Bits and Bobs 11/24/25

Another: controlling the pixels the user sees has an order of magnitude more leverage.

The combination gives an order of magnitude more strategic leverage than either alone, but both are very powerful.

LLM API providers don't have the state, it's stateless!^[ey]

They also don't control what pixels are on screen.

That's why OpenAI is moving so aggressively to own the vertically integrated consumer experience, complete with tons of state.

It's impressive they can get them that quickly.

It's cool, but it's a low ceiling.

These are little widgets, micro-apps with no data.

None of them have any differential pricing power, since the models are practically commodity.

But they do all have a shared interest in the inference cost not dropping to zero, to recoup their capital investment in training.

This is not too dissimilar from the Unix Wars.

There were a small number of extremely expensive Unix options, in a stable equilibrium.

Then Linux showed up, a high-quality free option, and it totally destroyed that equilibrium.

Earlier this week Claude refused to comply in a hilarious way.

I had a recipe named "Five Cheese Mac and Cheese".

But the recipe only actually listed four cheeses.

I asked Claude to add the ingredients for the recipe to my shopping list.

It refused, because there were only four cheeses, not the five as claimed, so something must be wrong.^[ez]

It's the best model, and still not transformatively better.

This is what it would look like if we were hitting the top of the s-curve.

Even if the models are actually much better, we already have more than enough quality for many tasks.

Similar to the logarithmic improvement in the number of triangles in a 3D object.

Past a certain point the cost keeps on going up and it just doesn't matter.

You can electrify things that used to not be able to move on their own, making them dynamic, almost alive.^[fb][fc]

Competing to get better quality for cheaper.

But ultimately it will just become a commodity.

No one will care where their tokens come from, if most providers have similar quality and don't store state.

One place this metaphor breaks down is that power delivery infrastructure has a natural monopoly in a way that APIs don't.

It's just a provider of a commodity.

Your LLM provider should be the same--although unlike electricity which has a natural monopoly, the LLM provider should be easy to swap out.

They wilt!

Model quality is getting saturated.

The differential quality comes from the harness now.

It's gotten way harder to do a vibecheck when they're all so good.

Long-running agentic toolcalling is where the incremental quality is visible.

But most uses just don't need the quality.

Andrew Ng has noted in the past that the quality jump from adding a good agentic harness to GPT3.5 was higher than the quality jump to GPT4.

If the harness is more important than the model, but the harness is easy / cheap to build and reverse engineer, that implies different strategic outcomes.

By wrapping the models and standing on their shoulders you can get further, with way less capital–but also less moat.

Step changes in those have profound implications!

Massive projects you can't do with just the squishy "muscle" of associative reasoning.

You need to give it external structure.

That allows you to page in and out things into the "CPU registers", which there are only a very small number of!

First, Anthropic's research shows it's remarkably easy to poison a model of arbitrary size with deliberately chosen malicious training data.

Second, if there's a model that everyone uses that has a subtle but consistent bias, that bias at society scale could lead to significant society-scale impacts.

The Ouija Board effect again: a consistent bias in a noisy signal, at scale, leads to large emergent macro effects.

Feeding my little remote Claude Code instances with little thoughts to keep them happy and productive.

But unlike a tamagotchi, at least they're producing output, and it's not just a game.

You feel like you're right on the edge of it working and don't want to lose the streak / mental energy.

But with vibecoding I can get back in the mode in a fraction of a second.^[fe]

Uh oh!

That is, do the hard thinking up front in design and research and speccing, and then the execution is mostly small questions.

If the LLM asks you big questions constantly, it quickly gets overwhelming.

You are constantly needing to page back in significant complexity, thrashing between workstreams.^[ge]

It's overwhelming and exhausting.

The basic approach is to start a normal interaction routine and then abort it.

A couple of examples of this:

Here's my mental model for what's happening.

When you start a stored social routine, your brain expects to simply execute it.

When you pull the rug out, the brain fritzes.

The prefrontal cortex is put to sleep but now you need to actually think, so you just go along with whatever was suggested.

In that stunned chicken moment we're extremely suggestible.

Uh oh!

This week I learned about "account ripening."

Bad actors need fake accounts that look real.

This used to be expensive.

LLMs make it orders of magnitude cheaper!

So if you have a long conversation where you say "not X" over time it has a high likelihood of thinking just "x."

It was always a powerful rhetorical trick, it's just most people hadn't noticed before.

But now it's kind of ruined by everyone knowing it's an AI tell.

Before, good rhetoric often co-occurred with good thinking.

But now LLMs allow applying good rhetoric to half-formed ideas, which makes signal of rhetoric quality less powerful.

Most models have a baked in reward function.

Assistant games try to infer what the user wants to do, based on their actions, and then help them do it.

Each action the user does helps update the model's priors about what the user's goals might be (or definitely are not).

Instead of a baked in reward function, they have a floating reward.

They're much harder to do, but potentially valuable.

That is, models trained on only healthy inputs that the model creator has permission to use.

Personally, I haven't invested much mental energy in it.

We have these models, imperfect as they are, and they aren't going anywhere.

We might as well figure out the ways to unlock as much prosocial power given that we have them.

Billions of users are held captive in a small set of one-size-fits-none products that are hard to leave and have no alternatives, so the competition to improve them evaporates.

Mid tech.

But what?

Even easier if a master gardener does the work for you so you don't have to be a gardening expert yourself!

Software that can change itself.

My friend Aniket imagines what it would be like in a world where software is fully dynamic.

The PM job is about making software that can be sold to users.

But in a world of infinite software, everyone can have software perfectly fit to them.

The idea that PMs will create one-size-fits-none software is downstream of software being expensive to produce!

PMs today are racing to use LLMs to do their normal process faster, to get an edge.

But that's kind of like the racoon washing the cotton candy.

Oops, all gone!

What would it have to do?

It would have to be something that no company could control.

That would have all of the little features you need,

That would allow you to collaborate with the people you want to without coordinating on which bit of software to use.

If you had an everything app that did everything for you and you could collaborate with everyone in the world you collaborate with, you'd never use the old big-box apps.

But consumers haven't gotten that benefit.

Someone should unlock the prosocial power of turing completeness for humanity.

Compounding isn't necessarily good, it's just powerful.

Compounding is amoral; morality comes from whether it's a thing that's good for society or bad.

Obvious, but not the end point.

The first mass-market use cases of the internet were chat and email.

Everyone "gets it" immediately.

But that's not all there was in AOL.

As time went on, the value of all of the information at your finger tips grew, and then those experiences could be interactive applications.

The secondary use case of "teleport anywhere, do anything" was harder to explain, but could diffuse out as people used it.

Because if you get more efficient, that labor is owned by someone else.

If there's only so much the company needs done, they need less of you.

But in your personal life, AI makes you the master of your own personal life.

The more you can achieve that is meaningful, the more you can achieve.

Then you can take the messy, amorphous reality, abstract it away, and have just a clean, easy-to-reason-about box.

We do it all over the place

Ben Thompson's analysis is excellent and widely read in the valley.

There's no surprise anymore, anything on his comprehensive radar is known to everyone.

So the things that surprise the industry will be things that Ben Thompson can't see.

If you have to sell a commodity, the play is to go for volume, since you can't go for margins.

Volume gets you economies of scale.

SCO was famously litigious and cynical.

They would have destroyed the progress on Linux.

I think the industry would be in a wildly different place.

We're in the world where Oracle bought MySQL and then ruined it.

When you get to 90%, you still have 0 return.

It's not until 100% the value unlocks.

Other problems are "incremental work unlocks incremental benefit."

Tightening vs innovation.

The first kind of problem is what is necessary for a technical breakthrough.

A nice characteristic of ecosystems: they have the "marginal investment gets marginal benefit" but also have compounding returns!

Before you hit that viability point, the idea will rapidly evaporate if you take your eye off for even a second.

It requires tons of convergent energy to will it into existence, to pull it from the amorphous space of ideas into reality.

But once you hit that point, it's rolling down hill.

This nests: adding a feature on a viable product is a fractal version of this. Immanence and transcendence.

Maintenance and innovation.

The trick is the thing that can be startable and become unstoppable.

That's where compounding loops come in.

A self-accelerating thing.

Static friction is an order of magnitude higher than rolling friction.

If you have a thing you want to do, just get started.

Figure out a way to give you the little burst of energy, the why-now, the easy bootstrap into it.

A gauntlet is an onboarding flow that is high friction.

Only the most motivated users make it through the gauntlet.

Some gauntlets are intentional (e.g. an early, rough open source project).

Some are unintentional.

If the gauntlet is too bruising, it possibly delivers zero users.

But you can tune down the severity of the gauntlet until you're left with a dribble, and then tune it up or down from there.

That's very hard to do!

A quality loop that is on the side can't ever get going.

Typically you have to do it with a different, more quotidian primary use case, and develop the quality loop as the bonus use case.

Over time as the quality improves (hopefully at a compounding rate) it might eclipse the original primary use case.

The Optimization Ratchet.

The benefit of the optimization is clear, direct, concrete, immediate.

The cost of the optimization is unclear, indirect, ambiguous, delayed.

This creates a clear asymmetry, an unstoppable gradient.

Each optimization step that is taken is extremely unlikely to ever be undone.

So things get more optimized, until they get overfit, hollowed out, and then become prone to catastrophic failure.

This is why society has gotten so over-optimized, to the point of being hollow.^[fi]

It's beautiful, and the closer you look, the more beautiful it becomes.

Each layer supports the layer before, and your appreciation only grows.

Resonant things are transcendent.

However, some people are proud of being addicted to working out.

The question is: are you proud of the action?

If you are, you're more likely to evangelize it.

Resonance is general phenomena.

Resonant Computing is the application of resonance in tech.

Resonant AI is the application of Resonant Computing to AI.

Resonant AI is the humanity defining question today.

Not just a thing they like, but a thing they feel nourished by, proud about, happy to evangelize to others.

It's not just technology, it's something much deeper.

That means it wants to accumulate as much of a user's data as it can

Also use that data in at least some ways that the user finds valuable.

That last part is aligned with the user's incentives, at least.

Don't drive something that matters off a cliff, or let them drive themselves off a cliff.

"I'm just giving them what the number say they want."

If your friend were drunk and said they wanted to go on a joyride, would you let them?

Everything is safe because it all aligns with your expectations.

The closer you look, the more comfortable with it you become.

Resonant privacy.

Gives you peace of mind.

If you have a principal agent problem: people who are not on board for the long term will choose the minor short term benefit at catastrophic long term cost.

Imagine a world where you were locked to a specific collective, for live, with no possibility for exit.

We only care about things on the time horizon we expect to be involved.

Renter mindset vs owner mindset.

The reason no country allows tourists to vote is because if you had only short-term users voting, the country would be destroyed.

So why doesn't that happen to public companies, which have a large number of "tourist" shareholders?

The reason everything isn't destroyed immediately in practice is because there's a mix of short- and long-term interest.

Imagine that every single shareholder was expecting to hold the stock for precisely three months and then sell it and never hold it again.

"This is a microcosm of why AI is waning in popularity with normies:

> "we're going to take out your jobs"

> offer no tangible solution as to what comes next / how normies ought to get by

> but "trust us bro, everything will be better with AI"

Out-of-touch hubris, unfortunately"

Dario is Edison.

Sam is JP Morgan.

Someone is going to be Henry Ford^[gj]

That's because they remove something invisible.

All of the onlookers won't see they did anything at all.

But all the competitors just die, what a crazy random happenstance!

Icky!

Innovation requires cults.

Maintenance requires spreadsheets.

That's mainly busy waiting, ready to go as soon as they're ready.

Enormously wasteful!

Retaining optionally creates hollowness^[fm][fn]

Resonance comes from boldly committing.^[fo]

Emergence is like magic.

Impossible to see directly.

Only possible to see when you blur your vision a bit.

Emergence is magic.

You can never pin it down, but it's real, powerful, inescapable.

You can only react^[fp].

Everything just pulls it in and there's nothing you can do.

Growth comes from challenge.

Challenge doesn't feel good in the moment.^[fq]

But afterwards you're glad you did it.

Bad challenge grinds you down.

Good challenge makes you stronger.

In the moment it feels like all challenge is bad, and after you're done most challenge feels like good challenge.

Doomscrolling is not comfortable, but it's also not challenging.

It doesn't force you to grow, change, update your model of the world.

It just says "Yes, you're right, the things you thought were bad are bad."

Challenge is not comfortable.

But not all discomfort is challenge.

Math typically chooses beauty over pragmatism.

In CS, there's no need to choose between pragmatism and beauty, you can have both.

I've heard this insight attributed to Alan Kay.

If any kid makes a breakthrough they can all copy it.

Also there are older kids to learn from and pull everyone up.

Emergence can't be seen in the details, only in the whole.

But you can't force yourself to be obsessed.

Imagine a zombie exec at a large tech company.

Post financial but have no other meaning.

Work fills in for meaning.

"What would I even do without this job".

The job tells them a thing to keep optimizing!

Then other people continually vote that it's useful to keep, preventing it from eroding away.

But it gets smoother through erosion and selective rebuilding.

The process of keeping is orders of magnitude cheaper than the process of creating.

This is the process by which everything of value emerges.

They have to change their own.

If they don't realize there's a hole in their understanding, they aren't yet ready to change their mind.