Bits and Bobs 6/15/26

Freeman Dyson writes:

"It often happens that a scientific revolution is accompanied by a change in style.

I like to use the names of Napoleon and Tolstoy to symbolize two contrasting styles: rigid organization and discipline represented by Napoleon, creative chaos and freedom represented by Tolstoy.

In the world of computers, Napoleon is the massive IBM main-frame; Tolstoy is the humble Macintosh.

The computer revolution was an escape from the Napoleonic ambitions of von Neumann to the Tolstoyan anarchy of the Internet.

Future revolutions will bring more such escapes."

Alan adds:

"The big AI companies are the apotheosis — literally, in the view of many who work for them — of Napoleonic science.^[a]

The open web and the world of hobbyist and small-scale devices (often built on the Raspberry Pi) are our remaining refuges of Tolstoyan computing."

What if we took Tolstoyan computing and applied it to using AI?

The ones that get accelerated are ones that can be overseen by one human.

The ones that get slowed down are the ones that require coordination between multiple humans.

The main difference is whether human coordination is in the loop or not.

Human coordination is the bottleneck.

If you're simply building normal software with it, you aren't seeing its full power.

With access to Fable-class models, you can now build 10x more powerful software.

Agents shift from tasks to responsibilities.

That allows Anthropic to make a step-change TAM claim.

A "Tool" prices like software.

A "Responsibility-bearing colleague" prices like labor.

If you apply it to small problems you can't go that far.

When a more capable model comes along, much of that structure that was necessary to get the outcome you wanted with the old model is now unnecessary or actively distracting.

The structure went from scaffold to cage.

This happens every time there's a paradigm shift in capabilities.

Tired: laptop closed meetings. Wired: "no-agent-feeding" meetings.

"Your new assistant can schedule a meeting but it can't fix our broken world."

Before you cross that rubicon, there's a ceiling of what can be done in the system.

After you cross it, anything is possible (with enough effort).

That open-endedness is power… which is also dangerous.

I'm seeing a lot of outlets say this is the way to make ChatGPT safe, but that's wrong.

It just cuts off the most useful functionality so it can't be used against you.

In today's security paradigm, you get powerful or safe.

To get around that you'd have to change the laws of physics.

"just asked codex to pull investors meetings from my calendar and draft an email, it went ahead and emailed all of the*"

Powerful or safe, pick one.

"Uhm, I just discovered Codex was searching my slack messages without any active thread in my Codex telling it to do so...

And I can't inspect the cursor or anything to tell me why it was taking its actions?"

Almost tautologically.

I don't trust you to be proactive if I don't trust you.

Trust is about track record in that context.

The more trust something earns, the longer leash it gets.

The more autonomy it can be granted.

Today's agents have the former, but not the latter.

See also: OpenClaw used Gavriel Cohen's code and exposed the AI Agent accountability problem.

PCWorld: OpenClaw was too dangerous for your PC. Now it's the blueprint.

Securing CI/CD in an agentic world: Claude Code Github action case.

A new attack: A Fake Bug Report Hijacks Your AI Coding Agent – and Nothing Catches It.

Another new attack: Phishing for Lobsters: How We Tricked OpenClaw into Spilling Secrets.

Yet another new attack: Compromise OpenClaw with Prompt Injections in Message Objects.

Indirect Prompt Injection remains a fundamental security challenge for AI.

A new technique: dangerous worms say forbidden things so the LLMs in security scanners will stop scanning them.

What the OpenClaw vulnerability reveals about the future of agentic AI security.

WebMCP Can Be Used To Hijack AI Agents, Chrome Warns.

Prompt injection is not a sufficient framing; agents fundamentally undermine the assumptions our old software is based on.

We need new physics of trust.

We need integrity-based security.

A few examples where people are realizing this.

Everyone Is Buying AI Guardrails. But Agents Have the Keys to the Car.

It was an idea designed by Jefferson.

Normally, parcels would be split up based on an understanding of the terrain: where the rivers are, where the valleys are.

But the United States frontier was too vast; it was much easier to simply project a cartesian grid across the US, sight unseen.

This created some oddities: parcels that required multiple bridges to connect, for example.

But it made the problem of partitioning the whole interior of the US tractable.

It was too hard to be precise, so they just sliced across the grain of the land.

This is what we did with the app's same-origin model.

The origin boundaries are tidy, easy to reason about for the OS, and simple… but they cut across the grain of our lives.

Before, it was not feasible to be more precise.

But now, with LLMs, we can be!

It's the head!

The physics of trust we use today for software cut across the grain of our human lives.

Vertical silos by use-case, not horizontal by people and relationships.

As a result, tons of software that should simply exist does not.

It used to be that there were certain domains that were too complicated to be operated by anyone who wasn't a specialist.

Now the LLMs have the background knowledge to execute problems that require specialist knowledge^[b].

The situated knowledge, curiosity, and judgment is what is most important for the human to provide.

Generalists who are curious and proactive will thrive in a new AI-first world.

Fixed role descriptions will fade away… generalists will adapt to what the problem domains at the moment, marshaling their agents who have all of the specialist knowledge necessary.

"Getting the machine to do work is the easy part.

The hard part is deciding when the work is good enough to leave the machine.

That decision is going to define a lot of white-collar jobs, and not because everyone will learn to code.

More people are simply going to start receiving work from machines they did not supervise.

The first time it happens, it feels like magic.

The tenth time, it feels like management."

They are very high volition.

They are well informed.

They are willing to do the work to move to a better option if it shows up.

How much do people want to join in?

Once they do, how hard is it to switch to any newcomer?

Network effects used to work for developers.

But now agents undermine both of these.

Developers all use agents.

Agents make it easier to create anything the product already has for themselves.

Also, agents make it easy to move to another one if another one pops up, especially if users have all of their own data.

So in the world of agents, the network effects for developer-focused products are less strong and less important.

"The forgotten part of Aesop's fable of 'The Boy Who Cried Wolf' is that the wolf did eventually come."

"What is so fascinating about Anthropic, however, is that while I am sure some executives at the company are thinking this way, I also totally believe that they — and the employee base broadly — also happen to believe that they are doing the right thing.

It's fascinating to observe: me, the rational business analyst, sees a hard-nosed but understandable business decision to cut off would-be competitors; Anthropic employees and advocates, the true believers, see a regrettable but understandable safety decision that ensures that responsible and thoughtful people — themselves, of course — will be the ones guiding our AGI future.

This is true alignment, and it's an incredible accomplishment.

I continue to think that part of what led to so much drama at OpenAI was the misalignment between what was a research organization and the tremendous opportunity that was dropped in their lap with the unexpected success of ChatGPT; that misalignment led to both the research organization losing talent and slowing down even as the company failed to fully capitalize on its early lead.

Anthropic, on the other hand, has somehow convinced itself that every decision that optimizes its business outcome is done purely for altruistic and culture-affirming reasons.

It's impressive!"

Kind of reminds me of Apple's strategy credit of privacy.

The Silo Wars are heating up!

The Max/Pro plans give you the all-you-care-to-eat access to n-1 generation models.

The n generation frontier models will be available at marginal costs.

That allows customers to self-discriminate into the more expensive models.

As the models become cheaper to serve, they open up to more customers.

"Agent as intern" is dead, now it's "agent as pair programmer" if you're lucky enough to be as capable as it is.

A friend commenting on my workflow: "That workflow is so two days ago."

An example of a post-Fable workflow.

You already see this with the engineers who can afford to pay $200 a month for Pro/Max plans.

Recent grads probably can't stomach paying that much.

The people who already have enough resources to be able to pay that cost without thinking will pull ahead.

This happens fractally down the stack.

The engineers who can use Fable-class models vs Opus-class.

The people who can use Pro/Max vs the people who are just paid ChatGPT subscribers.

The people who are paid subscribers to the chatbots vs the people who are on the subscription plan.

The people who can use the better model before others have superpowers.

It wasn't until later it became obvious how useful it would be for individuals.

For example, maybe in the first run-through, it decides to defer a key bit of work you want it to do.

Even if you correct it, now in the conversation history is the decision it made to defer work… making it more likely to do the same in the future.

Cleaner is to rewind the conversation to before it made the incorrect decision, and make sure your prompt makes the incorrect branch unlikely, then let it play forward.

That way, the agent thinks it's in a universe where it always made the right decision, and it doesn't ever get distracted by wrong decisions it made.

Kind of like the movie Edge of Tomorrow.

Turns out having the whole internet in your brain makes you schizophrenic!

If you can switch models easily it's less of a problem.

We didn't mind this too much with Google Search.

Not only do many queries have AI-generated results.

But also, every result above the fold is an ad, styled to be indistinguishable from the organic results.

This is what you get from rolling down a gradient.

Still, it would be shocking to see for someone from 15 years ago.^[c]

A wrinkle: many of the same people are still working on search from 15 years ago.

The villain isn't necessarily someone else who doesn't care about the ends as much as you did… it can sometimes just be a degraded version of you!

An off-color Red Dwarf episode on this same observation.

What advice do you give?

Even the deep insights you learned would be hard to impart.

You can only learn the deepest insights from experience.

This is the Hallmark card fallacy.

The deepest insights you will discover are ones you've already heard a million times before... you just weren't ready to hear them before.

That means you can't pass them onto others, no matter how meaningful the revelation is for you.

You're going to pay the cost, you might as well get the benefit!

Once you have something that works and is usable, you can do Throughline Tacking to discover the main goals of the product.

As you iterate, you make the product, and its implementation, more what it wants to be.

LLMs can help with this process, especially on implementation.

This allows you to race to getting something that works, and then iteratively make it more robust.

How quickly you can move in a codebase is tied to the rewrite count.

The rewrite count is how many times you'll rewrite the codebase in the future.

If the rewrite count is high, you don't have to worry about getting it perfect, just load bearing.

LLMs make it easier to do rewrites, so the rewrite count is higher.

That means that you can move faster.

"I think skills — the little markdown-and-prompt bundles that have become the dominant unit of 'thing the model knows how to do repeatedl' — can have a much higher ceiling, just like coding does, when you make skills that work on skills: Higher-order skills.

And the gap between the two shapes of composition, the two ways you might think you're combining skills, is the difference between getting cake and getting recipes for better cake."

If you're making it useful then you have to go deep.

(As long as they are competent at their job!)

They will think about the indirect effects of their actions and account for them.

"Sitting in the same seat" means they don't expect to be fired or want to leave because they don't like the role.

People who will still be in the seat have a naturally long-term perspective, since any problems in this org they create will become their problems.

Organization leaders spend so much time trying to "avoid wasted work."

Going from 20% wasted work to 10% wasted work requires 10x the effort.

Some portion of that "wasted effort", especially if the team is high-autonomy people who are acting like owners and expect to be sitting in the same seat in two years, will actually be acorns.

Instead of making sure no wasted work happens, how do you make sure that great things are happening?

There's so much more you could do!

This means that by default, team members will gravitate to different parts.

Because it goes from "one thing among many" to "the only thing that matters."

If a user thinks, "Is that a mistake?" then it is a mistake.

It keeps itself spinning, in a critical state, ready to be nudged one way or another by a thing thrown into its loop by the outside world.

You can have peer loops or parent loops.

The whole system gives you the fractal chain mail shape.

Each dot is a topic.

You're trying to get from left to right along the top row.

At each point, you may go from left to right if you're in the top row, or down.

But you may not go to the left or right on lower rows, you have to pop back up.

This helps make sure you don't get lost in the details.

They set the floor for team members to hit.

When team members are actually enrolled in the goal, it enables a ceiling.

Acting like an owner, not a renter.

In the former, you want to clear the bar.

In the latter, you do want to do your best.

Floor vs ceiling.

Writing good jokes is extremely hard to do.

It requires a deep understanding of people's context and priors.

Also, where the line is… and how to toe that line without going way over it or way under it.

"Poop!"

At which point the five year olds all burst out laughing.

… The actual punch-line is "dung."

"You can't use up creativity.

The more you use, the more you have."