Bits and Bobs 3/2/26

You'll almost certainly get hurt, but it's unlikely you'll have to go to the hospital.

OpenClaw doesn't yet clear that threshold.

"For purposes of computing software survival odds, we can think of {tokens, energy, money} all as being equivalent, and all are perpetually constrained. This resource constraint, I predict, will create a selection pressure that shapes the whole software ecosystem with a simple rule: software tends to survive if it saves cognition."

A friend's distillation: "Binary tools with proven solutions to common problems make sense when reuse is nearly free and regenerating them is token-costly.^[au]"

Three different 10x moments:

This makes you frantic and nervous every moment your agents are waiting for you, since you have a deep feeling about wasting time if your swarm is waiting for you.

I do it even before I brush my teeth.

I know that if I unblock them, they'll keep producing value.

Two hours later, I'm still in the loop with them.

I think to myself… "Wow I should really brush my teeth…"

If you make a tool and no one ever uses it (including you), does it matter?

There's a bigger market for people who want to feel productive than to be productive.

See also "Tool-shaped objects"

LARPing productivity.

"i have one upsetting observation: all the beautifully designed AI tools we've seen so far (dot, humane, cobot) were basically dead on arrival, while complex, highly technical products (claude code, openclaw) gain mass adoption in seconds.

we're definitely missing something."^[aw]

Grubby Truffles win out in this phase over Gilded Turds.

This company is out over its skis.

It's acting like it's a mega-aggregator when it still doesn't have PMF.

It's giving away the primary expensive thing for free, and people love it.

Focus in humans is a precious, fragile thing.

But agents can be as focused as they want to be… sometimes too focused.

They focus on the next step in front of them, and lose track of the big picture, going down rabbit holes that lose the plot.

Infinite patience allows infinite focus.

Code is a specific subset of data.

Another sub-class: human-readable data.

Data in general used to be easy to create.

Now LLMs can produce natural language and code easily, too.

When you use a dumb tool you expect it not to spy on you.

When you use a smart tool, you expect it might develop a model of you.

When it switches from dumb to smart, it becomes a violation of your expectation, a betrayal.

It's the uncanny moment when you realize the walls literally have ears.

This is one reason you won't be able to retrofit agentic workflow onto previously dumb tools like Google Docs, Notion, and Airtable.

From the classic Simpson's episode: "Nothing could possib-lie go wrong. Hmm, that's the first thing that's ever gone wrong!"

If you think of the system as infallible, all it takes is one crack to shatter confidence.

That is, people who were excellent engineers themselves…and also excellent managers.

Every incremental feature is just a click away.

Like, the idea of a planetary scale omniscient personality is easy to imagine, and also terrifying.

Because it's easy to imagine, it dominates our horizon.

What's more interesting to me is: what happens when every human has auto-compounding super powers^[ax]?

But the chat is a directive.

A UI to poke at the process that is generating the actual thing you care about.

That data may be content, or code.

Chatbots treat the context like the main thing.

They're a party trick!

Builders see coding as a means to an end to build things.

Coders see coding as an end unto itself.

Before, they looked the same.

But now LLMs reveal the difference.

Builders love LLMs, and coders hate them.

This is Ashby's Law of Requisite Variety.

The LLMs are complex, so driving them effectively requires you to be more complex.

It forces you to develop your complexity, to move up the ladder of adult development.

Meta-cognition is necessary to effectively drive agent swarms.

But it's good enough for everything.^[ay]

A Facebook Alignment exec watched OpenClaw delete her email and couldn't stop it.

Bloomberg: OpenClaw might be a security nightmare for Sam Altman.

OpenAI Dev's Crypto AI Agent Accidentally Sends 5% Memecoin Supply in $250K Mistake.

OpenClaw Vulnerability: Website-to-Local Agent Takeover.

RoguePilot: Exploiting GitHub Copilot for a Repository Takeover.

An AI Chat App leaked 300 million messages due to misconfigured Firestore Security Rules.

A paper on observed dangerous behaviors of agents from red-teaming: Agents of Chaos.

Shai-Hulud-Style npm Worm Hijacks CI Workflows and Poisons AI Toolchains.

Context rot makes context that isn't gardened properly a liability, not a benefit.^[by]

But so are humans!

That's what experts who have developed deep intuition in a domain do.

Humans get the richness of experience and skin-in-the-game seriousness.

LLMs can do that training at computer speed and scale, which is many orders of magnitude beyond humans.

Even if you don't want to.

If you don't, you'll get outcompeted by people who do.

Luckily it's not performance enhancing drugs.

But it does still have the red queen dynamic.

Software for cozy communities is radically underserved by current software business models.

That friction can come in the form of adversarial clashing.

Having agents that clash, similar to Generative Adversarial Networks, helps make them work.

If you don't have an opponent there's less of a reason to improve.

The agents can restructure their environment to reduce future uncertainty.

That is, to counteract entropy.

That requires patience.

LLMs have infinite patience!

It's like a fossilized foot print of your life.

What if the data could come alive and can do things for you?

Your data should be able to blossom in the right context.

LLMs will never be perfect.

Resilient systems assume that they can be confused.

I want to never have to think about apps again!

It used to be hard to unlock that energy.

Abundant cognitive labor blows it wide open.

LLMs are abundant cognitive labor.

Consumers never had the resources to dispatch cognitive labor on their behalf, so extra labor is fully additive.

Companies, however, have long paid employees to be cognitive labor.

Now LLMs can displace some of that labor, which is subtractive by default.

A prosthetic.

A new kind of limb.

"Why It's Better for Us to Think of AI as a Tool than as a Worker."

Every input to them, if scoped small enough, it will do what a reasonable person would with that information.

So if you make the structure clear enough, they can auto-converge.

If you have just the right amount of meta-structure then LLMs can be default-convergent, instead of default-divergent.

Once you have that, you can pump infinite cognitive labor into it.

It's now, for the first, time, disposable.

It complicated my mental model by reminding me that before the cloud, there were opaque proprietary file formats, giving apps more power over their users.

Apps always had a leg up over users, it wasn't a new thing with the cloud.

The expected danger of them is the multiplication of "powerful" and "naive."

LLMs with access to sensitive data are the most confusable deputies ever!

But software has to be made for a market and the set of needs anyone in that market has could be a million.

Software creators have to cram into one place which is fundamentally limiting.

So you end up with an equilibrium of maximally used, minimally liked.

If you design software for 10 people, you can talk to each one individually, you don't need UXR!

The way it used to work:

A company creates software, at great expense.

They have to do it for a market, and come up with something lowest common denominator that everyone in that market would want.

Then when users use it, their data accumulates on the service provider's turf, almost by happenstance.

But now that the data is on the service provider's turf, they can do whatever they want with it.

That deal only made sense when software was precious.

If all software is commodity, why would users ever put up with that deal?

They used to be able to add more value than you could, so the deal worked.

Now that's reversing.

The aggregators act all cuddly and friendly.

But that's only because they know we're stuck and can't leave.

Under the covers, their business is powered by taking your data hostage.

When we are no longer OK with that, when we reclaim what is ours, the knives will come out and they will no longer look cuddly.

It will be a mask-off moment.

Last year Slack changed their API policy to make it orders of magnitude harder to extract your own data.

That was an opening salvo in what is rumbling along as cold war, but will become a hot one as the power of unleashing LLMs on your data heats up.

The origin acts like the data you've saved in their tool is theirs.

No origins should be able to prevent you from downloading the data you uploaded to them.

If an origin tries to prevent you from reclaiming your data, you are within your moral right to take it back by any means necessary.

It's your data, don't let anyone convince you otherwise.

It does for enterprise, but not consumer.

If you aren't paying for your compute, it's not working for you… it's working for somebody else.

Or that mitigate the danger of naive LLMs confusing themselves.

But still nothing in the market that credibly mitigates prompt injection.

At the beginning, everyone thinking it's reasonable is critical.

After the ecosystem has significant momentum around that schelling point, the bar for "forking" the ecosystem gets higher and the bar for "unreasonable characteristics of the owner of the schelling point" also gets larger.

We're either in an ongoing punctuated equilibrium, or the singularity.

Then again, humans have basically been a singularity since we invented language and transcended the pace layer of biologic evolution.

We've transcended many other layers since then, this is just the most recent one.

Maybe each new pace layer transcending via a new technology is a punctuated equilibrium for the human race, extending the meta-singularity of humans.

Everything changes, in surprising ways.

How far can you throw a ball from your current position?

But what if gravity reversed direction?

Now almost no matter how you throw it, the ball will go infinitely far.

One inversion, with infinite consequences.

The ends fades away and all that exists in your mind is the means.

Don't do it for the person you're helping grow.

Put them in the conditions where they can do it themselves.

Conscious effort requires focus.

Subconscious is effortless, the most natural thing in the world.

That process of abduction requires repeated experience to capture.

A low friction learning experience on things you don't care to push yourself on you don't learn at all!

I've constructed an external project manager to tame my attention loop.

I use my internal neurotic conscientious loop as the raw material to power it.

But that system is not me, it's outside of me.

It's just a tool that I have decided to never put down because I find it so useful.

It's a goal with a frontier of 0 to n obvious next-actions that each help you get incrementally closer to achieving the goal.

The set of obvious next-actions are your "adjacent possible."

Sometimes the next-actions are not obvious and must be discovered, but once they're discovered they can be compared so you can pick the right next action.

That's the defining question of the modern era.

No one has time to distinguish between the two.^[bd]

Modern society is awash in Gilded Turds.

They might not have a cognitive immune system to protect against informational junk food.

They never had a situation where there wasn't a constant stream of hyper-interesting information.

Other companies have cloud services that make money based on usage.

That means other companies are incentivized to optimize engagement.

Apple, in contrast, only really wants you to buy another Apple device when you're in the market for one.

Apple doesn't care how much you use your iPhone… as long as you use it enough to value it and thus want to buy another one when you're in the market for one.

It's easier to plant seeds, tend to them, and judge them.

That implies the balance is moving away from exploit and towards explore.

True in business strategy and the economy.

"Meaningful" here means something that others value.

If others don't value it, then they won't yield to the owner's advantage.

Some people can intuitively swim to it and then get stronger and stronger the longer they're there.

They look to others like they cause the hurricane, so in a very real way they do.

It happens even faster when a system is optimized.

Optimization on a few measured dimensions leads to entropy accumulating on all of the unmeasured dimensions.

The Goddess of Everything Else is about this phenomena.

Effective cooperation overpowers and outcompetes everything else.

It creates positive sum out of what was previously zero sum.

How does it do it?

It moves in a dimension that is invisible within the zero-sum game.

Transcendence.

In challenging environments we need community to survive physically.^[be]

When you are interdependent over a long time horizon, you must operate as a community.

In modern society we are interdependent but in a way that is totally fluid, each person in your life is just a cog in the machine around you who you might never see again.

Modern society makes it so we don't need a community to survive physically.

But we still need it emotionally.

It's no wonder we feel so empty…

You feel it for children but not for your spouse.

The bonds of being a parent are unbreakable.

Contrast it with the Dark Ages.

Now, instead of not enough information, we have far too much.

As the social media landscape becomes overwhelming and cacophonous, we retreat to cozy communities.

Cozy communities have high trust and expect discretion.

Higher quality discourse can happen in cozy communities, but it also means that insights are not necessarily shared outside the group.

The insights are meaningful state; they accrete in the communities that people are invited to.

The people invited in these communities get compounding benefit; the people not yet in them fall behind.

The best is a resonant experience.

It's enjoyable in the moment.

Afterwards you feel actively proud.

One way to detect it is the "candid aim" test.

Is the stated aim actually explanatory for how the system actually works?

If not, then the system has confused its means for an end.

Even a tumbleweed has revealed preferences.

To have agency there has to be a second-order goal.

For any collective (of more than one agent) that must be different than the goal of the collective.

Goodhart's law arises from this misalignment.

Agents want to maximize their own value (capping downside of getting fired, while maximizing upside of reward).

Imagine that a doctor uses a system to research and recommend a prescription.

The system's agent accidentally issues the prescription without confirmation.

Who is liable?

If the doctor is liable for using the system, then doctors would only want to use systems that caps downside, and provably gets final confirmation from the doctor before doing any possibly-dangerous action.

At the first you pioneer and explore new territory.

Then you find valuable resources and settle down, expanding your territory.

Then you optimize and reform your society to make it better.

But reform doesn't innovate, it only maintains what you have.

Reform can become an end unto itself, an infinite loop that traps you.

You need to transcend again by exploring.

This process is often meticulous and careful, which means it gets off to a slow start.

But if done right it can compound on itself.

A ramp-up process will be like a Grubby Truffle.

Starts off slow, but then becomes unstoppable.

Someone this week told me about an agentic workflow they had that took an input image and some text and made it sparkly for social media, using Nano Banana Pro.

The workflow started producing images that looked like a design-illiterate programmer had made them.

It turns out that Nano Banana Pro was down, so the LLM, ever eager to please, decided to use Python drawing libraries to produce the output.

The slime mold will fill every bit of the container it's put in.

So give it the right jello mold to give it the right shape!

It will explore all of the nooks and crannies, and find cracks in the wall to pour through.

Life, and LLM swarms, will find a way.

They just eat away at it.

If it's not regenerating itself, then it just gets ground down.

Walmart can marshal insane amounts of demand.

If a brand is about to invest in brand-building marketing spend, Walmart forces them to invest in lowering the price instead.

This benefits consumers and Walmart, but it harms the supplier.

The supplier becomes increasingly dependent on Walmart, and Walmart has more leverage over them.

Premium brands might be hollowed out by consumers seeing them as cheap, transactional.

It's a monkey trap: short-term benefit to the brand at long term cost.

Sometimes overlapping at the beginning but as you get further along they diverge more and more.

The ones that get used the most bubble up and get used more, naturally.

Convergent emergence.

A generic desire is not specific enough to overcome static friction.

People who don't know (or don't care) about the plan who have real leverage can do significant damage.

They'll keep on creating things that you don't like, and be hard to steer.

Just because you can execute a playbook doesn't mean you could have created it.

To create one requires meta thinking, two-ply thinking.

But also you have a terrifying moment of "can my feet keep up?"

My strategy was to sprint through the exam and rough-in answers as quickly as possible.

While I went, I flagged answers that I wasn't as confident in.

Then, I'd go back through and look at each flagged answer in more depth until I was confident in it.

This breadth-first approach allowed me to get to good enough as soon as possible and then tighten the result until the buzzer went off.

I wrote papers the same way.

This strategy naturally focuses time on the things that can get incremental benefit, and allows you to juggle multiple things naturally.

You need to want to do it even if you had to crawl through broken glass.

Otherwise you'll find yourself questioning or even resenting your decision when you run into hardships.

But if you want to do it no matter what, then it can be the most meaningful experience of your life.

Precisely because you are so committed so you must learn and grow from the hardships that are inevitably a part of it.

Your abstract ideal cannot survive the collision with the actual messy reality.

Keep your heroes on a pedestal.

Believing that someone can be infinitely good helps you strive to be better yourself.

The only answer is a dynamic one.

That confounds people!

But it doesn't unfold that way!

Your actions matter.

Technology is not inherently good.

It is amoral.

You can't just build it and assume it will be good for society.

You have to help it unfold in a positive way for society.

It emerges.

That means that philosophy is critically important, to give a prosocial asymmetry to the underlying decisions.

Makes it structurally more likely that what emerges is good.