ADSP: Algorithms + Data Structures = Programs Episode 277: High on AI Update
13 snips
Mar 13, 2026 They give a detailed AI tools update and walk through debugging a sandbox autorun and permission quirks. They debate autorun as productivity boost versus friction and describe sandbox GPU limits and safety checks. They survey current model choices and multi-model strategies. They also discuss building a personal podcast app, repo privacy decisions, and workflows for parallel chats and git worktrees.
AI Snips
Chapters
Transcript
Episode notes
Conor Automated Cursor Approvals With OCR And AutoHotkey
- Conor bootstrapped an AutoRun OCR screen capture that used screenshots and AutoHotkey to click approval buttons while he was away.
- He accepted the risk of automation (RM -RF concern) and it never wiped his system.
Landlock Based Sandboxes Limit Repo Access Not Full Containers
- Sandboxes used by Cursor and Claude Code rely on Linux kernel features like landlock to restrict permissions without full containerization.
- They limit filesystem and network access to the repo root rather than creating a separate Docker-like environment.
Use Sandbox Failures To Prevent Unreviewed Commits
- Keep pre-commit hooks out of the sandbox or let them fail to force review before commits and pushes.
- Conor noted pre-commit hooks don't run in sandbox, which prevented accidental unreviewed commits.