Learning diary › Year 2025 › July, 2025 › 2025-07-11 [2025-07-11]

✍️source

- agent
    - Not So Fast: AI Coding Tools Can Actually Reduce Productivity
        - 19 percent productivity drop
        - it's most interesting how objections are addressed
            - The John Henry Effect: motivated to “beat the machine”, not observed
            - Underuse of AI: not observed
            - Cheating: can't explain a productivity loss
            - Overly optimistic time estimates: should cancel out
            - Skewed task definitions: randomly assigned
            - Dropout effects: incomplete tasks are evenly divided
            - Out-of-date tools: not observed
            - Inaccurate time reporting: use alternative metrics
            - Overuse of AI: carried away by exploring AI
            - Lack of experience with AI tools: prior experience reported, no improvements through the study observed
            - Difference in thoroughness: 47 percent more code by AI
        - quotes from First, They Came for the Software Engineers… - by TarenSK
            - "Typically, large productivity boosts occur for small, well-defined, greenfield projects, or when an engineer is first learning a new language or API"
            - "For other work, gains from using current AI tools are often far more modest – and potentially entirely offset by increased time needed for review, debugging, integration, and managing AI quirks."
    - AI agent benchmarks are broken
        - many validity issues are identified in uiuc-kang-lab/agentic-benchmarks
    - We're Light-Years Away from True Artificial Intelligence, Says Martha Wells, author of the Murderbot Diaries series
    - Vibe Kanban – Kanban board to manage your AI coding agents
        - tried, required too much GitHub authorization, settled on local-first backlog
- An almost catastrophic OpenZFS bug and the humans that made it
    - follow-up: What people said about an OpenZFS bug
        - wrapper types can be done, but they are very difficult to work with
        - unused variable warning can't catch it, as it is only a dead store: we store a value that is never read
        - static analysers all catch it, but they suffer from noise
    - OpenZFS Bug Ported to Zig #zig
- FP8 is ~100 tflops faster when the kernel name has "cutlass" in it
    - the PR
    - "that is a very usual way (known practices) of vendors applying specific optimizations for known things"
- Pangolin – Open source alternative to Cloudflare Tunnels
    - client: fosrl/newt: A tunneling client for Pangolin
        - a custom user space WireGuard client
    - may be cheap to self-host on RackNerd
    - awesome-tunneling: List of ngrok/Cloudflare Tunnel alternatives and other tunneling software and services. Focus on self-hosting.