Contact support

We reply by e-mail, usually within two days.

Google reCAPTCHA checks this submission against abuse; data is sent to Google. The script loads only once you open this form.

← All posts

The AI that usually says nothing

· Lukas Wojcik

The summary feature has been offered 2,061 counters. It has written 262 sentences. Every one of the remaining jobs ended with the same result, and it is not an error message even though it is filed as one: too little data.

1,796 — too little data 262 sentences 87.1 % 12.7 % 2,061 jobs, of which 3 are still queued

Eighty-seven per cent of the time, the answer is nothing. That is not a defect and it was not a surprise; it is the same finding that killed an earlier version of the broken-counter watchdog, arriving from a different direction.

Where the threshold sits

A counter qualifies for a sentence if, over the last thirty days, it recorded traffic on at least ten days and saw at least fifty visitors. Below that there is nothing to say that would not be noise. "Your busiest weekday was Tuesday" is a statement about a coin landing heads twice.

Most counters here are below the line. That is what a service for small sites looks like from the inside, and any feature that summarises behaviour has to decide what to do about it. There are three options and only one of them is honest:

  • Lower the threshold and write sentences about noise.
  • Write vague sentences that are true at any volume, which is a different way of saying nothing while appearing to speak.
  • Say nothing, visibly.

The cost of the ones that do get written

The 262 finished sentences were all produced by the same model, mistral-nemo, and took 12.8 seconds each on average — 5.4 seconds at the fastest, 54.3 at the slowest.

That range is the reason this does not happen while somebody waits for a page. The sentence is written in advance, stored, and reused for seven days before it is rewritten. A statistics page that took thirteen seconds to appear would be a worse page than one with no summary on it.

Why the refusals are stored

A rejected job could simply be deleted. Keeping it is what makes this article possible: the ratio between 1,796 and 262 is a description of the audience, available at any time, at the cost of one query.

It also prevents the same counter from being offered again the next day and refused again, which would turn a threshold into a queue that never empties. And it is the honest version of the feature's own status: this is not a summary engine that occasionally has nothing to say. It is a summary engine that usually has nothing to say, and the number is written down rather than left to an impression.

Advertisement