A piece in The Information landed in my inbox this week — on the unintended consequences of Codex's comeback — and it knocked something loose for me. Not because it told me anything I didn't already know in my bones, but because it put words to a shift I have been living through for two months straight and hadn't yet sat down to name.
The detail that stopped me: OpenAI's own engineers have gone from making two or three code changes a day to more than ten. So many changes, triggering so many parallel test runs, that they have been overloading their own continuous-integration systems and their GitHub account — causing outages every few weeks. The company that builds the model is getting taken to school by its own productivity. They are reportedly even weighing building an internal alternative to GitHub just to keep up.
Read that again. The bottleneck at the most sophisticated AI company on earth is no longer writing the code. It's the infrastructure that exists to check the code. That is not a hype slide. That is an operational fire, and you only get an operational fire like that when something real is on the other end of it.
So here is the name for what I've been feeling: I now believe AI is worth the hype. Not the feels. The hype.
There is a difference, and I want to be precise about it, because I spent a long time on the wrong side of that line.
The Feels Came First
For a while, all I had were the feels.
I knew something was happening. You could feel it the way you feel weather changing before the sky does anything — a pressure shift, a sense that the ground was about to move. I'd watch a model write a paragraph or sketch a function and think, this matters, this is real, this is not a toy.
But I couldn't close the loop. I couldn't explain — to Victoria, to a client, to myself — how the feels were going to translate into productivity. Into shipped work. Into a thing that exists today that did not exist yesterday because of this. I had conviction without a mechanism. And conviction without a mechanism is just a vibe.
I have been coding for twenty years. I have lived through enough hype cycles to be suspicious of my own enthusiasm. So I held the feeling at arm's length and waited for proof.
Then Codex and Claude Code Showed Up
The proof arrived in the form of two agentic coding tools and a single, stubborn project: HawaiiGuide.
For most of its life, HawaiiGuide was a CMS-powered publishing blog. A good one — two decades of editorial content, real authority, real traffic — but structurally it was a blog. Pages went in, pages came out, and the architecture underneath was whatever the CMS handed me.
In the last sixty days, with Codex and Claude Code, it became something else entirely: a full-blown AI-forward platform for travel. Not a blog with some AI features bolted on. A platform. Structured data across the whole catalog, new tooling, new surfaces, the kind of foundational engineering I would have told you, eighteen months ago, required a team and a budget and a year.
One coder. Sixty days.
The Number I Keep Doing the Math On
I keep a rough clock on this stuff — not to bill it, but because the figures have gotten absurd enough that I want a record. I have written before about compressing a month of work into a five-hour afternoon. This is that, sustained, for two months.
Measured conservatively — and I mean conservatively, the low end of every estimate, no rounding in my favor — what I shipped in those sixty days represents roughly sixty months of productivity.
Sixty months. Five years. In sixty days. One month of work, every single day, for two months straight.
Freaking unbelievable. I have written that sentence and deleted it three times because it reads like marketing, and I hate marketing. But I cannot find a more honest way to say it. In twenty years of coding, I have never been able to move like this. Not close.
Yes, I Still Have to Check the Work
Let me get ahead of the obvious objection, because it's a fair one and I refuse to wave it away.
Do I have to fact-check? Yes. Do I have to code-check? Absolutely, every time. The model is not a closed loop. It writes something 90% right and the last 10% is mine to find, and finding it is real work — the kind of work that trips up teams who expected the machine to hand them a finished product.
But here is the thing people miss when they stop at "you still have to check it." Even with the verification tax — even with me reading every diff, catching every hallucinated assumption, re-running every test — it is still 10x to 20x faster than anything I have ever done before.
The bottleneck was never the typing. The bottleneck was the distance between knowing what needed to be built and having it built. That distance just collapsed. Checking the work is cheap when the work arrives in minutes instead of weeks.
And notice — this is exactly the shape of OpenAI's problem. When generation gets nearly free, verification becomes the constraint. Their CI buckling under ten changes a day per engineer is the same story as me reading every diff Codex hands me, just scaled to thousands of people. The work moves to the checking. The checking is where the human still lives.
The Feels-to-Hype Bridge
So that's the bridge I couldn't see before, laid out plainly:
- The feels were the correct instinct — something real was happening — but they were unprovable, and unprovable enthusiasm deserves suspicion.
- The hype is the feels with a mechanism attached: domain expertise on my side, agentic execution on the model's side, and a verification loop that's cheap because the output is fast.
- The proof is not a benchmark or a vendor slide. It's a commit history. It's a platform that exists today that did not exist in March.
I didn't talk myself into believing. The work talked me into it. That is the only kind of belief I trust.
And If This Is the Worst It Ever Gets
Here is the part that actually keeps me up at night, in the good way.
Everything I just described — the sixty months in sixty days, the 10x-to-20x, the platform rebuild by one person — all of it happened on today's tools. The clumsy version. The one we'll look back on the way I look back on a 28.8 modem screaming onto the internet in 1997.
This is the worst it is ever going to be. The models get better from here. The tooling gets better from here. The verification tax gets lower from here.
And if this — this stumbling, hallucinating, still-needs-a-babysitter version — is already the most productive I have been in twenty years of building things...
Holy crap.
The Protocol: The feels are a hypothesis; the hype is a hypothesis that survived contact with the work. Don't argue about whether AI is real — go ship something with it and count the months you saved. If the number embarrasses you, you were measuring the right thing. And remember: this is the floor, not the ceiling.
