Tagged text and encodings
- codecard refuses to render a snippet that names a codebase. It missed all 569 Kotlin package lines I gave it
- For every Belarusian page in Common Crawl there are 400 in Russian
- My live Whisper transcriber made 2.6 times more mistakes than Whisper on the whole file
- A secret scanner found 2 secrets in typed notes and none in the Whisper transcript of the same call
- A model behind my bank statement regexes turned 300 loud failures into 249 silent ones
- The tenant came back as the lessor after a round trip through Russian
- One in 7 "Belarusian" pages in mC4 is in Russian
- GPT-3 reads Belarusian at 6.5 times the price of English
- I asked my notes 15 questions they cannot answer and got 13 answers
- OpenAI's biggest embedding model costs 75 times more than the smallest and loses ArguAna to a 22M one
- A commit built by hand: 187 bytes
- Reading UTF-8 as 1251: "Привет"