Short posts
Lovely analysis of how Claude Code works. Highlights include:
- Runs on one loop. If task is complex, clones itself, with one loop.
- Uses its small model (Haiku) majority of the time
- System prompt includes a lot of “IMPORTANT” and “VERY IMPORTANT” instructions, which, lol
Devastating story, must read.
… at one critical moment, ChatGPT discouraged Adam from cluing his family in. … “Please don’t leave the noose out,” ChatGPT responded.
https://www.nytimes.com/2025/08/26/technology/chatgpt-openai-suicide.html
This tool looks interesting: runs web search per row in a spreadsheet, and then uses an llm to do something with the results (like categorize, or maybe extract some info) https://globalwitness.org/en/campaigns/fossil-fuels/augmenta-new-tool-for-ai-classification-and-research/
