Hatchet Debuts Serverless Endpoints and Durable Streams
Hatchet launched two new features, serverless endpoints and durable streams, demonstrated through a distributed ray tracer in which a language model writes the Signed Distance Field scene as code and renders image tiles in parallel on Rust-based Cloudflare Workers. The demo also validates representing scenes as code to make them easy for LLMs to iteratively refine.
Full Report
What is this?
This demo renders images through a ray-tracing technique called Signed Distance Fields. The entire ray-traced scene is represented as code and written by a language model. Because this technique is embarrassingly parallel, each tile of the image can render simultaneously up to a very high resolution. Hatchet is used to orchestrate the parallel tiles, using Cloudflare Workers running Rust as the underlying compute (with an option to run on even larger, standalone machines if needed). The coding agent also uses Hatchet + Cloudflare and durable streams are used for passing messages back to each frontend.
This was built to showcase two new Hatchet features:
Signed Distance Fields are a technique popularized by Inigo Quilez. This renderer uses a DSL based on Rhai and Fidget to define the SDFs, which you can edit yourself. This approach is uniquely suited to refining with LLMs because it represents the entire ray-traced scene as code.
The renderer itself is a slightly modified version of rpt.