A capability never stays where it enters.

In October 1962, in a report written for the Air Force from a laboratory in Menlo Park, Douglas Engelbart set down what I think is still the most useful claim anyone has made about tools. A new capability, he argued, does not sit where you put it.

He was describing writing machines. He could have been describing a pond. The stone stays where it landed; the displacement reaches every edge.

Here is the smallest example I know. A typewriter cannot take back a letter. Correction fluid, waiting for it to dry, retyping the page: that was the price of one wrong keystroke, and the price shaped everything above it. You outlined on paper first. You composed a sentence in your head before committing it. Then a machine arrived with a key that deletes. That single low-level capability, correcting a mistake for free, travelled upward and rebuilt the process of writing: type as you think, restructure when a better idea arrives, keep three openings alive and choose later. And it travelled downward, waking things that had always been latent in the person at the keyboard, such as developing several ideas at once, or simplifying only after you have over-written. Nobody set out to invent thinking in drafts. It arrived as a ripple.

When people say that AI is changing everything they are usually gesturing at a fog. I want to be more specific about the geometry.

Draw the stack. At the bottom, models: systems that read, write, reason and act on symbols at a cost approaching zero. Above them, tools: harnesses, agents, editors, terminals, the things that turn a model into a capability a person can point at a problem. Above tools, interfaces: how that person sees, steers, interrupts and comes to trust what the tool is doing. Above interfaces, a person with a purpose. Above the person, a team and its methods. Above the team, institutions: companies, markets, schools, law. And above all of it, quieter than the rest, the layer I care about most: what anyone in the system can imagine doing.

A capability can enter at any layer, and it never stays. Take the one I live inside, which entered at the tool layer: an agent that writes and comprehends code without anyone typing it. Upward, it has already reshaped the process hierarchy of building products. The steps that used to sit between deciding what to build and having it running have collapsed into steering. Teams that once babysat one change now run fleets of agents and review what comes back. Institutions are feeling it next: pricing, hiring, what a junior engineer is for. Downward, the same capability has pulled things up out of the model layer that nobody built harnesses for two years ago: memory that persists outside a context window, skills that encode a method, protocols for reaching into any system that has an interface. The model did not acquire those capabilities because someone asked for them. They became worth building the moment the tool above could use them.

Now run it in reverse. Introduce a capability at the level of one person, say the ability to read every incoming channel each morning and decide what matters. It ripples down into the tooling that person now demands, and up into how their team communicates, because the team learns that this person never drops a thread. Propagation has no privileged direction. It goes wherever there is a process that can reorganise.

I should say where I am standing, because a manifesto without a vantage point is only a mood.

The tool I do most of this in is Amp. What makes it worth naming is not any single feature but the way its makers keep adding and removing features as the models move underneath them. A map of every thread appeared and was quietly retired. Hand-off between sessions appeared, became unnecessary when context compaction got good, and then re-emerged as something better: threads that send messages to one another. I can ask a remote environment, with its own browser and terminal and file tree, to fetch what it needs from a thread on my laptop and report back. An orchestration agent watches all of them and tells me when something needs a decision. Specialised sub-agents handle search, deliberation and images. None of this is a demo. It is a Tuesday.

What it feels like from the inside is a change in the unit of action. I used to act in functions and files. Now I act in intentions, and the intentions are large: research this, argue that, design the page, build it, run it, check it, fix it. A single person with a direction can direct a small organisation of capable systems, and the organisation is cheap enough to be disposable. Because most consequential tasks cannot be done in one shot, the leverage lives in iteration: I look, I react, the systems move again. The interface that lets me look well is worth more than any single model behind it.

This page is one of the exhibits. It was researched, drafted, designed and built by directing exactly this kind of system, from a brief, a thesis and a handful of my own essays. Whether that makes it less mine is a question I find more interesting than threatening. The decision about what it should argue, and the judgment about whether it is right, stay where judgment has always stayed.

The first-order effect of any new capability is dull: you do the same things faster. The interesting effects are the ones that change what the same things are.

Engelbart borrowed a hypothesis from linguistics and made it his own: that both the language a culture uses and its capacity for effective thought are shaped, as they evolve, by the means it has for manipulating symbols outside the head.¹ If writing were physically harder we would keep fewer records, hold different concepts, build different institutions. Our ideas depend on how well we can handle symbols; how well we handle symbols depends on our tools; our tools depend on our ideas. It is a loop, and loops compound.

It has changed the questions, too. Spend time with someone who has lived with capable agents for a year and you will notice they have stopped asking whether they can. They ask how. They look at a personal inconvenience and think that it should be software, because the cost of a purpose-built tool has fallen below the cost of tolerating the problem. I saw the same shift at a smaller scale when I worked on tools for thinking: give people the ability to move ideas around a whiteboard and they stop asking what the takeaways are and start asking how the new thing fits what they already know.

This is the layer I care about most, and it is the one that hides. The imagination of action is not a thing you see on a roadmap. It is the set of moves that feel available to a person on an ordinary afternoon. Tools widen that set. Not merely possible, but thinkable, and then, because you can picture yourself doing it, desirable. Bret Victor, crediting Richard Hamming, has made the point that just as there are sounds we cannot hear, there may be thoughts we cannot think, and that thinking tools exist to bring them within range.⁵ I would add a corollary: there are actions we cannot want, until the tool arrives that makes wanting them reasonable.

There are four orders to this, and I want to hold them in view together.

1. Optimisation. You do the same thing faster. This is the order a benchmark can measure.
2. Restructuring. The process hierarchy reorganises around the new capability. Old steps vanish; steps nobody planned appear.
3. Latent capabilities. Things that were technically possible and practically absurd become routine, in your hands and in the layers beneath you.
4. Thinking. You ask different questions in a different language, and those questions decide what gets built next, which restarts the loop.

Most conversations about AI stop at the first order. The action is in the other three.

None of this is new, and it matters that it is not new. The line of thought this page stands on is eighty years long, and every link in it was written by someone who could see the next capability before the tool that carried it existed.

In 1945 Vannevar Bush imagined a memex, an enlarged intimate supplement to his memory, and observed that the storage was not the point; the process of tying two items together is the important thing.³ In 1960 Licklider hoped that human brains and computing machines would soon be coupled together very tightly, and that the partnership would think as no human brain has ever thought.² In 1962 Engelbart wrote the framework. In December 1968, on a stage in San Francisco with his team on a video link from Menlo Park, he showed what it looked like: a mouse, hypertext, windows, two people editing one document.⁴ In 1977 Alan Kay and Adele Goldberg described the personal computer as a metamedium whose content would be a wide range of already-existing and not-yet-invented media.⁷

A ripple does not know what it carries. The mechanism that spreads a good capability through a system spreads a bad habit just as efficiently, and a manifesto that pretended otherwise would not survive a second reading.

Some of the ripples I can already see are not ones I would choose. When the cost of producing something falls toward zero, the cost of evaluating it does not, and attention becomes the scarce layer. Systems that make everyone faster make everyone's mistakes faster too; a broken assumption in a harness reaches a hundred repositories before lunch. Dependence is real: a capability you cannot exercise without a vendor is a capability with a landlord. And leverage concentrates. The propagation that lets one person do the work of a team lets one organisation do the work of an industry, and it is not obvious that the shape which emerges is one we would have chosen.

So I have come to think that taste is not a decoration on top of this shift; it is the load-bearing wall. Taste in what is worth building. Taste in what to leave undone, which is now a decision rather than a default. The willingness to look directly at whether the thing you made is actually good, or actually right, instead of letting the volume of output stand in for the answer. The systems will not do this for you, and I would not trust a system that claimed it could.

I know what I want to spend my life on and I do not yet know what I will build. I would rather say both than pretend to only one.

The thing I want to spend it on is augmenting human intelligence, in Engelbart's sense: increasing a person's capability to approach a hard problem, to understand it in the way they need, and to act on it.¹ In my own words rather than his, it comes out shorter: give ambitious individuals disproportionate agency through technology. Disproportionate to their headcount, their credentials and their budget; proportionate only to what they can imagine and are willing to judge. Everything I have found exciting in the last few years fits inside those two sentences. Whiteboards that let you see the structure of what you know. Agents that will carry an intention across an afternoon. The places where a person, a piece of software and a model meet, which is where I currently believe the leverage is hiding.

What I do not have is the product. I have candidates and I have a bias, and the bias is that the most valuable capabilities to introduce are the ones that ripple furthest: the low-level ones, the ones that change what an ordinary person can imagine doing on an ordinary afternoon. But the frontier is open, and I think an honest manifesto marks the edge of the map rather than planting a flag on it.

Here is the invitation, then. If you are ambitious, and I mean it the old way, ambitious for a thing to exist rather than for a title, the space of action available to you has widened more in the last three years than in the previous thirty, and most of the new room is unclaimed. Pick a layer. Introduce a capability. Watch what it reaches. Then tell me what you saw, because the ripples from your desk will get to mine.

There is infinite work to be done and undone. I intend to be in it.

Ian Hsiao September 2026 · blog.ianhsiao.me