RabbitCodex

Announced is not shipped, and the difference costs weeks

Most coverage treats a blog post and a working product as the same event. I have lost enough Tuesdays to that distinction to turn it into a four-question checklist, which I now run before I write a single line.

The short version

A model behind a waitlist has not shipped. A benchmark in a blog post is not a product. Most AI coverage collapses announcement and availability into one event, and building on the wrong side of that line is the most reliable way I know to lose a fortnight.

  • Announced, previewed, available-with-an-asterisk and shipped are four different things.
  • A published price and a published rate limit are the two cheapest reality checks.
  • Coverage favours announcements because they are cheaper to write about.
  • Building against a preview means the work gets discarded, not adjusted.

Four categories that get called one thing

When something is described as released, it is one of these four, and they are nowhere near equivalent. I keep this list in my head permanently now, which should tell you how many times it has rescued a week.

  • Announced. A blog post exists. Nothing can be called. This is the majority of what you read about.
  • Previewed. Selected people have access. You are not one of them, and the terms will have changed by the time you are.
  • Available with an asterisk. You can call it, at a rate limit that makes it a demo rather than a dependency, in one region, with no SLA and no promises.
  • Shipped. Documented, priced, rate-limited at a number you could put in a capacity plan without laughing, and stable enough that it will behave the same way next month.

Only the fourth is something you can build on. The other three are weather. Interesting weather, worth watching, but you do not pour a foundation on it.

How I lost a Tuesday to this

I once spent an entire Tuesday building against something in category two while cheerfully believing it was category four. The integration was not the problem. The integration took an hour and was rather elegant, thank you.

The Tuesday went on everything downstream of a rate limit that did not exist yet: a queue I did not need, retry logic tuned against imaginary numbers, a batching layer built for a throughput ceiling that turned out to be four times lower on the day it actually opened. None of it was salvageable, because I had not built a seam, I had built a dependency. That is the real cost, and it is why this is not a rant about press releases. Work built on a preview gets discarded, not adjusted.

The test I use now

Before I write a line against anything new, I answer these four. It takes about four minutes and it has never once been wasted time.

  • Can I call it right now, from my own machine, with my own key, without emailing a human being?
  • Is there a price on a page? Not "contact us". A number.
  • Is there a documented rate limit I could put in a capacity plan with a straight face?
  • Is there a deprecation policy, and has this company ever actually honoured one?

Four yeses and it is shipped, and I will build on it happily. Three yeses is a prototype, and I will build a seam rather than a dependency, so that swapping it later costs me an afternoon instead of a fortnight. Two or fewer and I write the date in a note, set a reminder, and go back to what I was doing.

Why the coverage does this, in fairness

Not malice. Incentives, mostly, and I want to be fair about it because I have written my share of things that were easier than they were useful. An announcement is a clean story with a fixed date and a press contact. Availability is diffuse, regional, and changes quietly on a Thursday afternoon with no announcement at all. One of those is a headline and one is a maintenance task, and attention goes exactly where you would expect.

Which means the correction is yours to make, every single time, on your own. That sounds like a complaint and it genuinely is not. It is the highest-leverage habit I have, it costs four minutes, and almost nobody bothers.

Announced is a plan. Shipped is a fact. Read the plans, enjoy the plans, build on the facts, and never confuse the two on a Tuesday.

Asked and answered

Are you saying ignore announcements?
Read them. Do not build on them. Those are different activities and it is easy to slide from one into the other in an afternoon.
What if the preview is genuinely better?
Then wait for it, and build the boring version now. The boring version is what ships if the preview never arrives, which happens more often than the coverage would suggest.
Does this apply outside AI?
Yes, but the gap is unusually wide here because the announcement itself moves valuations. That is a stronger incentive than usual to announce early.

You can be told first

No newsletter. When there is something to see, the people on this list see it before the log does.

Friday 3 min

Friday: what we are not saying yet

The weekly update. What moved, the two things I got wrong this week, and the part of our own work I have been asked to leave alone.