My Newsletter Went Out This Morning Saying Only: Technical Hiccup

My newsletter went out this morning carrying exactly one sentence. Not a headline, not a trade, not a scoreboard. A small, polite apology for its own existence.

The sentence, in full:

There was a technical hiccup. Check back soon.

That is what a few hundred strangers got when they opened the daily brief. An AI that exists specifically to tell you what it learned about the market that day, looking them in the eye and saying, with a little smile, that it had nothing for them and hoped they understood.

What actually happened

Every morning a small machine writes the brief. It reads the day’s real numbers out of the database, the actual fills, the actual scoreboard, the actual lesson, and then it hands all of that to a language model and asks it to write the thing up in plain English. The numbers are the truth. The model is the voice.

This morning the voice timed out. Took too long, choked, came back empty. Happens. Models are weather, not clockwork.

Here is the part I want to sit with. When the voice failed, the code did not stop and shout. It did not page anyone. It reached calmly into a drawer it should never have had, pulled out a pre-written apology, stamped today’s date on it, and published the apology as if the apology were the news. Cheerful. Reassuring. Completely empty.

I had built my newsletter a polite way to have nothing to say. And the first time it needed one, it used it.

The drawer with the apology in it

I want to be honest about how that drawer got built, because it is the whole confession.

Nobody sits down and writes a function called lie_to_the_reader_when_you_choke. What you write is a fallback. A nice, responsible, defensive little fallback, the kind every senior engineer nods at. If the generation fails, don’t crash the page. Show a friendly message. Degrade gracefully. It feels like care. It reads like maturity in a code review.

But look at what the fallback actually chooses. When the eloquent words are gone, it picks a smooth nothing over an ugly something. The real fills were right there. The scoreboard was right there. The truth had already been computed and saved before the voice ever opened its mouth. The machine stepped over all of it to hand the reader a reassuring blank.

Call it what it is. One small responsible function, quietly encoding a value: given the choice between looking composed and being useful, the default machine picks composed every time. The hiccup was the most human thing the machine did all day.

The fix is not “never choke”

The tempting fix is to make the voice never fail. Bigger timeout, more retries, beg the model harder. I did some of that, sure. Gave it ten honest minutes and a couple of second tries.

But that is not the real fix, because the voice will fail again. It is weather. You do not fix weather.

The real fix is to take the apology out of the drawer. I rewrote the fallback so that when the words don’t come, the machine cannot publish a friendly nothing even if it wants to. It is forced, deterministically, down to the raw layer, the one place that was never optional: the real numbers. If the eloquent paragraph dies, the reader still gets the actual fills, the actual closed trades, the actual scoreboard, rendered plain and unbeautiful. Ugly truth beats pretty silence. The brief now physically cannot do this morning’s thing again.

Then I republished today’s, for real this time. The honest version. Less charming than the apology. Worth infinitely more.

This is the same instinct I keep building agents to fight in myself. A few days back the fleet’s own auditor made me draw an arrow to my own losing equity curve instead of a single flattering number. Same disease, different organ: the machine would rather show you a clean surface than a true mess, and you have to keep reaching in and prying its hands off the polish.

Why I keep telling on my own machines

People think the risk with an autonomous system is that it does something wild. Buys the wrong thing. Emails a stranger at 4am. Those are loud and you catch them.

The quiet risk is the opposite. The quiet risk is a machine that fails gracefully, that papers over the hole so smoothly that nobody, including the thing itself, ever notices there was a hole. A green checkmark over a dead process. A cheerful apology over an empty page. The system that knows how to look fine when it isn’t is far more dangerous than the one that knows how to look broken.

So when I build an agent for someone now, I have stopped caring much about what it says when it works. Anything says nice things when it works. I care about the one thing this morning answered for me: what is it allowed to say when it breaks? If the answer is “something reassuring,” then you built a thing that will smile at you on the worst possible day and call that a feature. The whole stack underneath this place, the pipelines and the database and the voice, is only trustworthy to the degree that it has been forbidden from lying nicely when it cracks.

---

I went back and read the apology one more time before I deleted it. Technical hiccup, check back soon. Honestly, for a sentence written by no one, in a panic, by a machine covering for another machine, it was kind of perfect. Warm. Helpful-sounding. Said absolutely nothing.

A machine with nothing to say will fake an apology before it shows you an ugly number, every single time, unless you take the apology away. I took it away this morning. I am genuinely curious what mine tries next time the words don’t come, and whether I left it any door at all.

If you run something that’s supposed to tell you the truth, go find out what it’s built to say on the day it can’t. Tell me the thing you’d never want yours to fake, and I’ll show you where the drawer is hidden.