Bits and Bobs 3/17/25

Very powerful, and if you need a general result in an open lawn it's fine.

But if you're trying to do specific detail work near the flower bed, watch out!

It's a play to increase stickiness by offering something convenient that makes it harder for developers to switch to other models.

A monkey trap.

It also makes it harder for other platforms that abstract over multiple models to offer the same kinds of features, because now users can get them for free… if they just commit to one model provider.

It's OpenAI trying to change the game from stateless, easy-to-swap LLM providers, where the only competition is on quality and cost of the model, and move it into more of an ecosystem value proposition.^[tj]

In a document, you can change any part of it at any time.

In a chat, you can only append new things at the end, you can't change the parts in the middle.^{[to][tp][tq][tr][ts][tt][tu]}

The only adaptive part in a chat^[tv] is at the far edge of it as you and the LLM append tokens.

If the LLM went in the wrong direction in a chat, you have to wrangle it back to a path you like, which might be hard if it has momentum in a distracting direction.^[tw]

If an LLM interprets a doc incorrectly, edit the doc and try again^[tx].

The app model gives us simplicity but also centralization.

Your data can't leave, so your use cases must flow to it.

The app with the strongest gravity attracts everything; a gravity well.

"App" is different from an application.

An application is possibly general-purpose, and stores its data in the filesystem.

An app is a single-use bit of software, with its data model welded shut inside of itself.

This makes apps safe to install but also fundamentally limited in how they can interoperate.

Apps are fixed, single use, little islands.

Even if you could generate apps on demand, it still wouldn't fix the problems of the app model.

Agents take our agency.

Tools extend our agency.^[tz][ua][ub]

Tools that are incidental, human-scale, tied directly to my needs in that moment.

Not too big, not too small.

Not for some average market participant, but for me, in that moment.

Just Right.

It was so hard to do that only people who specialized in it (engineers) could even do it.

But nearly everyone has situated needs where having Just Right tools created on demand would help them live a more whole life.

If you call it "software," lots of people who could benefit from it won't realize they need it.

Another reason to call it "tools."

The tool creator itself is enchanted.

The tools it creates are also enchanted.

Vertical Saas is an extremely powerful business model.

It turns out that if you can create an "operating system" for a given niche of business use cases, it can be extremely valuable.

Within a given type of business (e.g. gyms) the needs are often very similar, so it's possible to make a one-size-fits-most "operating system" for that niche.

But individuals in their personal life are just too different.

Our email, our calendar, our financial data, our health data, all of it is fragmented out in a lot of tools that don't help us do anything productive with it.^[uc]

It's not possible to make a one-size-fits-most operating system for our personal lives.

But LLMs could allow the creation of a suite of tools perfectly bespoke to a given user, to help be the operating system for their lives.

To do it properly would require a bullet proof privacy model.

When you visit a site or an app, you have to go to it.

Email is where that service comes to you.

If you go to a site and get distracted halfway through your task, you forget to complete it.

Email is where all of the origins can come to talk to you.

Email is naturally async and text based.

Each thread can advance on its own timeline, interleaved with other tasks.

Why hasn't email been a center of our personal operating systems?^{[ud][ue][uf][ug]}

It's hard to make email interactive enough.

How much more could you get done if every email were interactive, if you could reply to every email and get a human-like response?

"Wait, back in the day you programmed on PUNCH CARDS??!"

But people did it even though it was like crawling through broken glass because it was the only way to get those valuable results.

There was no better way, so you did it without thinking.

What are the things that are like crawling through broken glass today that we do anyway?

There was a concept not well covered by existing words.

As soon as a word popped into existence (especially from a prominent influencer) that everyone thought was a good word for that concept, it was locked in.

We needed a word, so a word emerged.

"I see vibe coded games as something like folk art, i.e. the value is in the creative process and not the final product . i bet there's way more people who want to make slop N64 games than play them in 2025"

The point of vibe coding is not sharing the result, but the process of creating.

Vibe coding just on its own is fun but a low ceiling.

Vibe coding on your most important data is a powerful unlock.

A place where you import your personal data to vibe code on.

People want it so badly that they're crawling through broken glass to get it.

Using tools like Cursor plus MCP to hack it together.

Someone will figure out how to make it easy, and be something that can make your data come alive proactively, and safely.

Lots of data, no tools, little danger for LLMs.

No data, lots of tools, little danger for LLMs.

But when you have both that's when it gets potentially explosive.

Data is gas, tools are air. LLMs are the spark.

You got practice with figuring out and tweaking software.

But now you just take what some billionaire gives you.

Similar to Toqueville's observation of how practicing democracy in the small develops the muscle for democracy in the large.

I've previously thought of it as the shift from PCs to the cloud.

Someone I was chatting with this week localized it to the shift from the terminal to the GUI.

Software is our digital home.

But we aren't allowed to change anything.

You can't put a post-it note on the wall!

It's where we live but it's not home.

"The real underlying problem is that humans just absolutely love slop".^[uj]

A key battle about who defines what good is: rockism vs poptivism

Algorithms have pulled us more towards slop, and LLMs accelerate that.

At some point though might the pendulum swing back?

This section is riffing on some ideas from Robinson Eaton.

Similar to the old joke about the "average family" having 2.3 kids.

An app is designed for a market not a person.

This is why it's one-size-fits-none software.

You want as many customers to collectively share that fixed cost.

The higher the fixed cost, the larger the market necessary to make a project viable.

A related observation:

If you don't pay for your compute it doesn't work for you.

This is Clay Shirky's concept.

The Coasian theory of the firm is that finding a price has a cost.

You need to create a cell membrane to create boundaries between things that need to transact, and that has a cost.

The Coasian Floor is the smallest size of project a company might undertake.

Software in the app paradigm has a very high Coasian floor.

Feeds decide what we will focus our attention on.

They allow us to sift through way, way more information.

This is a necessity in the modern world where we are inundated with more information.

But the tradeoff of a feed is that the entity deciding where to focus our attention doesn't work for us but rather a company who prefers we just keep watching the feed.

Having feeds also leads to more information being created.

Because we can now handle more information, more information is created.

Similar to the "equilibrium of misery" for traffic.

Add a new lane to a highway, and you get more traffic to absorb the capacity.

Just figure out the dopamine drip.

The Algorithm is great at figuring it out, emergently.

A dystopic outcome: locked inside a box with a super-god AI Clippy… who is dopamine hacking us.^[ul]

Meaningful things, hopefully, not lizard-brain things.

Software should be an extension of our agency.

But it's too expensive to create and so we bifurcated it.

It's created by others to serve others' purposes.

We've been alienated from our tools.

This misalignment has eroded us, because the software that others bother to create for us optimizes for what we want, not what we want to want.

If we fix that fundamental misalignment that is broken it can all be fixed.

The web / social media is an adaptive emergent beast that does not have your best interests at heart.

Be careful how much time you spend in its embrace!

Now with LLMs, it's only sensing that's left.

That is, what things might matter to you in your current context.

That's both a sensor problem, but also a "user having their data in a place they feel comfortable letting the assistance layer see it."

LLMs pull us towards the centroid; this process uses LLMs in a systematic way to skate towards interesting outcomes, not average ones.

The technology exists within a broader social context.

Users have to choose to use and adopt a technology.

A technology that is "correct" but doesn't get adopted doesn't have an impact.

The choice process is a social one.

As Vitalk has put it: The Most Important Scarce Resource is Legitimacy.^[um]

Some of the processes are technical and go faster with better technology.

Some of the processes are social, and don't go faster because the clock speed is the human mind which doesn't change.^[un][uo][up]

A common thought: "Technology moves so quickly now, so this new disruptive thing will be adopted quickly."

But the main constraint might be the social process of diffusion.

Diffusion in a system of bits is faster than diffusion in a system of atoms.

In a fast changing environment, by the time it gets 7 plys deep into the organization it's out of date information.

The org is executing on something that is now wrong.

The streetlight fallacy of organizational attention.

Decisions require making a choice in imperfect information.

If you have imperfect information, it's possible–likely even–that you'll make mistakes.

The fix is not to make no mistakes, it's to make sure mistakes are quick to recover from.

Making decisions knowing you'll likely need to revisit them later helps you learn.

Related to the friction of working with bits vs atoms.

Looking at a big, audacious, amorphous long-term goal, and figuring out a thin slice that is concrete and does something useful that takes a step in that direction.

If you can run this loop very quickly, you can roofshot your way to the moon.

Leadership is not just summarizing the bottom-up beliefs of all of the constituents and doing the average.

That average, consensus view of various plausibly viable options might not even be a viable option itself.

The bottom-up insights provide ground-truthing, a set of constraints to operate within and balance.

The same is true for setting a product vision and UXR; UXR doesn't tell you what to do, it tells you the constraints of what's viable.

People have to queue, raise their hands.

In real collaborative contexts, conversations can slip into smaller sub conversations without people even realizing they're doing it.

They just kind of turn to someone and then if it's too loud they edge away from the main conversation just a little bit.

A percolating sort that creates and extends and attenuates naturally without anyone in the system individually having to think about it.

Totally fluid and emergent, allowing surfing the efficient frontier for a set of conversations in that moment.

Most tools have no affordances for it, or only weird top-down "breakout rooms" affordances.

Even tools like Gather require people to do something explicit to fork or rejoin a conversation, it's binary, not continuous.

A closed system can't be infinite.

An independent test suite that makes it easier to have multiple implementations.

The implementations themselves might not even care to be interoperable, but if the community wants to use it as a schelling point, a pirate testing suite can help create and maintain alignment emergently.

To appeal to more people it moves towards the lowest common denominator.

If it's too welcoming it auto-extinguishes by diluting itself.

If the movement has too many purity tests then it auto-extinguishes by not reaching sufficient scale to matter.

If the system dopamine hacked me to want that thing, does that count?

If any of the competitors in the market is assuming it's a winner-take-all market, then everyone else must, too.

Because if it is a winner-take-all market, then anyone who doesn't play to win will be knocked out of the game.

It might very well turn out to not be a winner take all game, but it's safer to play like it is, just in case.

Everyone playing like it's a winner take all game (e.g. deploying significant capital to get usage) makes it so anyone who doesn't play that way can't compete.

So everyone believing it's a winner take all game effectively makes it behave like it is one.

A few apple-related emergent process examples:

They don't have a global sense of state, only local state being progressively summarized at higher layers of abstraction.

They can get a sense of average but not count.

System 1 in our minds also has this average-but-not-count ability.

Most of them don't work and fade away, evaporating quickly.

But every so often one that is better is found; it has a consistent ground truthed bias that makes it stand strong amidst the noise.

This is conserved, naturally.

So the system is jiggling constantly trying everything, it's just the only ones that cohere and stay around to be built on are the ones that are effective, so they're all that we see as macro-scale changes.

This is similar to how the principle of least action arises in physical systems, as I recently learned from this mind-bending Veritasium video.

"Ditch trends. Avoid playbooks. Reject boredom.

Embrace the niche to rediscover authenticity."

Very Rockism vibes!

Never trust a shark.

When they act in a way that takes advantage of your assumed good faith, the shark would say the fault is on you.

"How dare you double cross me and abuse my trust!"

(Unapologetically) "I'm a shark, sucker! That's what I do!

By asking questions sometimes you can nudge them and they unpack their own underlying problem.

Meanwhile you're engaging so they feel heard and held.

Because they have a mental model of how a given thing should work and their collaborators don't, so unless they have the same very high fidelity mental image of what it will be, they'll be hard for everyone else to collaborate with.

You commit a form of fundamental attribution error, with an implied "that other person is actively trying to thwart me."

Which is almost never true.

The worst thing you can do to a victim when helping them is give them the "yeah the world is out to get you".

The best thing you can do is coach them to see the constraints on themselves and others and help them get to a solution mindset, with an iterated series of successive "why" questions about how the other parts work.

Within an organization, change requires a period of ambiguity.

Moving from a thing that is known to at least be coherent to something that is new.

A bit of a leap of faith.

Leap of faiths require trust.^[uw]

Sometimes those ideas just distract and randomize and make people think less of them.

Internal processors in contrast might self-censor ideas that would have turned out to be great.

Every so often an idea that is burped up is game changing.

"What is that novel thing I see in the distance? Oh, it's that thing I already know!"

Gives you a little roller coaster bump in your stomach that's enjoyable and feels meaningful.

"Gilded age" is one of those terms that we learn in school and think of as an atomic unit, just a meaningless set of sounds that we associate with that time period.

But when you unpack the construction of the phrase you understand the deeper insight.

Over time the process can create gilded turds.

When that force is turned on communities, you get a bunch of people so upset with the system they want to smash it, even if doing so would hurt themselves too.

Adults might call that "flow state" or "being in the zone of proximal development."

Optimize to feel "that sparkle feeling" as much as possible.

The sparkle feeling has a half life that is independent of the scope of the achievement.

Stretch too hard and you'll often be frustrated and give up, and when you do succeed you'll feel it more acutely but it will fade just as fast as it always does, so you won't feel much more of the feeling.

If you don't push hard enough the acuteness of the feeling will be less.