i just yell at computers in the woods. software engineer @ NVIDIA, writer, musican, father of three. opinions are my own.
vanitas vanitatum, omnia vanitas.
@simonw It is building a 3D voxel-based solver for routing the 40+ cables for a parametrically specified digital twin of a local inference server we're designing, alongside designing the full topology of the server (8 RTX 6k PRO GPUs via a two PCIe switches). A lot of components.
Implementation Rules
- This is a brand-new codebase. Treat the docs/spec as authoritative and all
implementation as AI-produced until proven against that source of truth.
- When code behavior conflicts with the docs/spec or exposes a bug, RCCA the
issue and fix the canonical root cause. Do not preserve previous behavior with
compatibility shims, aliases, or transitional paths unless the user explicitly
asks for that.
- Remove exploratory scaffolding once it has served its purpose. Do not let
prototypes, placeholders, or workaround behavior drift into production.
- Use Biome for linting and formatting.
- Do not create or leave any source-code file over 300 lines. Split source
files before they cross that limit. Documentation and generated output are not
covered by this limit.
- Keep implementation slices small and usable. Prefer a runnable vertical path over
broad scaffolding.
@Shaughnessy119 Security is more important than ever. I'm building one of these for myself and very quickly realized that giving an agent a) access to your data, b) agency, and c) access to the Internet is a recipe for utter disaster. Adopting an airlock architecture helps but doesn't solve.
@thekitze I've been texting Bareclaw images through Telegram, I had to take a few mins and set up media support but now I can text it pics / docs / whatever and it can respond with them too (incl. if it makes something).
https://t.co/lHIRpptioz