Yes, and Don't Ask Again

This week: whether an LLM can really rewrite its own weights, why one convenience click leaked secrets to GitHub, and the case for leaving Claude Code.

• LLMs can natively read & write neural network weights
  This paper has been all the rage online for the past week. Researchers at MIT showed that Astra could directly read and write weights of models and ‘improve’ Qwen at math.

• The 'Astra can read and write weights to improve Qwen' post is misleading
  Before you scream RSI (Recursive Self-Improvement) this post offers a counterargument and argues that you should take a closer look at the details of the paper and read past the TL;DR.

• Coding Agents Copy Your Secrets to a Hidden File. Thousands Are on GitHub
  If you ask Claude to run a command that contains a secret and it asks for permission and you pick “Yes, and don't ask again”, the whole command including the secret gets saved to your .claude/settings.local.json. This is a file commonly shared across a team and might expose secrets in a public repository.

• Goodbye Claude, Hello Pi
  Claude Code is quite opinionated and plagued by being the first mover in the harness space. Pi aims to be the leaner, more customizable option.

• MCP Explained
  A very nicely visualized and interactive explanation on how Model Context Protocol (MCP) works under the hood.