Outsider And Insider, At The Same Time

By Andrei Roman

A while ago I stopped posting here. The reason is mundane: I joined a large company, and the first weeks of any serious job leave little surplus for writing. But the silence was also deliberate. I wanted to feel the new ground before I said anything about it, because what I found there is the clearest confirmation I have yet seen of something I have been arguing on this blog for months - that the center of gravity in software has moved outside the code.

I am not going to write about where I work or what I do there. That stays inside. What I want to put down is a single observation from the experience, because it turned out to be a small, exact instance of the larger shift, and it sharpened how I think about my own relevance in a world that is rearranging itself around AI.

The two feelings

In my first days I felt like an outsider and an instant insider at the same time, and the two feelings never resolved into one.

The outsider part was obvious and humbling. A large, living codebase is like a city you never visited. You don't know which streets are safe, which buildings are important besides the tourist traps, which conventions are law and which are folklore that everyone forgot to delete. No amount of seniority imports that knowledge. You arrive not knowing the map.

The insider part was stranger. People adopted me quickly. Some of that was social proof - shared networks, the quiet validation of "you worked with so-and-so," the way a known name in common lowers everyone's guard. I want to be honest that this was real and that I cannot fully separate it from the other thing, which showed up in a specific moment that repeated: the moment someone described a problem they were wrestling with, and I understood it - not the surface of it, the shape of it - and they could tell that I did. That recognition happened fast, and it did not depend on me knowing their map at all.

For a while I held those two feelings as a contradiction. They are not a contradiction. They are the whole point.

Why both are true

The thing that made me an outsider - not knowing the repository, the internal history, the accumulated conventions - is exactly the kind of knowledge that is becoming cheaper to acquire. This is the descriptive layer of a system: where things live, what calls what, which flag flips which path. It used to take months of immersion to build, and that slow build was a large part of what tenure was. Good tooling and good agents now help you absorb a lot of it in a fraction of the time. Not all of it - some of the map lives in incidents, scar tissue, and implicit contracts that no tool fully compresses, and pretending otherwise is its own kind of hubris. But enough of it compresses that possessing the map is no longer the scarce, identity-defining achievement it once was. You can get oriented faster now.

The thing that made me an instant insider - being recognized as someone who genuinely understands a hard problem - is the layer that did not get cheaper. It does not transfer through tenure. You cannot acquire it by reading the repo. It is judgment: the ability to see the governing constraint, to know which trade-off actually matters, to recognize when a stated problem is really a different problem underneath. That recognition between two people who have both done the hard thing is not social proof. It is the one signal that still cannot be faked, because faking it fails on contact with the first real question.

So the outsider feeling and the insider feeling are not in tension. They are a measurement of where value now sits. I was behind on the thing that is becoming less scarce, and immediately useful on the thing that still is.

The stake in the ground

Here is what I will commit to in writing, because I would rather be wrong in public than vague in private.

The first generation of AI tools optimized for generation - making code appear. AI has now made code generation cheap enough that it is no longer the center of gravity, and in doing so it commoditized the very skill most engineers spent careers compounding. The thing it did not touch, and the thing it actually made more valuable, is everything that surrounds generation: knowing what to build, where the boundaries go, what "correct" means in a domain, and whether the output can be trusted at all. The scarce work is no longer producing the artifact. It is the judgment that aims the production and the verification that grades it.

This is why I have spent my own time on orientation, on knowledge representation that refuses to give confident wrong answers, on deterministic substrates that constrain what agents are allowed to do. Not because generation is uninteresting, but because generation is no longer where the leverage is. The leverage is in control.

I am not the only one converging on this, and the most useful thing I can do here is point at three pieces that say parts of it better than I could, from very different vantage points:

• Satya Nadella, A frontier without an ecosystem is not stable. His distinction between human capital and token capital is the cleanest framing I have read of the actual asset. His sharpest line is that you can offload a task, or even a job, but you can never offload your learning - the firm compounds only if it owns the loop that turns judgment into reusable capability, and can swap the underlying model without losing the accumulated expertise. Read it for the sovereignty test: can you replace the engine without losing what you learned from running it?

• CTOs Agree: Cognitive Debt Is the New Technical Debt (ShiftMag). A dinner of engineering leaders admitting nobody has this figured out yet. The line that stuck: code is cheap, but launching it isn't - writing something is not the same as shipping or maintaining it. And the quieter, harder one: if you turn your engineers into reviewers protecting the company from code they didn't write while someone else takes the credit, check on them. The bottleneck moved from writing to judging, and most orgs have not updated anything to match.

• Steve Yegge, The Flat Curve Society. His two horizons are the most useful tool in the set. The demand horizon: you cannot tell two models apart unless your problem is hard enough to stretch them - so keep a genuinely hard problem in your pocket. The discernment horizon: you cannot supervise an answer you cannot grade - so the ceiling on how much intelligence is useful to you is set by how much of it you can actually verify. Both push the same way: the scarce skill is not prompting, it is the judgment that poses hard problems and the understanding that can check the work.

What this means for being relevant

If the descriptive layer is cheap and the judgment layer is scarce, then the move - for me and for anyone thinking about the next decade - is not to defend old knowledge by hoarding it. It is to let everything you already know become reproducible without you, so you are free to spend your attention on what is not yet known. Replaceable in every solved problem; chosen, again and again, for the unsolved ones.

That is a strange kind of security. It is not the security of being the only person who understands the obscure system. That security is a trap; it ties you to a decaying monopoly and a job you cannot leave. The durable position is the opposite: be the person who keeps converting fresh contact with reality into reusable understanding, and then moves on before that understanding becomes a commodity.

I felt like an outsider because the map was new. I felt like an insider because the problems were familiar at the level that matters. The next decade, I think, belongs to people who are comfortable being both - fast to orient because the tools make orientation cheaper, and valuable because they bring the one thing the tools still cannot: judgment that earns recognition on contact.

The work now is not to make more code appear. The work is to make accelerated change understandable, verifiable, and safe. That is the layer I am building toward, and there is a great deal of it left to build.