I keep being told the voice guide is unnecessary: show the model what you have already published and let it work the voice out. I gave claude-opus-5 8 of my 11 English essays and asked for the guide. Over 5 runs it wrote 92 entries and 2 of them are mine. All 16 checks score 0 across 12587 words of published prose, and for 6 of them even the baselines score 0, so there was nothing there to find.
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.
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.
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.
I wanted the whole sky seen from Proxima Centauri, built from a catalog. Gaia DR3 does not have the 7 brightest stars within 10 parsecs, so the sky is Hipparcos 2007 with Gaia's distance to Proxima. With catalog parallaxes Alpha Centauri B lands 70 degrees from A. The Sun is a 0.38 star in Cassiopeia, the median naked eye star moves 0.50 degrees and Sirius 27, into Orion.
I crawled the first 1000 domains of the Tranco list to see how many sites serve source maps with their original code inside sourcesContent. 87 served a map with source text. After sorting them by hand, 55 serve their own code, among them GitHub, Atlassian and Trello, while 19 only leak through an embedded vendor script.
I counted how many sites in the HTTP Archive crawl load a third party script whose own Wappalyzer description claims AI or machine learning. In November 2022 it is between 0.22% and 0.72% of mobile origins and almost half of it is 2 vendors, Ezoic and Dynatrace. The growth over the year came from new catalog entries: the technologies already detected in January went from 0.58% to 0.54%.
The simplest live Whisper transcriber cuts the stream into 5 second pieces and gives each to transcribe(). On 19 minutes of LibriSpeech read speech, Whisper base at the 9 December commit gave 5.79% wrong words on the whole file and 14.89% on the pieces, at 4 times the compute. Longer pieces help at the cost of delay and the previous text as a prompt helped little. Cutting in the quietest spot between 4 and 6 seconds gave 10.01%. Even 30 second pieces stay worse than the file, because transcribe() can no longer go back to a cut word.
I took the 4362 npm packages with at least 1 million weekly downloads at the end of 2022 and counted who can publish them, who asks for money and how much money is public. 47% of the downloads come from packages with 1 maintainer, 19% of packages ask for money and only 13 READMEs found by my word list say anything about the people. Open Collective shows 1.65 million dollars for 2022, 16 per million downloads. 53% of it went to 5 collectives.
In January the author of colors and faker broke both packages. I ran the code that is left, including a 1.4.2 tarball from the Wayback Machine, against the 9 top npm packages that depended on colors. 7 of them would hang with 1.4.2 and most reacted within 5 days. But in December both names still belong to the author and most of the downloads still go there.
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.
I sampled 2 hours of every day of GH Archive in 2021 and 2022 and counted new GitHub repositories whose name or description is about AI. Among repositories with a description it is 1.36% in 2021 and 1.44% in 2022, so no jump. What changed is the stars: generative repositories went from 0.10% of all stars in 2021 to 3.07% in December 2022.