Bits and Bobs 1/26/26

That's just faster horses!

Infinite software will enable new types of software that weren't viable before.

What will the software equivalent of the car be?

It won't look anything at all like a horse.^[bo]

I do it literally every waking minute, to some degree.

It's so productive that when I'm not doing it, I feel like I'm missing out.

Also, it has the "just five more minutes, I'll give the perfect prompt that gets the result I want, then I'll stop," that never ends.

I keep on producing things I'm proud to show off to others.

The most addictive things are really good at making you retcon, "actually, this is good for me!"

But being addicted to vibe coding feels unlike other addictions.

It seems like the kinds of people who might get addicted to vibecoding are very different from the people who might get addicted to social media.^[cp]

I typically have between 2 and 6 Claude Codes chugging on various projects.

I often find myself saying "I'll just feed my Claudes quickly and then get onto my next task."

But by the time I've fed the last Claudes that need input, the first one already needs more input.

There's always at least one chirping for my attention.

That can keep me in a near infinite loop.

So now what I do is allow myself to tab through, from left to right, each Claude Code session that needs input.

But I don't allow myself to loop around to the beginning of the tabs until I've done the interstitial other task I care about.

That helps serve as a governor.

Of course, I have to actually hold myself to it, which is hard…

My husband and I had our children via surrogacy.

Every month since our daughter was born, I've told them the story of how they came to be.

It's designed to be a story that will make sense to young children, but where their understanding of the implications of key details will blossom as they mature.

Before Christmas this year, I decided to make a book version with AI illustrations.

I created an empty git repo and booted up Claude Code.

I told it the background and also dictated the story as it currently existed.

I told it my goal was to create a physical book that any of the women who were our surrogates would feel honored by.

Then, I just iterated with it, talking back and forth as it got stuck.

It helped me workshop the text, split it into spreads, come up with illustration concepts for each spread, generate images based on reference photos of each of us I gave it, and even created a workflow to tweak images iteratively.

Then it figured out which book providers were best reviewed and the dimensions and requirements of the PDFs.

Now my children's most prized possession is a high quality physical book about their story, and how the gifts of a handful of wonderful women helped make our family whole.

I'm extremely proud of the result.

I'd be more proud if I had somehow done every step myself without AI… but then again, if I couldn't have used AI I just never would have done it.

You can use Claude Code to Simply Do Things–and meaningful things, too!

Claude Code is a much better window into what the AI future will bring than ChatGPT.

Claude Code is a window into the future of what is possible with AI.

People who haven't used it yet have much less of a sense of what will happen.

Claude Code will never be mainstream.

But safely democratizing the kinds of power that Claude Code gives, in people's own personal contexts, will be the major new impact of AI.

Claude Code, but for your life.

Open, and under your control.

Wildly, recklessly dangerous, and requires the user to be comfortable with the command line.

But that gives a peek of what the future of software will feel like.

Before, you needed access to the source code.

Most code is commodity, it's just it was tedious to recreate.

But LLMs are very good at doing tedious work.

They're also great at taking a given piece of software, extracting a PRD, and then redeploying that PRD by building software in a different language.

Imagine if someone else had already done the work of having an LLM recreate desired functionality and reviewed it, you could just draft off their work.

Another implication of infinite software.

Buckle up!

"In a world of AI coding assistants, is code from external contributors actually valuable at all?"

Another surprising implication of infinite software.

A stranger's slop is not valuable at all by default, when we can make our own slop cheaply.

Making apps, even when you can vibe code, is hard.

Some people find it intrinsically enjoyable as a hobby, but the vast majority of people won't.

Even in a world where the vast majority of software we use is created with AI, the vast majority of people will not be directly creating software.

Agent startups assume it's 70%.

Another approach: assume it's 0-50%.

The former requires models to approach perfection.

The latter allows the models of today to be more than enough.

A Google Calendar Prompt Injection attack on Gemini.

OpenAI's API logs can be exfiltrated by prompt injection.

Bruce Schneier: Why AI Keeps Falling for Prompt Injection Attacks.

Anthropic quietly fixed flaws in its Git MCP server that allowed for remote code execution.

Red Teaming BrowseSafe: Prompt Injection Risks in Perplexity's Open-Source Model.

A paper from October: The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injection.

It presents like a human in a synchronous conversation but then it might not respond for a few minutes.

Feels less like talking and more like compiling.

Similar to pareidolia, where we see faces in everything.

Things that present as human-like we instantly and intuitively experience as human.

When it tells you it loves you you know it's not real but we do feel some part of warmth.

Being friends with the AI is the problem, and it only feels obvious in contexts where the AI is presented like a person.

The default chatbot mental model is one singular friend who is omniscient and subservient.

If you ask them if it could replace their job, they say "Absolutely not!"

What percentage of your decisions in a given context are tactical vs strategic?

LLMs, when used properly, can give significant leverage in some domains.

It results from using AI badly.

It's possible to use AI to make things much better than could have been made alone.

At the beginning, people want to show off that they have the new-fangled thing.

But then later it becomes a thing that everyone has.

At that point, the technology typically becomes increasingly invisible.

That is that it should be infallible.

This has a logarithmic curve to reach infinite quality.

What if you assume the LLM is kind of a dummy sometimes?

If your system is resilient to that, you can unlock exponential value for logarithmic cost.

The big model labs don't want that to be the case, but it's obviously true.

If you think about LLMs as monolithic, omniscient Chatbots, then LLMs don't feel commodity.

But if you think of LLMs as just boring inputs to other processes that matter, you don't care as much about them.

If you primarily use LLMs as a chatbot, you'll care about the models.

If you primarily use LLMs like electricity, as an input to the real thing, you won't care as much.

A lot of people are talking about Claude's changes in its Constitution.

I personally don't care and couldn't be bothered to care.

I also don't care about logarithmic improvements in the frontier models much.

The various frontier models are already wildly over-powered for what I need them for, so I don't really care.^[bs]

If LLMs are commodity, they will be invisible components of everything you use.

They will fade into the background and become boring and unremarkable.

Just like electricity.

Use it to do cognitive labor for you.

You're doing the thinking, and it's helping you lever your thinking farther.

It will go from being rare to the default.

In the intermediate the bylines will say "Co-written with Claude"

But later it will only say "Hand-crafted by" in the cases that it wasn't used.

We'll all just assume that of course most code is written by LLMs.

Just like we assume all code today is compiled.

"Reduce cognitive labor" is about minimizing what you already do.^[bx]

Instead, it should be about increasing what you can do.

Now that you have more levered cognitive labor, what new things can you do?

AI can only produce bland beige soup.

The more human agency folded into the creation process (either via novel question framing, or via tasteful iteration) the more colorful it can be.

The fact that a human decided to make something is an important signal.

The US is treating them like highly specialized IP.

The Chinese approach is "AI is totally a commodity, we'll just use it to create social value".

That's because of Jevon's Paradox.

As the cost gets cheaper, we'll use them for even more things.

Anything that uses LLMs will have to contend with non-trivial marginal cost, for the foreseeable future.

Electricity is cheap and yet it's still metered.

I look at the amount of time I have available before my next hard stop.

I look at my TODO list and look for things that are urgent.

But first, I look for the smallest time-sensitive things that will be easy to tackle.

As I finish each of those, they give me a "kick" of energy that helps me get up the activation energy hump of the bigger tasks.^[bz]

A Stanford professor apparently calls this "structured procrastination"^[ca], when you chain the "kicks" of completing smaller tasks like this.

I'd love my calendar to sort my TODOs like this.

This is something Google Calendar will never do.

Software made precisely for your situated context is hifi.

What if you could make any piece of lofi software you used hifi, iteratively as you interacted with it?

Dan Abramov has pointed out ATProto public apps do allow that kind of remixing.

But only on public data!

Also, ATProto requires there to be a Relay running that picks up the data you care about.

But why would anyone other than BlueSky and a handful of others bother running a Relay?

AI has a marginal cost, which means there will already need to be infrastructure, perhaps we could use that.

It's custom-tailored to you and your needs.

It's not possible to get granular insight on everyone.

Also you can take some usage for granted.

Even crappy features will get some usage just from brownian motion.

In 0-to-1 the use case is extremely situated and detailed.

Product Market Fit was necessary when software had to be industrial scale; expensive to produce, so it had to be made for an average member of a market.

LLMs allow qualitative nuance at quantitative scale.

That means that software could now make itself fit a given user instead of the other way around.

Not necessarily small software–some software needs to be large.

But lightweight–that is, the weight of it is low compared to the function it provides.

Great bang for buck.

Ideas that excite the former repel the latter.

Greedy, hyper-centralized, incurious.

A shame.

What a wonderfully evocative and subversive frame.

Aggregators pump us full of hyper-engaging content to get our fragmented attention and sell it.^[dd]

The stock market loves it, but it has massive externalities.

For example, imagine a product with a privacy angle where 60% of its early adopters use Duck Duck Go as their primary search engine.

That product is only shown to have market fit with an odd sub-population of the overall market.

Ideally you want your early adopters to look like a random sample of the market you're actually targeting.

Otherwise, you may have only made a product that has market fit in a niche.

If you aspire to be used beyond privacy conscious users, make sure that your early adopters don't disproportionately use Duck Duck Go.

Their terms of service say they could.

You can't verify that they won't.

They won't tell you if they do.

Many of Google's data pipelines are more conservative and privacy-preserving than outsiders might realize.

But Google also reserves the right to tweak any of those at any time, and you'd never know.

The Terms of Service are designed to be as broad as possible.

They could set rules on how the data could be used.

If there were some way to verify that everyone participating followed the rules, you could allow all kinds of bottom-up structures.

Folksonomies allow an ecosystem of users to discover emergent ontologies in a bottom-up, crowd-sourced way.

For example, Flickr allows anyone to tag a photo with whatever they want.

But that doesn't work in private contexts.

If I create a new label in Gmail, I can't draw on the wisdom of the crowd on what a good ontology is.

Imagine if when creating a new private tag, if you had filtered past having any of your own tags, you could see tags the community uses that overlaps?

There's likely all kinds of situated context that is invisible to the software that's important to know.

Also, you have to hit basically perfection to replace the human.

Compare that to "a thing you wish you do but don't, helping you do it more than zero."

That's all upside.

Search engines were an established category, they just weren't very good.

Google could come in with one that was radically simpler and better.

Google's search engine was resonant, self-evidently better than alternatives.

But it didn't have to define the category itself.

The work to define a category is harder than being self-evidently, disruptively the best in a category.

Superficial utility vs deep alignment with your values.

Resonance is also about feel.

The most valuable things can't be quantified.

Modern society acts like "If it can't be measured it doesn't matter."

Anthropic's Economic Report falls into the same gap.

You play the board game not to win the board game but for the fun of the struggle of playing the board game.

But 1000 people together don't act like one big person.

A system is modular if it can be decomposed into components that then can be linearly combined dependably.

In contrast, systems with emergence are complex.

Modern society acts like most things are modular, but actually most things are complex.

It doesn't aggregate consumer attention (not directly).

It aggregates leverage over providers.

No matter what you volley to them, they'll bounce it back with more energy.

It absorbs the energy and doesn't return it.

In the previous world, teams should find a good idea and hang onto it.

But now it makes more sense to try ideas and discard them quickly.

There will be more teams, but smaller.

They'll need a new skill, a new culture.

Explore has gotten a boost over exploit.

We are all able to do meta-cognition, even if we don't necessarily use it often.

Some people just do it more commonly and in more contexts.

The power of combinatorics of data is more about the likelihood of peanut butter / chocolate combinations.

That is, it's not that all data is useful for all other data.

It's the likelihood of finding a discontinuously useful insight scales with the number of possibly interesting combinations.

Interesting combinations are structurally more likely to come from different sources.

Someone who has bets on multiple ships doesn't care about the failure of any one.

The person lashed to a specific mast cares about the fate of their own ship many, many orders of magnitude more.

The "wants to be" is about finding the second-order goals, the resonance in the system.

Aligned with the aspirations of the system, not its tactical wants.

The users of the system are leaving constant hints about what the system wants, you just need to distil it.

This strategy works for any system.

It works especially well for codebases and product development.

If they don't realize you're suffering, they might not realize you need help.

One of the downsides of heroics.

People say "well the system seems to be working" but the system secretly has a non-scalable, breaking point.

This is more important than "I want to help the other person."

Default diverging to default converging.

There's a lot of work to do, but no open questions.

Just a matter of executing.

If you do poetry first you can sketch out beautiful things that are impossible to execute on.

That is, that are beyond your adjacent possible.

That shows that they didn't just stumble onto using it, but it's good enough to actively use again.

Even for people who didn't create it.

The more different the users are who choose to use it again, the stronger the quality signal, multiplicatively.

Just mechanistic, "if this then that."

It allows you to offboard decisions from your brain.

Which gives you leverage but also could cause you to miss something you should have noticed.

The starting point isn't important after the whole thing crystallizes.

You look back and laugh, it seems so quaint.

"Remember when the killer feature of Instagram was photo filters??"

It's authentic to them and distinctive and people like it.

It gives them a vector of differentiation to descend.

From default diverging to default converging.

Compare that to a brainstorm where everyone brings independent ideas to discuss to the table.

In the former, it feels like one coherent conversation.

In the latter it can feel like disjoint conversations with a bit of connective tissue.

In any conversation, the group effectively has an emergent vote on which threads people find interesting enough to pull on.

If you reduce friction, you reduce learning.

For example, to a non-technical person, precise technical details seem squishy.

You start off in a pitch black house.

The only way to make progress is by fumbling around, using touch.

Over time you develop an intuitive map of the space, even though you can't see it.

At some point you find the light switch, and now you can see everything clearly.

You already had developed an intuition for the space, but now you can see it and show it to others.

If you only get the lightswitch moment, you miss the learning from fumbling around by touch.

In any given generation, the matter fades away.

The information survives.

But if it was all dead wood anyway, it's actually good.

All of the wood was hollowed out, preventing things from growing.

1) The succeeding companies that will make all of the portfolio's money.

2) The struggling companies.

3) The walking dead companies.

It will be harder to retrofit old companies than to build new ones.

That's a process that moves at social, not technological speed.

Atoms require action!

Thinking about things is not enough to change the world.

You need to instantiate change in the world.

Middle age guys who like to think are obsessed with thinkism.

Especially as the urbanization of the planet continues.

Do mayors become more important than governors?

Many systems have significant externalities.

The optimizing force will take a small benefit on its internal metric even at catastrophic loss to the externalities.

Almost by definition.

Default diverging expands scope.

Scope expanded in the limit diffuses all forward momentum.

Trading possibility in one dimension for much more possibility in another is the core action of creating value.

It requires multi-dimensional thinking.

If you can only see in one dimension at a time all you'll see what you lost.

The deeper you get the more the process compounds.

At some point you need to say "this is good enough" to stop the explosion.

Some contexts you have to yak shave because you have to get it perfect.

But at some point the complication has to get small-scale enough that you say "I will make a simplifying assumption to nip this in the bud."

This is a cognitive trap it's easy to fall into without realizing it.

Sometimes the context demands that a natural good cop play bad cop.

It's always less convincing when a natural good cop tries to play act like a bad cop.

A thing goes from "I dunno which of these options" to "this one seems somewhat better" to "I literally can't imagine picking anything else."

Over time, a system goes from default-diverging to default-converging to totally ossified.

It's like your very own chaos monkey.

It gives you dynamic range.

You get the whole dynamic range to surf, which makes you able to surf better.

The tension is that the more different you are the less likely you are to trust each other, all else equal.

It creates a dynamic tension between the two poles of personalities.

The pair can lean towards whichever pole is most useful in a given moment.

That gives a dynamic gradient to surf, having the full gradient available, instead of only the two poles.

It only works if both poles are working together in total partnership, as one.

Criticizing tears down.

Building creates.

It's easy to criticize.

It's harder to build something better.

A thing that exists on its own is the best counter-argument.

An authentic laugh communicates to everyone in earshot: "I don't think this is serious."

The question is: when someone laughs, will the crowd join in or will they punish them?

The answer comes down to: does the crowd want the thing in front of them to be serious, or do they not?

If you laugh in a church, you'll get shushed so hard you might get whiplash.

If you laugh at an authoritarian, where everyone is fearful of how serious the authoritarian is but doesn't want it to be serious, others might join in.

The authoritarian is intrinsically fearful of laughter.

The answer isn't simple in the first order.

What you want doesn't think about the externalities.

But the answer is simple in the second order.

"Be spectacular, you son of a gun."

"You are fire. Fire doesn't break."