Capabilities expire. Principles don't.

I picked up Ethan Mollick's Co-Intelligence again in 2026, a book written for 2024, and ran into a sorting problem almost immediately.

Plenty of sentences still teach. Plenty of others now read like museum labels for demos that used to feel sharp. The capability snapshots have aged. The operating advice about how to work with machines has not, or at least not as fast.

That gap is this essay. It won’t be a book review. It’s more like a filter for how I read AI writing when the demos expire before the principles do.

It isn’t that surprising now

A lot of 2024 AI writing was built around astonishment. Look what the model can do if you give it a role. Look what happens if you prompt carefully. Look at this chat where you talk the work through step by step while you keep the destination in your own head.

Those moves mattered. They taught a generation of people how to get useful work out of chat interfaces. But they also had a short shelf life.

Prompting and prompt engineering already feel very 2024 to me. The role-play, the careful framing, the long back-and-forth where I stay the task manager and the model stays the clever intern: that interaction mode is not gone, but it is no longer the frontier of what feels surprising.

What feels normal now is closer to handing an agent a problem and letting it figure out the tasks. Agents already take over parts of my to-dos. The work still needs direction and judgment. The demo that used to stop me mid-scroll does not.

That is the perishable layer. Interaction snapshots. Capability screenshots. "You will not believe what happened when I typed X."

Principles travel further

The parts of Co-Intelligence that still feel solid are the quieter ones. Invite AI to the table. Give it guardrails. Build firsthand intuition by putting machines into real work instead of only watching demos from a distance.

Those are not model release notes. They are postures for exploring, directing, evaluating, and redesigning work around systems that keep getting more capable. The concrete recipe changes. The need for a posture does not.

I have been writing my own versions of that durable layer without always naming it that way. High agency, low authority is a commit-point principle: let agents prepare end to end, keep sending and shipping behind an explicit human yes. Bring agents to the systems is an architecture principle: take intelligence to the places where work already lives, instead of rebuilding life inside a new destination app. Compound judgment, not better artifacts is a practice principle: use AI so your next decision gets sharper, not only so today's draft looks cleaner.

Those essays are adjacent to this one, not the same claim. Compound judgment is about what should improve inside me while generation gets cheaper. This piece is about what should survive in my notes when the public demos age out.

Dan Shipper's After Automation lands in the same durable zone from another angle: cheap competence does not erase the human job, it relocates it. Someone still has to stay ahead of the frame. Geoffrey Litt's Understanding is the new bottleneck makes a related demand for agent-written work: understanding is not only verification after the fact, it is how you stay able to participate in the next creative move. Those claims will still matter after whichever model currently impresses me gets replaced.

The stake is how I read

When I save an AI essay, I am trying to notice whether I am bookmarking a trick or an operating idea. A trick is "this prompt pattern made Claude do X in March." An operating idea is "keep authority narrow at irreversible edges" or "prefer firsthand intuition over spectator takes."

Tricks can still be worth saving. They just need a different mental expiry date. If I treat every shiny interaction as evergreen knowledge, my notes rot in public. If I treat every principle as sacred wallpaper, I end up with slogans I cannot test.

A rough check I use while reading: would this still help if the model doubled in capability next quarter? If yes, it is closer to principle. If it only helps me recreate today's demo, it is closer to a dated capability note. Keep both. File them differently in your head, even if the bookmark tool looks the same.

Nate B. Jones's friction maxing is useful here as a reading posture more than as a production workflow. Most AI content removes friction: ask, get the answer, move on. The durable reading move is the opposite of passive consumption. Hunt for the claim that would still be interesting after the screenshots stop surprising you.

What I refuse to overcorrect

Retiring demos too aggressively is its own mistake. Capability observations can stay useful as before/after evidence. "This used to require careful prompting; now an agent decomposes the work" is a real historical signal. Erasing it to sound timeless would make the notes worse, not better.

"Principles" can also become vague motivational copy if I cannot point at a falsifiable edge. High agency without a commit point is just vibes ✨. Inviting AI to the table without guardrails is maximalism dressed up as curiosity. Mollick's invite is easy to misread as "use AI for everything forever." My own reading of that advice is narrower: get firsthand intuition. That is not the same as permanent maximal use.

And not every piece of AI writing needs to be sorted into this two-bucket system. Some posts are news. Some are product notes. Some are entertainment. The filter matters most when I am trying to learn something I expect to still use six months later.

What I want to keep

When I read about AI now, I try to separate the perishable capability layer from the durable principle layer.

Keep the demos as dated evidence. Keep the principles as portable posture: invite the machine in, give it boundaries, stay responsible for judgment, bring agents to existing systems, compound the next decision rather than only polishing the current artifact.

The models will keep moving. The screenshots will keep aging. The useful reading habit is noticing which sentences were built to survive that.