When You Can't Go Back

Last week I was halfway through a feature when Claude started giving shallow, off-target responses. Enough degradation that I had to stop leaning on it and start thinking through things I hadn’t thought about in months. An Android API I used to know cold, I had to look it up. A state management pattern I’ve implemented dozens of times, I stared at the screen trying to remember the shape of it.

The tool came back an hour later. But the feeling didn’t leave. I had just watched myself hesitate at my own job, and the reason was that someone else’s infrastructure had a bad afternoon.

I can do my work without AI. I know this because I did it for over a decade. But “can” is doing a lot of heavy lifting in that sentence. The way I work now, the pace I maintain, the scope of what I take on, all of it assumes the tool is there. Remove it and I’m not just slower. I’m a different engineer, working inside a role that was shaped around a capability I don’t fully own.

The Subsidy Era Is Ending and Nobody’s Ready

My Claude Max subscription costs $200 a month. At API pricing, the tokens I consume are worth roughly ten times that. Anthropic eats the difference.

Cursor was even more aggressive. I was on their $60-a-month business plan and routinely burning thousands of dollars in actual compute. A single prompt once consumed over eleven million tokens. Then Cursor enforced limits on frontier models, and the same number of requests suddenly produced twenty-eight times less output. My monthly quota now evaporates in about 25 sessions.

Remember when Uber was cheap? When Airbnb undercut every hotel? When DoorDash delivered for less than the cost of the food? They burned investor money buying market share, subsidizing your rides, stays, and deliveries until competition thinned out. Then prices found their real level. AI coding tools have been running the same playbook.

And the moats are thin. These companies have little defensibility besides the quality of the models themselves, and there are only a handful of providers. The dynamic is familiar: enormous market power concentrated in a few hands, and everyone downstream is dependent on decisions they can’t influence.

I felt this concretely when Claude Code shipped an update that removed the question-asking feature I relied on. I liked being asked clarifying questions before the agent dove into implementation. It made the collaboration feel deliberate. Then it was gone, replaced by a product decision I had no say in. Later they shipped Opus Advisor, which consumes tokens at a completely different rate. These aren’t bugs. They’re business choices made by someone else that reshape how I work, every day, without asking.

That is what dependency looks like in practice. Not a dramatic failure. Just the slow realization that the terms of your productivity are set by someone else’s roadmap.

The Return Path Closed Before Anyone Noticed

This pattern is not new. What’s new is the speed.

In the Middle Ages, scribes were essential. They were the people who could read and write, employed by monasteries and kings who often couldn’t. Then Gutenberg built the printing press, and within a generation the economic infrastructure that supported scribes dissolved. Literacy spread. The monasteries shifted. You couldn’t go back to being a scribe even if you wanted to, because the world that needed scribes no longer existed.

What happened to them is more interesting than “they disappeared.” Many became editors, typesetters, proofreaders. Their understanding of what made a good book didn’t evaporate. The ones who identified with the act of transcription were replaced. The ones who identified with the craft of making good books found more power than they’d ever had, because now they could reach readers at a scale that was previously unimaginable.

Nobody reads handwritten books anymore. Nobody reads typewritten ones either. We optimized, and many more people are able to share their books with many more readers. That is, on balance, a good thing. But the return path for the scribes closed permanently. The skills didn’t just become rare. They became invisible.

Telephone operators tell a sharper version of the same story. At their peak in the 1940s and 1950s, connecting calls was one of the most common jobs for women in America. Automatic switching arrived, and the role didn’t evolve into something adjacent. It ended. The skill didn’t transfer. The infrastructure that supported it vanished within years.

Typesetters had it even faster. Hand typesetting lasted over five hundred years, from Gutenberg through the 1960s. Machine typesetting replaced it and lasted about ninety years. Phototypesetting followed and lasted roughly forty. Then desktop publishing arrived in 1985 with PageMaker and the Apple LaserWriter, and within a decade the physical craft of arranging metal letters into pages was irrelevant. The knowledge of kerning and layout survived in new forms. The hands that could set type at speed had nowhere to go.

Each transition compressed faster than the one before. Scribes had a generation. Typesetters had a decade. Telephone operators had even less. I don’t know exactly what that suggests about the current transition, but I notice that the pattern only accelerates.

The Woodworker and the Programmer

I watch a YouTube channel called Blacktail Studio. It’s run by Cam, a Portland-based woodworker who is the sole owner and sole employee of his shop. He went from a novice who ruined his first epoxy table to selling pieces for $10,000 to $30,000, with nearly three million subscribers watching him build. One person, power tools, vision.

His work is still handmade. But he uses CNC routers, precision jigs, specialized epoxy systems, tools that would have been unimaginable to a furniture maker two generations ago. If you took those tools away, Cam could still build a table. It would take dramatically longer, cost dramatically more, and he’d need a team to match what he does alone. His lifestyle, his business, his output all assume the tools are there.

The parallel to programming almost doesn’t need spelling out. But the important thing is what happens on the other side of the equation: more people get better products. Cam’s tables are better than what a team with hand tools would have produced a generation ago. The consumers benefit. Some of the people who would have been on that team are doing other work now. The net outcome is arguably better. The individual displacement is real.

Cam’s identity isn’t “tool operator.” It’s the vision, the design sense, the understanding of materials and grain and what makes a piece worth the price. The tools amplify that understanding. They don’t replace it. If his identity were “the guy who operates the planer,” he’d be in trouble.

The same split applies to engineering, and it has applied before. Coding went through this compression many times: punch cards to assembly, assembly to FORTRAN, FORTRAN to C, C to higher-level languages, then frameworks, then libraries, each layer abstracting the one below. At every step, engineers stopped needing to know the previous layer in depth. At every step, someone argued that this was the last abstraction we’d need, that this time we’d reached the floor of essential knowledge. What makes us think the current JavaScript frameworks are the end state? They are not. They never were.

Choosing It Anyway

I am choosing the dependency. Eyes open.

Not because I’m naive about the risks I just spent a thousand words describing. Because the alternative is falling behind, and because the tools genuinely expand what I can do. The content workbench I built for our learning team. The Flutter apps I ship as side projects. The architecture diagrams I produce in thirty minutes that used to take days. The work is better. Not just faster. Better.

It is wrong to think we are at the end of history. Every generation before us thought it. The scribes didn’t see Gutenberg coming. The typesetters didn’t see PageMaker coming. The telephone operators didn’t see automatic switching coming. We will not predict what comes after the current AI tools, and we will look equally surprised when it arrives.

This doesn’t mean mindlessly adopting whatever appears. Some of these tools will matter and some won’t, and waiting often reveals which. But the underlying shift is not going away.

More is possible now than at any previous point. The people who stay mindful about the dependency, who understand what they’re trading and why, who keep choosing deliberately rather than being carried along, are not helpless in this transition. They’re navigating it.

But navigating isn’t the same as controlling. I choose the tools. I don’t choose the terms. The subscription price, the token limits, the features that appear and disappear, the roadmap that reshapes my workflow without consulting me. Those belong to someone else.

What I Can’t Name Yet

The dependency is going to deepen. I know this because every previous version of this transition only went in one direction. The scribes didn’t go back to copying manuscripts. The typesetters didn’t go back to metal letters. The telephone operators didn’t go back to the switchboard. And I am not going back to writing every line of code by hand, because the version of my job that required that is already receding.

There are things I used to know about Android development, patterns and APIs I spent years mastering, that I can’t recall with confidence anymore. They’re fading the way a language fades when you stop speaking it. I notice the gaps sometimes, in meetings or when reviewing someone else’s code, and the feeling is strange. Not panic. More like watching a tide go out and realizing you can’t remember what the beach looked like before.

I’m choosing to let those details go because what I’m gaining is real. The scope of what I can build, the speed at which I can think through problems, the range of projects I can take on. These are genuine expansions, not just efficiencies. I believe that.

What I can’t name yet is the cost. Not the subscription cost. The other one. The thing you lose when your competence depends on a tool you don’t control, when your expertise lives partly in someone else’s infrastructure, when the knowledge that once defined you quietly becomes optional. I know something is being traded. I just can’t see all of what it is.

The dependency will keep deepening. The tools will keep improving. The return path will keep closing. And I’ll keep choosing to walk forward, because that’s the only direction the path goes.

I just want to be honest about the fact that I’m not entirely sure where it leads.