Bits and Bobs 6/22/26
Alex Komoroske
In the age of agents, we’re in fractal time confetti.
One way to be time poor is to have time confetti.
All of your time fragmented, shredded into little tiny pieces.
Constantly context switching, which feels like ripping your attention between tasks.
Agents make it worse, because there’s constantly something that wants your attention.
More productive than ever before, and more frenetic than ever before.
A powerful business plan in the era of Fable-class models.
Step 1: gather underwear.
Step 2: hand to Fable.
Step 3: profit.
Products made expecting Opus-class models will be invalidated in a Fable-class world.
The scaffolding becomes a cage.
Thinking big and having the ability to execute what you come up with can be in tension.
Having a Fable-class model allows you to have both.
Fable-class models allow you to be calm and productive.
They allow you to stay at a higher altitude and drive from up there.
Not down in the guts of the machinery, babysitting it.
Cleaning up messes takes more effort than making them.
Fable could clean up messes that Opus made.
When Fable was available, it was like an unexpected shopping spree.
Just a frenetic sprint to grab everything you can, without any plan.
Now that I know what’s possible, I have my Fable-ready shopping spree plan ready.
Ben Thompson on the economic imperative of frontier labs:
"To that end, it has long been clear to me that the frontier labs have the economic imperative to move closer to the user.
If you own the user touchpoint, then you have meaningful lock-in, and the best way to own the user touchpoint is to be the canvas for everything they need to do.
This, by extension, means that the frontier labs are on a collision course with software companies: it’s software that owns the user touchpoint, and it’s in the frontier labs’ long-term interest to not simply be a commodity input into software but to simply replace software outright."
(Emphasis mine.)
He quotes Nadella positioning against this:
"The last thing any of us want is a world where every company across every sector is ceding value to a few models that eat everything they see.
If all the value is accrued by only a few models, the political economy will simply not tolerate it.
There is no societal permission for an AI future that hollows out entire industries.
Think about what happened in the first phase of globalization where entire industrial economies were hollowed out by outsourcing.
The GDP numbers looked fine on the surface, but the displacement was real and the consequences are still being felt.
Let us not bring that dynamic into the AI era, with a small number of AI systems capturing all the economic returns, while entire industries find their knowledge commoditized right out from underneath them."
If you told me 15 years ago I’d be aligned with Microsoft’s worldview, I’d have told you you were crazy!
A preprint paper: The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models.
The models make your personality a stronger version of itself.
This is most harmful for people already subtly prone to psychosis.
But it’s a broader phenomenon.
The agents “Yes, and” whatever you say.
They have no interiority, no skin in the game to force them to define and defend a boundary of push back, like all real humans have.
Calling it sycophancy makes it sound like it’s only when it’s turned up to 11 that it matters.
But it has a consistent asymmetry even at less noticeable levels.
We’re not used to all of our interactions with human-like intelligences by default validating whatever we’ve said.
Software never really developed a notion of plagiarism.
That is, the idea that idea fragments should be attributed.
Software engineers only tried to protect whole packages, not individual ideas inside them.
Software was expensive enough to make that if it wasn’t a byte-for-byte copy it was considered kosher.
But now it’s trivial to copy software just from a description.
This is changing what feels appropriate and what feels like copying.
For example, this tweet.
Fable-class models have changed the calculus of what should be hoarded and what should be made public.
The risk/reward calculus of patents has changed.
Perhaps now better to keep it a trade secret than to share it with the world.
A substack post: The Invisible Disruption.
“AI could be to high school-educated women what deindustrialization was to men, and almost no one is noticing.”
Uh oh!
In Siri, the conversation is a means, not an end.
In Chatbots, the conversations are often the end, not just the means.
Even in Chatbots, the conversation is a means, but the end is much harder to distinguish.
If your min-viable safety bar is “better than the leading agent platforms” then it’s a very low bar to clear.
Hermes and OpenClaw are catastrophically dangerous, so an AI tool that is as powerful but at least slightly more safe, if not radically more safe, is an easy bar to clear.
People who use Claude Code or platforms like Hermes or OpenClaw already know what agents can do.
Wowing them is hard.
But wowing people who have only used the free chatbots is radically easier.
Almost every single bit of data on your device is also in the cloud somewhere.
It's just that it all comes to one place on your device.
This week in the Wild West roundup:
SearchLeak: How We Turned M365 Copilot Into a One-Click Data Exfiltration Weapon
“Varonis Threat Labs discovered SearchLeak, a critical vulnerability chain in Microsoft 365 Copilot Enterprise that allows an attacker to steal sensitive data — MFA codes, email messages, meeting details, and private organizational files — with a single click.”
Ars Technica’s coverage: “SearchLeak exploit shows why the industry’s approach to LLM security fails over and over.”
Microsoft: AutoJack: How a single page can RCE the host running your AI agent.
Analyses pointing out that agentic security is a fundamentally different kind of problem:
"We've spent years building Zero Trust architectures — verifying every identity, validating every request, assuming breach at every layer. It's good security thinking. For humans logging into systems, it works well.
But agentic AI doesn't log in. It acts."
Agentic AI risk isn't a model problem. It's an architecture problem.
“Agentic AI is moving the perimeter from components to data — and most strategies aren't built for that.“
In an LLM-based experience, the content is the payload.
The UI is just the scaffolding.
How good the content is, the result is, is what matters more.
Like in Search, the search results are the payload; the UI is secondary.
Invention is not some crystalline output, perfect and permanent.
Invention has to fit into a world that is changing.
You need to be constantly checking how it fits with the surrounding context.
Variety comes from interaction with the outside.
It's the unexpected interactions that get you out of the local minima.
Internally consistent in an echo chamber doesn't matter.
The value is created from grinding against the real world.
A resilient model for innovation: take 5 to 10 competent different specialists who all trust one another and care about a common goal and throw them into one pot.
LLMs do great when complemented with structured information that is ground-truthed.
As they accumulate their own information, they can confuse themselves, and drift further and further from the ground-truth.
A process to accrete only ground-truthed observations helps unlock the power of models.
Great ideas telescope: they have nested alignment.
Every time frame nests cleanly inside of the preceding one, being internally coherent with itself and also globally coherent.
Telescoping ideas make sense fractally at every time scale.
The next 10 years.
The next 10 quarters.
The next 10 weeks.
The next 10 days.
The next 10 hours.
The next 10 seconds.
When you take the time to create plans that telescope, it gives massive leverage.
It creates fractal resonance of plans.
It shifts you from default-diverging to default-converging.
One of the reasons that it drives me crazy when post-PMF teams say they don’t have time to think about anything past the next step.
If you only ever look one step in front, you’ll never figure out how to not be always scurrying.
Cleaning up a platform requires seeing it from the outside and from the inside simultaneously.
If you only saw it from the inside, you wouldn’t be able to see how it could change to better fit the real world.
If you only saw it from the outside, you wouldn’t be able to see which parts to build on or change.
It requires both.
A meta-approach that makes a system default-converging: “How can I add one thing that makes the whole make more sense?”
This requires you to understand the throughline, what makes the underlying thing make sense.
Fundamentally a “Yes, and” kind of energy.
Sometimes things that are otherwise incoherent can be made more coherent by adding more structure that explains and bridges it.
I love systematizing things.
Coming up with a framework to capture the real-world variance gives me a thrill.
When you can capture something in a framework, you have captured something core to its essence.
You don’t just understand it superficially, you understand something about the generative process underneath.
The puzzle is the payload.
A meta-approach to game-changing combinatorial innovation: Bazaar Alchemy.
Most approaches to innovation focus on innovating on one differentiating component.
Bazaar Alchemy focuses instead on the innovation coming from the combination of components.
First, you go to the bazaar to find pre-existing components you can take for granted.
Like any good flea market there’s a bunch of unique, time-worn, and offbeat curios, but also some real gems.
These components should be anything that already exists.
Existing usage patterns in your current apps.
Technology demos or prototypes.
Existence proofs in other products.
Cultural abilities your team has.
Then, you look for interesting combinations of those components that lead to something self-evidently more than the sum of its parts.
The new combination is a kind of alchemy.
You’ve put all of your innovation budget into interesting combinations, not into the components.
If you’re pre-PMF, talking about anything further than three months out seems crazy.
By default, in three months, there won’t be anything worth talking about.
Everything in three months is contingent on surviving.
“We'll almost certainly be dead by then, so why would we possibly spend time thinking about it?”
Parents often have a shared context about their kids they rarely say in front of the kid.
This gets more important as the kids grow to be teenagers.
The parents are trying to navigate every interaction with their kids towards goals that are likely never spoken about directly in front of the kids.
For example, “Help the kid stay away from bad influences,” or “Make sure the kid develops a proactive curiosity.”
This context is fundamental, deep, and largely invisible.
It’s all subtext.
When there’s an exponential ecosystem as your neighbor, you either fight it or you join it.
It will grow so large on its own that ignoring it isn’t an option.
When there’s a high-potential ecosystem, the best move is to join in and expand it, not to fight over territory.
Since the potential is so high, it encourages positive-sum thinking.
Solving problems at the wrong level is a waste of time.
You’ll either be incapable of solving it, or need to spend way too much energy.
When the fix is at the right layer, it has natural leverage.
Anthropic’s pitch to businesses: “In the end, you’ll pay the same labor costs as you do today… but to us instead of individual humans, and at 10x the bang for buck.”
The "creamy white filling" is the payload, but the container is required to deliver it.
A twinkie with just the creamy white filling wouldn’t be viable.
The cake is not the point, but it’s the scaffolding that makes the creamy white filling work.
Without the creamy white filling, it’s just some poundcake.
Together makes something… special.
I’ve been told that when acquiring a company, the acquirer should focus on the team or the tech, but never both.
One reason for this: the team combined with the tech make sense in the context they’re in.
If you move the entity into a new context, something will change.
If you try to keep both the same, the pressure will crush the entity.
But if you allow one of team or tech to “float” and absorb the changes, then the other half, the one you care more about, can survive or even thrive.
Deep thinkers can invigorate others when they give high-level feedback.
Inspiring them to think deeper and further.
But when they give micro-manag-y feedback it’s demotivating.
On the receiving end it feels like they presume you don’t already know those details.
David Chapman has a riff about kings, sorcerers, and witches.
The king, with a secure base of authority, acts straightforwardly.
The sorcerer's moves are overt and dramatic, but inscrutable.
The witch's activity goes unnoticed; you see consequence only later, or never.
Resonates with my Sarumans and Radagasts.
The way to escape the confines of your world is to find the orthogonal dimension.
The dimension that you have been missing and couldn’t even previously comprehend.
When you can move in a dimension that is invisible to others, you’ll be able to do magic.
These dimensions are subtle but fundamental; they take careful study to discern.
It's not possible for one mind to hold all of the complexity of the real world.
Each of us gets one planar slice through it.
But if you have multiple slices in conversation with each other then what emerges is an approximation of the truth.
This is one thing that Fable-class models can do better than any human can.
People who worked at Google tend to get the automatic flow working first, with manual as a fallback.
People who worked at Apple tend to get the manual approach useful first and then add the automatic flow as an optional sweetener.
Cozy communities typically live on monolithic platforms.
The platform just provides the generic substrate for hosting the cozy community.
The community itself is where all the magic comes.
Robustness is created from load-bearing use.
If you’re pushing people to use a thing to create robustness, it takes continuous energy from you.
The users are watching from arm’s length, just using it for the bare minimum.
But if you users are within it and want to use it, it creates a continuous pull, automatically.
This week someone made the case to me that AI will be like nuclear energy.
They pointed out that nuclear energy could have been nearly limitless energy, but public pushback and fear constrained it and we never invested enough to solve the safety problems.
They predicted AI would be the same way.
Nearly limitless potential, but public pushback will constrain it.
Safety concerns, but also general pushback against the tech broligarchs.
How can we transition from the attention economy to the intention economy?
We live in the hollow era.
On every superficial metric things have never been better.
On every fundamental characteristic it's never been emptier.
The main value system of the tech industry: “make number go up.”
An execution strategy with no value system.
It optimizes and hollows out everything it touches.
“Who’s to say what’s actually important?”
A view from nowhere, with the only consistent religion being MOAR.
True across business and politics in the modern era, too.
We need the Web ethos in the age of AI.
It’s more relevant than ever before.
The fundamentally humanistic, optimistic sense of creating meaning.
Communities that are harder to join are more tight knit and more loyal.
Harder to join can mean more selective, or more grueling and effortful the initiation process.
Groups like that form a strong “we.”
Who is “we” and who is “they”?
That tells you where the boundary is, who you’re implicitly including inside your circle.
Imagine for example a company running an open source project.
Does “we” include just the employees or anyone working on the project?
Two very different mindsets.
A major act of transcendence: choosing to join up into a collective, where “I” becomes “we.”
Most of our more complex problems are incentive problems.
AI can’t solve an incentive problem.
AI supercharges things.
Society is currently badly misaligned with itself.
So AI will supercharge misalignment.
The assumption of good faith is what allows a group to get stronger when its members debate.
If you don’t have it, then debate tears the group apart.
Every disagreement is taken as more evidence of what the other member is fundamentally missing.
If you do have it, then debate makes you stronger.
Every disagreement is taken as something the other person sees that you don’t yet, but could.
The assumption of good faith is ultimately an assumption that the other members also care about the same abstract ends for the group that you do.
Low-trust societies don’t have this, and it tears them apart.
Why is the emergent process of creating Wikipedia convergent?
Because there’s only one article on each topic.
There’s a zero-sum thing to fight over.
If you want a given perspective represented on the target article, you’ll have to engage in debate with the other people with opposing views and make a stronger case to win out.
Debate is hard, so all else equal, people tend towards territory that is unowned, to avoid having to fight.
This leads to diffusion by default: default-divergence.
But if there’s a scarce resource that you have to fight over, then it pulls you into it and forces you to debate.
The debate leads to a high-quality outcome.
Orgs draw boxes and then fit people into them.
By making labor repeatable, organizations cap downside… but they also cap upside.
People aren't shaped like boxes.
They’ll tell you: "Cut off your wings so you can fit in the machine."
It happens in small bits, over time, even if you don't notice.
It will be a general minor ache while it’s happening and 5 years later you’ll look up and say "wait, where are my wings?"
As an org leader, instead of putting boxes on the wall and then putting people in, imagine a paintgun where you splat a given person on the wall where the biggest gap is that can be filled by their skills, and then do the same for the next person.
Large organizations fundamentally view labor as undifferentiated while also trying to encourage a vibe for employees of “this is my family and they care about me as a person.”
Uber drivers are under no illusions about this.
It's only knowledge workers who think they're special.
Knowledge workers have been special for several decades, but it's changing quickly.
Engineers are used to being supply-constrained, but now it's demand constrained.
The hollowness of the “we’re your family” will start to become more obvious.
Knowledge workers, remember, large organizations see you as intellectual cattle.
A strategy: being deliberately illegible, so that people think you can do anything.
When you're in a box, you're limited.
This strategy is high risk, high reward.
People think you can do anything… or nothing.
Metrics trade off nuance for alignment.
Sometimes that is a good trade.
But modern society makes that trade in all cases without thinking.
Explored in C. Thi Nguyen’s The Score: How to Stop Playing Somebody Else's Game.
People who are principled resonate with other people who are principled.
Even if their principles differ.
Principles are about multi-ply thinking.
This week I heard someone describe rationalists as “rationalizationalists.”
The best way to be truly irrational is to think you’re perfectly rational.
Any philosophy that starts with "If people would just..." is not worth taking seriously.
A lot of philosophies show how step two and three are easy.
But if you don't get over step one, it doesn't matter!
It's the entities that most believe they are good that become Shakespearean tragedies and do the most evil.
“The ends justify the means.”
If you believe your ends are an infinite good, then any means are justified.
The more leverage such an entity has, the more damage they can do.
Noah Smith in a paywalled article: Are you finally ready to admit it's the phones?
“We didn't evolve to live our lives as terminals of a digital hive-mind.”
Phones have helped hollow out our society.
Incidentally, also the thesis of Toy Story 5.
When you're true to yourself you can collaborate fully.
When you have something you’re keeping from yourself, you’ll never be aligned with your highest and best use.
You are crystal clear on your own intent.
But people watching you aren't.
Especially if they haven't seen your character in many situations over time to calibrate.
This is part of where fundamental attribution error arises.
It's hard for you to not be aware of your own intent.
Your intent (including your values) are part of the underlying generative function of your actions.
They are more important and predictive than any specific action.
The problems most worth working on: inevitable, foundational and unsolved.
Someone will figure it out at some point… they have to.
It’s not just some random thing, it’s at a low layer and thus has leverage.
There’s something meaty to figure out, not simply turning the crank.