← @bgSponge

Building software factories (with no slop)

Summary

David argues that software engineering must shift from reviewing human-readable code to an interpretability model that treats codebases as black boxes with inspected intent and behavior.

  1. Slop results from unbounded generation.

    High-quality models like Sonnet-3 have mastered coding, but quality degrades when supply exceeds the fixed rate of human inspection, turning reviews into mere formalities.

    Slop is what happens when generation is unbounded and verification is bottlenecked on a human reading the output.

  2. Code readability is a temporary phase.

    Current AI writes human-readable TypeScript or Python, but future systems may emit machine code or AI-optimized languages that are opaque to humans.

    Maybe we invent new languages purpose-built for AI to write - far more token-efficient, optimized for a reader that isn't human.

  3. Engineering focuses on solved problems.

    The profession survives AI because engineering is about solving business problems through systems, whereas programming is merely the act of writing code.

    an engineer's output is a solved problem, and code was only ever the medium

  4. Treat codebases as black boxes.

    Instead of reading implementation, engineers should use domain-specific pieces with defined inputs, outputs, and sensors to interrogate behavior through the interface.

    You never read the implementation. You don't need to.

  5. Review decisions instead of code.

    David prompts agents to produce a ledger of silent decisions, ranked by confidence, allowing engineers to audit the most consequential choices rather than thousands of lines of code.

    I explicitly prompt it (via a skill) to hand me a ledger of every decision it made where the spec was silent

Read the original ↗Save to your Sponge

Summarized with Sponge. Quotes are the author’s own words.

More from @bg

All readings →