When AI starts building itself, "being able to do it" is no longer what's valuable

—Anthropic has laid "AI accelerating AI" out in the open. For someone carrying ten projects single-handedly, it's a rather uncomfortable mirror.

Today I watched Claude run an entire job end to end.

Core updates on two servers, reboots, verifying each service back to life one by one; midway I crashed an astrology service running under someone else's account by rebooting it myself, and Claude traced it all the way from a single 502 alert to the nginx config, to which user's pm2 that service was hidden in, to why it didn't come back after the restart—then fixed it, added a boot-time auto-start, and confirmed it won't go down again next time. Through the whole thing I was basically only responsible for saying "okay, do it."

A year ago, this would have been impossible.

Also today, I read Anthropic's piece "When AI builds itself." There's a number in it I'd normally just take as news—it says Claude now writes, of their ownover 80%of their production code, up from single digits in early 2025. But paired with what I'd just watched with my own eyes, that number suddenly landed on me and felt different.

This piece is my honest attempt to say what this article actually means for "someone like me." Not the big question of whether AI will destroy the world—that's not my subject. It's something smaller, more personal, and harder to dodge.

Don't let the phrase "building itself" scare you—the thing to really watch is another line

Anthropic's piece has a sensational headline: recursive self-improvement, AI designing and training its own next generation, with humans barely able to get a hand in. It even seriously calls for the world to preserve an option to "verifiably pause frontier development."

I'll set that part aside for now. For one, that's frontier-lab–level stuff, far removed from someone doing marketing in Taipei and writing a bit of code at night; for another, the article itself is half positioning—"our AI is so powerful it builds itself" is both a warning and a kind of persona. The numbers are real, but don't forget who's saying them, and why.

What really deserves your attention is the quiet curve in the middle of the article:

The length of task AI can reliably complete doubles every four months.

• Opus 3 (early 2024): a four-minute task
• Sonnet 3.7 (early 2025): ninety minutes
• Opus 4.6 (early 2026): twelve hours
• Extrapolated to 2027: week-long tasks

This isn't abstract. I live on this curve every day. The fact that I can single-handedly keep a customer-service system, divination, an election-prediction tool, a blog, and a conversational system grown from my own writing all running at once—the reason that exists isentirely because of this curve. Stretching "the length of time it can do work" from four minutes to twelve hours is, in effect, magnifying what one person can carry by a hundredfold.

Whether recursive self-improvement comes, and when, I don't know. But this curve has already taken effect on me. It isn't a prophecy; it's my daily working environment.

The mirror: when building gets cheap, what's stuck in me can no longer hide

And then comes the uncomfortable part.

A while back someone forwarded me an article about a person who used AI to spin up sixteen projects at once, mostly "write me a little script to do such-and-such," and an hour later had neither the scripts nor the original problems solved. Reading it, something clutched in me—because I've got ten on my hands too.

I used to comfort myself like this: I have a lot of projects because I work fast. AI lets one person do the work of ten, so I do ten things—reasonable.

But this Anthropic piece tears that excuse apart.

If "being able to do it" is rapidly losing value—code quality already on par with humans, AI reviewing its own work catching a third of production incidents—then "I can do ten at once" is no longer proof of any ability. Becausedoing was never the bottleneck。

The bottleneck was always the other two things:deciding which one is worth doing, and actually finishing the one you chose.

Not one of my ten projects is stuck on "I can't build it." They're stuck because I never thought through which one matters, stuck because I start one, find it interesting, then get pulled away by the next interesting thing. When building is cheap, the cost of starting a new project drops to nearly zero—so I keep starting. Every one is real, every one half-baked.

When building gets cheap, it won't solve your problem. It will magnify the problem you already had. If you were already someone easily distracted and bad at finishing, AI will let you get distracted and not finish with unprecedented efficiency.

That's what the mirror shows: I thought AI gave me the ability to "do more," when really it just threw a sharper light on the two holes in me—"deciding" and "finishing."

The part that won't automate away happens to be exactly what I've been working on

But this mirror also shows something that leaves me on steadier ground.

Anthropic itself lists three futures: capability growth stalling (which it thinks least likely), AI accelerating development but withhumans setting the direction, and full recursive self-improvement. That middle one—"AI does the building, humans decide what to build"—is probably the version we'll actually live in for the next few years.

In this version, what gets automated away is "doing." What won't automate away is three things:

deciding what to do, taste, and verifying whether it's right.

Deciding what to do is a value judgment, not a computation. Taste is knowing which one of a thousand workable options is the right one, and that takes a whole set of standards that are hard to articulate but are unmistakably yours. Verification is what today's astrology pitfall taught me—AI can take the whole thing to completion, but a judgment like "did we miss inventorying a service under someone else's account" still needs a person watching. It can do the work, but it doesn't know what it missed.

And these three things have one thing in common:none of them is a skill—each is a person.

That loops back to the thing I've been working on that others often think is a waste of time—saomin.tw/me, an "echo" grown from my own words that responds in my own way. I used to frame it as something very personal and existential: whether, after I'm gone, I can leave behind something that still responds.

But this Anthropic piece showed me another side of it. If the future really is "AI does the building, humans decide the direction," then "how a person judges, chooses, and what they believe"—their decision-making system—becomes the scarce thing that won't automate away. On the surface me.saomin externalizes a person's voice; at its core, what it externalizes is a way of judging.

I'm not saying this thing is worth a lot. I'm saying that what I've been doing on instinct happens to point in the direction value is concentrating toward. That puts me a little more at ease. Not because I'm smart, but because the subject I'm working on happens to be "the person," not "the doing."

That "pause" and alignment isn't my fight

The article spends a large part at the end on governance: frontier labs across multiple nations agreeing to pause under the same conditions, verifiable stopping mechanisms, misalignment compounding when "models build models," humans losing control.

These are all real problems. But honestly, this isn't my scale.

What I can do, what a frontier lab can do, and what a nation can do are entirely different things. Loading civilization-scale anxiety onto a solo builder only leads to paralysis—you start to feel everything you do is too small to matter, and then you do nothing.

The only part that touches me is a small reminder: I'm building something that "answers for me in my own voice." In an era where AI is gradually becoming everyone's gateway to understanding the world, a "thing that speaks for you" needs more care about whether what it says is really you, and whether it presents inference as fact. But that's a matter within my own valley, not the world's.

The world's fight, leave it to those who fight the world. My fight is in these ten projects of mine.

In closing

Having finished the article, I didn't come away more anxious—if anything, a little awakened.

It didn't tell me anything new. It just put what I feel every day plainly, in data:Building will keep getting cheaper; deciding what to build, finishing it, and verifying whether it's right—will stay expensive, and will stay a human matter.

So what I should do isn't to spin up an eleventh project to prove how efficient AI makes me. It's the reverse—cut the sprawl, concentrate my fire on the few things where "my judgment is the differentiator," and just maintain the rest, letting AI carry more of the building.

AI starting to build itself won't liberate me. It just took away my excuse. I used to be able to say "I don't have time, I can't finish it alone." Now it handles almost all of the doing, and what's left is only the question I keep dodging:

what do I actually want to do, and then, am I going to finish it.

This question—no model can answer it for me.

Further reading: - It wants to be AI's upstream; I just want to leave an echo - How to build an LLM that's truly your own - I built myself a knowledge base, then refused to let it speak for me - Original article:When AI builds itself — Anthropic Institute