After the Frontier

Friction and the Populated World

I detest surveys. I go out of my way to avoid them, and the last one I completed had an Amazon gift card attached to it. So when a popup from Claude last night asked if I would be interested in sharing my thoughts with Anthropic, you'd figure I'd be hitting command-w hard enough to void my laptop's warranty.

Instead, I was intrigued.

The interviewer was itself an AI, naturally. It asked what I had most recently used AI for. Anthony Beevor's Rasputin, as it happened, although that book is less about Rasputin than about the Romanov dynasty entering the terminal phase of a very long systems failure. From there we had wandered through the Russo-Japanese War, the spectacularly botched murder of Rasputin, the still more ghastly execution of the Romanovs, and eventually Tsesarevich Alexei's dog Joy, who escaped the family's fate, lived out his life in England, and ended up buried beneath what became a car park.

This is more or less how I use these things.

I was then asked for my most memorable moment with AI.

I couldn't think of one.

Which surprised me more than any answer I actually gave.

There had certainly been landmarks: early chatbot novelty, models suddenly able to digest enormous documents, the dawning need for a sycophancy sniff test, coding agents crossing some threshold from parlor trick to useful coworker. But no single moment stood out anymore.

I compared AI to a refrigerator.

When was the last time refrigeration made my life better? This morning, probably. Cold drink. Almond milk. Blueberries. Breakfast becomes possible because some appliance hums away in the background, and unless it starts leaking water onto the floor I don't regard that as an event.

Somewhere along the way AI seems to have become like that.

Not invisible, exactly. Ambient.

This week I had an agentic IDE promote an ECS service into staging. It did it quickly enough that I almost immediately discovered the thing I had forgotten to put in the specification: check whether related services had picked up new environment variables or secrets. Redeploy. Fix the spec tomorrow before production.

Meanwhile I am still waiting for production credentials from the DBAs.

The machine is fast. The humans around it become visible.

Not necessarily because the humans are doing anything wrong. Production systems contain permissions, ownership, approvals, secrets, upstream dependencies, people who know which thing cannot be changed on Thursday afternoon because somebody once learned why. An agent can gallop ahead with the Terraform, but eventually it reaches the pass and waits for the rest of the posse.

This may be why I've started having problems with the word frontier.

A frontier is wonderfully photogenic. Mountain ridge. Horse. Duster coat. Vast unclaimed distance.

"The Marlboro Man died of cancer, y'know," whispered my inner smart-ass.

The romance works better if you aren't mopping up after the gunfight.

Frontier AI is a perfectly serviceable technical term for the models at the moving edge of capability, but the metaphor keeps dragging other things in behind it: wilderness, lawlessness, rugged individualism, the assumption that the horizon exists in order to be pursued.

What happens behind the rider?

The dirt road gets paved. The lone sheriff and the vigilante mob eventually give way to a courthouse. Somebody installs plumbing. Somebody maintains it. Somebody writes down which department owns the culvert.

The frontier seems very interested in movement and curiously uninterested in persistence.

Then a friend sent me a delightlfully overcaffeinated essay arguing that Apple is secretly winning AI because enormous open models can increasingly be run on "consumer-price" Macs rather than rented from a hyperscale datacenter by the token. Strip away the "Nvidia is dead" theater and the underlying idea is fascinating.

What if the destination isn't the frontier at all?

What if it is the homestead?

You plunk down Darth Vader money for a Mac stuffed with unified memory, put a model on it, unplug the network cable if you like, and discover that last year's miracle has become a desk ornament.

Of course the homestead will be obsolete in two years.

But maybe that stops mattering at some point. If a quantized model is good enough to follow a detailed spec, inspect a repository, produce Terraform, write a Python script, run tests, and hand back a diff worth reviewing, how much better does the next model need to be before the old one stops being useful?

That seems less like racing toward the horizon and more like deciding when the tractor in the barn already does the job.

Then Sherry Turkle's Artificial Intimacy arrived this week.

Timing.

I am admittedly allergic to the word authenticity, and "artificial intimacy" immediately made me think of bartenders and les femmes de la nuit. Humans have been purchasing bounded forms of listening, attention, companionship, flirtation, confession, and care for a very long time.

Turkle's more interesting question isn't whether machine companionship is fake. It is what happens when companionship becomes extraordinarily smooth.

The machine is there. It listens. It does not need to get home. It does not have a headache. It does not suddenly decide you are wrong and would rather discuss something else.

Does that change what we expect from everybody who does?

A couple of weeks ago, Nancy and I flew to San Francisco to see Carla dal Forno. It was a stupidly expensive undertaking involving airfare, a hotel during the Dreamforce conference, food, rides across town, and all the usual costs of placing two physical bodies someplace they were not previously located.

It was also magical.

We wandered into a tapas place because two guys were playing guitar outside. Sat at a sidewalk table. I plunked around on a jaw harp for a minute while Nancy filmed me and tried not to laugh. After the concert, the server recognized us walking past and asked how the show had been.

Then our Lyft arrived, a Tesla with vibratingly loud hip hop playing inside, and late-night San Francisco briefly acquired the uncanny sheen of a cyberpunk role-playing game.

None of this was efficient, and none of it needed to be.

That doesn't mean friction is secretly virtuous. I would still like those production database logins. A bureaucratic delay doesn't become meaningful merely because another human caused it. Discord servers reduce enormous amounts of friction and is still thoroughly populated by people who can disagree, vanish for three days, misunderstand the joke, remember the joke six months later, or derail the whole discussion into something worse. Or better.

Maybe that's the distinction I keep circling: Not human versus machine. Not physical versus digital. Not friction versus convenience.

Perhaps the question is what kind of friction we are removing, and what disappears with it.

Some friction is waste. Some is scarcity. Some is bureaucracy. Some is simply the resistance created by another person having a vote. And some experiences may depend on a world that remains populated by things we did not specify in advance.

The Anthropic interview began by asking me for memorable AI experiences. Perhaps what I learned from it was that the memorable part is becoming harder to isolate.

While the frontier keeps moving somewhere out there, I find I'm increasingly interested in what is quietly taking up residence behind it.