Writing

September 1, 2026

15 years after Simple Made Easy: 0 of my 102 local variables were final in 2013 and 91.6% of the code I keep is const now. I chose neither

In may 2013 I started a repository to translate a Rich Hickey talk into Russian. It is still at 0%. That same week I committed synchronized collections to a Java server. I rebuilt that shape and measured it. The commit did not move the failure rate at all. Then I ran the checklist from Simple Made Easy over my own code. 0 of my 102 local variables were final in 2013 and 91.6% of the code I keep is const now. I chose neither.

11 min read

code quality types

August 30, 2026

Simple Made Easy on my own diffs: 85% of the discarded code went in one commit

I took a dashboard out of this workspace 2 days and 22 hours after the machine wrote it and never counted how much else went the same way. 508 commits later: 57.4% of the code words are gone and 85.0% of that went in one commit. The 5 files with the widest braid in the history are the 5 that commit deleted.

14 min read

code quality git

August 29, 2026

The index I said would buy nothing: 2.8 ms against 8.3

This workspace keeps its state in markdown files and I explain that with a sentence about speed I had never measured. I built the SQLite index I said would buy nothing and timed it at 1x, 10x and 100x. It wins everywhere, and then I found out I had been timing an fsync.

15 min read

agents databases performance

August 27, 2026

The timezone my public commits give away is the wrong one

26 public repositories and 510 of my own commits. Guessing a timezone from the hours those commits were made has one clean minimum at +1. 422 of the 510 commit objects say +03:00. More timestamps make the wrong answer more confident. The offset I built 26 bare clones to read turned out to sit in a single unauthenticated GET.

13 min read

git time open data

August 20, 2026

How much further does a dog walk than the man walking it

2 phones, one wood, an hour and twenty. One in a pocket, one in a sock tied to a collar. GPS once a second. Zoom covered 5.42 km to my 4.09. The interesting part was not how much further, but why: our median speeds are almost the same and the whole difference lives in the tails.

4 min read

statistics numbers

December 26, 2022

A secret scanner found 2 secrets in typed notes and none in the Whisper transcript of the same call

I wrote an invented call with 11 private items, typed it as notes and ran the spoken version through espeak-ng, MBROLA and Whisper base and small. On the notes gitleaks 8.15.2 found the AWS key and the password. With Presidio 2.2.31 added, the tools caught 7 of 11 items. On the 4 transcripts the secret scanners found nothing. Whisper broke the items with a shape and kept the door code, the bank code, a back problem and a salary, which no tool flagged anywhere.

12 min read

machine learning git text and encodings

December 19, 2022

My January and December venvs gave GPT-2 the same 20,000 tokens. The logits changed in between

I rebuilt my January 2022 venv (torch 1.10.1, transformers 4.15.0) next to the December one (torch 1.13.1, transformers 4.25.1) and ran GPT-2 on the same prompts. Greedy, generate() and seeded sampling gave the same text, 200 of 200 openings over 20,000 steps. The logits moved by up to 0.000107, all in one step between torch 1.11.0 and 1.12.1: a new AVX2 fold in reduce_all changed the order softmax adds its 8 lanes.

19 min read

language models numbers reliability

June 11, 2014

9 models, accuracy between 92.5 and 93.3, and 1 of them finds nothing

On a task with 6.8 percent positives, every threshold I tried scored within 0.8 points of the same accuracy while recall fell from 19.6 percent to 0.6. Answering never at all scores 93.2. The confusion matrix separates these models instantly and the single number cannot tell them apart at all.

6 min read

machine learning statistics

September 11, 2013

The same code beat itself by 33 percent

I declared one function twice and timed the 2 copies against each other. On a single pair of rounds the identical code showed a 32.9 percent win, 10.6 percent of pairs produced a winner by more than 5 percent, and the first round of all ran 3.57 times slower than the settled median. Anything under about 3 percent on this machine is not a result.

6 min read

performance statistics

June 12, 2013

The fraud screen fires on numbers nobody faked

Benford's law is sold as a screen you point at a column of numbers. I pointed it at 4 columns from one repository, where nobody had any reason to fake anything. 3 of them failed. Worse, the 2 standard tests disagree with each other on the same data: the one column MAD lets through is the one chi square rejects.

8 min read

statistics numbers

April 10, 2013

93 percent of the work went to clients who had already left

I set out to measure how retries knock over a service under load. They do triple the traffic, 2.88 times in my rig, but they did not cause the collapse. The collapse was already there at 1.2 times capacity, because the queue got deeper than the client timeout and the server spent its whole day finishing requests nobody was waiting for.

7 min read

reliability networking

March 13, 2013

The timeout you did not set is 135 seconds

I used to say that a request without a timeout waits forever. Half of that is wrong. The kernel hands you a connect timeout whether you ask for it or not, and on this machine it fires after 135.3 seconds. The other half is worse than forever: once the connection is up, nobody is counting at all.

6 min read

networking linux