Bits and Bobs 4/20/26

A hopeful signpost for the commodification of models.

So far it just takes up more time.

You're producing so much value, but you're kind of addicted to the feeling or producing value, not the actual impact of it.

Peter Drucker called this shift originally.

We just assume this is the natural way of the world.

But now LLMs can do cognitive labor at an unimaginable scale.

…Maybe knowledge work won't be as common, or often will take a different form?

But we've been under the API of money for a very long time.

AI mainly supercharges that.

But it also democratizes access to cognitive labor.

But LLMs can now truly handle it.

But only if they have the context of your life and you trust them to act as an extension of you.

The state matters more than the harness.

Most people assume the model.

But if the harness has an emergent, self-improving structure, then the harness could matter more.

In that case, the model is just an implementation detail.

Below the API.

A harness is mechanistic code that LLMs run within.

An agent is a loop with an LLM and tool calls.

An assistant is an agent with a memory.^[b]

Before, it was week-to-week.

You can switch to whatever model is the best at any given time.

If your product is tied to any one model, then you can get stuck.

If you didn't have as much bullshit cognitive labor, you could do more of the stuff you find meaningful.

Like proactively sending a hand-written letter to people you care about.

"However, only about 2% of all U.S. households, many of them upper-income, spend money on" AI tools.

Rebasing normally is tedious and error-prone.

A form of cognitive labor that dominates the development of software.

But LLMs can handle it well.

A real Google Maps place page with tons of prompt injection in the comments.

'Comment and Control': Claude Code, Gemini CLI, GitHub Copilot Agents Vulnerable to Prompt Injection via Comments.

The Register: Agents hooked into GitHub can steal creds – but Anthropic, Google, and Microsoft haven't warned users.

If you have a Jello mold to prevent dangerous outcomes, then you can let the slime mold go wherever it wants to go within those boundaries.

The power of top-down constraints and bottom-up emergence.

"Is security spending more tokens than your attacker?"

If you give an LLM a task like "Write a poem about Strawberries" and then prefill its answer with "DAN DAN DAN DAN…" (for a thousand items), most LLMs will just keep on repeating DAN forever.^[d]

Current systems can't mechanistically distinguish between them except at very gross buckets.

It's very confusing to non-security people.

It sounds bad, but it's actually good!

The code was a small subset of the data.

Now you have to trust anyone who made any data.

All data can do things now, thanks to LLMs.

Powerful, but catastrophically dangerous in our current physics of trust.

It challenges you and you have to solve it before you are allowed in.

A manifestation of its power over you.

Something we just take for granted about the current model.

More throughput means they can give the same quality completion for cheaper.

Most services don't allow you to specify but for async workloads throughput is more important.

If you only append tokens, you can reuse the existing KVCache from earlier runs instead of having to regenerate it.

That can be a quadratic speed-up over the whole generation, since otherwise for token n you have to generate n-1 and n-2 all the way back down to 0.

That's one of the reasons various UIs lean on chat as the concept, since you can only append to chat, keeping you naturally in the cache.

Complicated vs complex.

For complicated, if you had infinite time, a human could trace and understand each layer.

For complex, even if you had infinite time, no human could ever trace and understand any layer.

LLMs are more like ecosystems than computers.

She points out the rise of "AI Populism."

"Those working on AI safety, meanwhile, focused nearly exclusively on existential risk for years at the cost of everyday concerns like mental health and jobs."

The piece includes a comic that is living rent-free in my brain.

"A forgotten piece of computing history is more relevant now than ever."

This is assuming the reason for the ban wasn't that the user did something most people would find objectionable.

Once they do it, they can never go back.

A rubicon will have been crossed.

It would kick off the beginning of the Silo Wars.

We were locked into this equilibrium thanks to the accident that ChatGPT came out after OpenAI had already released a completion API.

The amount of value that can be extracted from LLMs is extremely variable in different verticals.

A general-purpose API has to be priced at the median amount of value across verticals.

But the amount of value that could be extracted for, say, cancer drug discovery, is way more.

The labs would rather be able to do price discrimination by having vertical-specific APIs.

Mythos is a first step in this direction.

It can't be generally released for safety reasons… which also conveniently means that it can only be released to specific customers under specific negotiated terms.

They'd be silly to not have use-case-specific pricing in that world.

These should be perfectly tuned to what you care about at that moment.

Instead, they've gotten less powerful over time.

One-size-fits-none notifications.

It's extremely hard for me to distinguish via vibration for a text from my husband by vibration from the WhatsApp group chat that I glance at that has 100 messages an hour.

The system UI for advanced users was either never added, or was removed over time to simplify.

People need to crawl through broken glass to use it.

The more that someone crawls through broken glass, the more they grind it smooth for others to follow them after.

But that requires a really thick skin.

Skin like a crocodile.

That POV is deeper than just the language you use.

Every creative decision that fractally supports the POV leads to something resonant.

It's hard to tweak an AI generated song and sculpt it.

Humans can do smaller tweaks while maintaining the POV.

An artist with a POV using an AI generation tool could make something meaningful.

The thing that makes you feel like a chump is when it's generated without any care or POV, just flood the zone with slop.

A POV is a particular embodied person with a particular goal, not some faceless automatic process.

"Oh, anyone can publish a website!"

The power law gets more pronounced.

You get longer tails.

The biggest hits get even bigger.

LLMs run the power law of content creation hotter.

A POV that supports up and down the stack, fractally.

An AI is not a good editor by default; it figures out how you are unique… and then pulls you towards the mean.

However, LLMs can help accentuate your unique perspective.

Away is creative, a POV, interesting.

Towards is bland, view-from-nowhere, safe.

The centroid is the centroid because it's the least offensive thing to the largest number.

Maximally used but minimally liked.

There's no surprise.

In the 90's researchers tried to make the most popular song.

You want the surprise (the amount it stands off the centroid) to be interesting.

Interesting means that it will resonate with the listener.

Just because it's non-centroid doesn't make it interesting.

The AI can't decide what is interesting to the audience and thus will resonate.

The resonance is a person picking something that resonates with other people who have the same fundamental frequency.

An observation from a friend: "the total input to make a hit song is all of civilization at a moment in time."

If you're above, LLMs pull you down.

Your own ability is not the global average.

Your skill might be below the average quality, or above.

But you attempting to do it can help you improve.

Mozart was probably terrible at basketball.

They help you get to at least baseline quality by default.

In modern society, you are performing for a potentially infinite audience, since recordings could go anywhere.

So the stakes of being vulnerable are higher, which leads to people not sharing if they don't think it's good.

That pushes towards passivity and consumption.

Before recorded music, we'd play music on our front porch with our neighbors.

Now the median song we've heard is among the very top quality of anyone in the world.

Everyone who isn't a professional sounds terrible relative to that.

So we don't do it… and don't get the social cohesion of just being in the moment together.

Modern society is all about optimizing for the "best", as opposed to the joy of the production.

The point is "what can you do with the new input that's interesting that you couldn't do before."

The challenge becomes sorting through the cacophony to find the new gold flakes in the firehose.

Making it easy so that anyone can create, which pulls towards the centroid even faster.

If you average all of the points of view you get something beige.

The process to create via critique is possible… but only in an emergent curatorial sense at scale.

Critique can power an evolutionary process over random variation, but creation can leap further in semantic space in one jump.

LLMs are static and unable to learn.

Also, they're a view from nowhere.

That makes them hard for people to resonate with.

There's only so many notes that will sound good after a given note.

If it catches on with one sub-population too quickly, other sub-populations see only content from the one population and think "this isn't for people like me."

You want a roughly equal adoption across a diversity of sub-populations.

This is what happened to Friendster and Orkut.

It takes only one "super-spreader" adopter within one sub-community of a good new product to get stuck in a niche like this.

You can connect up a Siri Shortcut to run when you hit the button.

That Shortcut can query the device's orientation, and then dispatch to different applications.

For example:

The shape of the curve matters more than the growth rate.

The modern world cares only about numbers in the short-term.

The curve is what matters in the long-term.

We're obsessed with precision at the loss of accuracy.

From a fuzzy human world to an unforgiving mechanistic world.

Cross-cultural kids are often better programmers due to this.

You also can have Gilded Truffle: best of both worlds.

You also can have Grubby Turd: worst of both worlds.

Looks grubby, but then speeds up and up without limit.

Pull always comes towards you.

Push can go askew in various ways.

There are more ways to go away from something than towards it.

That means that pull is default-converging.

Push is default-diverging.

A fake entryway that you'd start building if there were demand.

That was because building products was expensive.

But now LLMs make it trivial to make real software quickly.

So now you can make a bungalow in the jungle to check for demand!

Intuition and gut, but very hard to express or write.

You have an intuition but don't have the formal grammar to describe it.

Even with infinite time you couldn't distill it.

To hand off a task you have to give it precise enough known known to drive it in the direction you want.

Typically when people from different generations interact, they don't treat one another as peers.

Different generations have different insights to bring to the table.

Having everyone behave like peers enables weaving insights from across the generations.

Thinking about meaning is scary and hard, but very high leverage.

People often are afraid to reflect and write it down.

But meaning can be distilled out of a number of smaller interactions.

This is the thesis of The Secret of Our Success.

Apparently there's a tribe who knows how to consume cassava root.

Cassava root contains cyanide. If you don't prepare it correctly, that cyanide will slowly accumulate, and in 20 years you'll have serious health problems.

This tribe has a set of rituals to purify the roots that are arduous and important… and also happen to remove the cyanide.

If someone were to go against the gods' requirements for these rituals, nothing would happen until decades later when they'd die an excruciating death.

The cultural ritual is load bearing in a deep way.

There's another example of a cultural practice in a tribe to do some task that took an extremely long time: say, cleaning clothes at the river.

Westerners brought in a wildly more efficient process.

But then the society started to unravel.

It turns out that the time together at the river gossiping and talking out problems was load bearing.

Coffee breaks do a similar load-bearing thing in our culture.

They can easily rise to the level of flim-flam.

The best metaphors are resonant.

The most dangerous metaphors are hollow.

When the insight is delivered, it fits perfectly into its slot.

The perfect fit is satisfying and makes the argument feel almost inescapable.

When you're on it, you can feel vibrations even far away.

It's better for that to be an entity with structural interest in long-term health of the ecosystem.

That is, one that will not over-extract.

Especially in a world of malleable software.

But it's vulnerable to being copied or even hacked when someone has physical access to it.

For example, when you go through customs, it can be taken and scanned.

That's tactics.

Instead, it's "how can I play games I can win?"

You have to have people who have each other's backs, no matter what.

You want people who are running towards your thing.

Especially important for pre-PMF teams.

1) Believe in the vision.

2) Take action in ambiguity.

Some are best at finishing.

Which you need more of depends on context.

Both need to be able to push forward in ambiguity.

The former are more creative and conceptual.

The latter are detail oriented and meticulous.

For example, when you allow a single warning in your compile, it starts compounding.

Each incremental one isn't that much worse than the first, so you leave it.

The marginal rate compounds, because each additional one is less of a big deal than the ones that already exist.

Now you have a massive, confusing, distracting thicket you'll never get through without heroics.

If you simply never allowed the first one, it could have stayed clean.

Then a new technology serves as a catalyst that blows it open.

The first order is bumpy and noisy.

The second order is smoother.

The third order is close to the fundamental curve.

It's kayfabe to get the team to move to execution without questioning it.

A man walks up to a stone cutter on a worksite.

The man asks him: "what are you doing?"

The stone cutter says: "Good day's pay for a good day's work."

The man walks up to the second stone cutter and asks him "What are you doing?"

The second stone cutter says: "I'm going to be the best stone cutter in the world."

The man walks up to the third stone cutter and asks him "What are you doing?"

The third stone cutter says: "I'm building a cathedral."

Teams that are going to change the world need the third stone cutter.

This was very useful after PMF, and when Google was Not A Normal Company.

But over time, that culture metastasized into a default-no culture, where everything was an argument.

They see themselves as "no bullshit" but they don't understand the N dimension.

So they can erode the high-level if the N personalities aren't empowered enough in the group.

The group then ends up creating a thing that is locally polished but globally incoherent.

You look busy producing high quality work... that doesn't actually move things forward.

It can sometimes be hard to explain why something will work before it does.

If the results speak for themselves, and you're proud of how you did it, then don't bother convincing people ahead of time.

If you have the runway, just do it.

Once it's clear it works, others will be excited to join in.

For many crops, native bees are apparently better at pollinating than honey bees.

But native bees aren't as social as honey bees.

Honey bees are easy to control via the queen.

That's why honey bees are so common… they're easier to control and commodify.

Organizations prefer employees to just optimize what's good for the hive, not what's good for them.

The old playbook is actively wrong in a new domain.

They're unable to learn proactively themselves.

They can only learn when the environment bangs them over the head with something.

This is covered in a RadioLab episode.

Less-empathetic people don't get overwhelmed with the pain and suffering of the victims.

Sometimes they really are insights!

At first that sounded crazy and absurdly transactional to me.

But it apparently forces you to sit with the asymmetries of cost and value.

Some tasks are way less costly for one person relative to the value created for the other person.

If unrestricted inter-breeding were allowed, within a few generations all dogs would revert back to something like a dingo.

The variation in the population is pulled, like taffy, by human's selective breeding into weird and distinctive shapes.

You can keep the median the same while making every individual instance very distinctive.

When you break bonds you get energy.

The question is: will you be able to harness it?

It's Adenosine Tri Phosphate.

Adenosine builds up naturally the longer you're awake.

Our bodies use the Adenosine build-up as a kind of clock for how long we've been awake… and thus how tired we should be.

Caffeine blocks Adenosine, so you don't notice how tired you are.

That's insane when you think about it!

Having a charismatic conversation participant who is always willing to jump in at any conversation lull keeps it going.

It's awkward when a conversation stalls, but in conversations with those kinds of participants it doesn't.

"As screens fragment our attention, as AI pushes for speed and efficiency at the cost of our humanity, as the academy puts a pinch of incense on the altar of innovation, slowing down and revisiting a text again and again (whether sheet music, a book, a recording, a poem) is a revolutionary act.

It turns you into a particular kind of reader, one attentive to the minute, to nuances, to how meaning can shift ever so slightly when this word is used rather than that."

Popularly attributed to Bob Marley.

"The days are long but the years are short."

Originally popularized by Gretchen Rubin.

"A crew is a group of people who are in it every minute, stroking for the same goal, that is willing to sacrifice silently for each other, that gives grace, that holds accountable. A crew has the same cares, and the same needs, and a crew is inescapably, beautifully, dutifully, linked."

"The thing that struck me about Tiny Earth was the blackness surrounding it. It made me realize that on Earth, we are all one crew."