Show HN: Ristretto, an OSS sandboxed code playground/notebook written in itself | Hacker News Reader

Show HN: Ristretto, an OSS sandboxed code playground/notebook written in itself

ristretto.codeberg.page

54 pointsbenatkin6 comments

That is super cool.

A lot of the design work we are doing in louie.ai is the LLM version, where users can talk to / control / extend the system. So more of an OS. But bootstrapping all the way as done here is intense, and proves a lot of flexibility, way more ambitious than we considered. Awesome to see!

Thanks. I hand wrote the code, often using naive approaches so it’s hopefully simple to understand. The only dependencies currently are CodeMirror and ProseMirror. I only used LLMs explain a few things to me in the process. I have also tried running code generated with an LLM in it. I am interested in and learning LLMs though, and one of the goals for it is to “AI and Web3 all the things \o/”. https://codeberg.org/macchiato/ristretto/issues/2 Perhaps LLM generated code will be included but I’ll want to mark it as such. The container with the iframe sandbox and CSP allowing following links only with a warning could be good for directly running untrusted code coming from an LLM on private data.

Edit: oh, that’s right, the colors in colors.json are from an LLM. https://fosstodon.org/@bat/111423328778967733

Great job, but were you aware of a project name collision, even though they're not necessarily in the same worlds:

https://github.com/xfce-mirror/ristretto

I was actually thinking of another project named Ristretto that I was thinking was maybe in a similar space but didn’t think it was close enough. https://ristretto.group/ There is a rust crate for it and mine builds in Deno which is based on Rust. This one also seems to be in a different enough space. Either way, this is just a component within macchiato.dev and I named it Ristretto because it’s intended to be a more minimal environment, while with macchiato.dev and resources.co I plan to work with more complex environments.

A powerful thing for LLMs is this means (I believe!) the metaobject protocol around most functionality -- introspection/reflection/dependency injection -- can be made quite powerful. So then all of a sudden, the LLM can act on everything. E.g., today you can ask chatgpt to run code in a sandbox, but not run code to edit your chat session.

A more powerful mode would be the functional equivalent: let the LLM not only simulate acting on everything in just the neural model, but generate a copy for agentic models to test manipulating them in reality, and backtrack as needed.

Nice!

Reply on news.ycombinator.com