordinate.

Process is more important to be robustly tolerable than precariously optimal.

There are a lot of different processes that all work reasonably well.

What's important is not which process you pick, just that everyone picks a compatible one.

When driving, it doesn't matter if you drive on the right or left. What matters is that everyone in your area picks the same one.

They also include things like:

Riffs

Documents

Relationships

The more conditions they blossom in (where a reader has an "Oh!" moment), the more robust they are, the more important they are to invest in making into a durable format that can be shared easily.

"I mean I'm optimistic the official plan will work, but... there sure are a lot of challenges!"

A statement that is safe, and also always true, and somewhat subversive.

It allows the other person to glom on the additional challenges they see too.

When used in a private, candid environment, it's a magnet for disconfirming evidence, giving others permission to point out challenges they see.

No one wants to talk out of school about a thing they were told in confidence or know officially.

But when they have just a theory of a thing they will happily share it with others.

Once they're officially briefed, they can no longer share it, and also can't disconfirm it.

So they'll go silent and try to move on from gossiping on that topic if it comes up.

Their refusal to continue gossiping about a previously gossiped topic is a canary it's about to come true.

Someone working with intention can likely find little ripples to accentuate into larger waves to ride, because there isn't a strong opposing force, just randomness.

A very different situation: the bottom-up entropy of any large organization… but with a very empowered senior lead with a specific, and perhaps on some topic wrong, perspective.

Those are way, way harder to make good things happen without convincing that key person.

You can't find waves to surf to good outcomes, because they all emanate from that one person's perspective.

When you don't see a path to something great, you're an order of magnitude less motivated.

Large organizations, by their very nature, have constrained paths to greatness.

The naturally-occurring eddies of interpersonal coordination slow down any clear shots.

That means that in large organizations, there are few obvious paths to greatness, and it's harder to get intrinsically excited and in your unstoppable flow state.

That's not to say that the person can't do it, just that they can't see their potential in that context.

The board of virtual stickies kept remarkably organized, emergently.

I noticed the implied algorithm we were all following was simple:

1) Nudge around things to have less overlap.

A kind of simple force-directed layout

2) Move things where their new neighborhood is clearly a closer semantic fit than their current one

It has a kind of simulated annealing kind of behavior.

At the beginning it jumped around a lot, but later it became mostly small tweaks.

I bet you could automate the second part of the algorithm to some degree with embeddings.