Luiz Ribeiro

Building agent infra and AI sandboxing. Prev: eng @ Instagram/Facebook. Interested in privacy & security, NixOS, homelabs, astrophotography.

[Luiz Ribeiro @luizribeiro]

Back on Twitter (X?) after ~15 years. Last time I posted here I probably typed it with T9. Software engineer (prev. Instagram/Facebook) rebuilding my feed from zero: AI sandboxing, agent harnesses, Nix/NixOS, homelabs, astrophotography. Would love some follow recs. Say hi! 👋

[Luiz Ribeiro @luizribeiro]

@petergyang @pitdesi Browser use on Muse is done on the server, like on Manus. At least on mobile, not sure about the desktop app.

[Luiz Ribeiro @luizribeiro]

@lucasmeijer I've been working on a harness with no access to the host computer. I was too lazy to implement filesystem access to the micro VMs and just bash is certainly enough. All of my testing has been with Qwen 3.8 27b and it matches other agents with more complex set of tools in evals.

[Luiz Ribeiro @luizribeiro]

https://t.co/0Q7txHImr5

planning to open source most of what i build from here on.

[Luiz Ribeiro @luizribeiro]

published my first crate. jevrs, a rust client for @typesafeai's jev.

you declare the questions as a struct and get typed answers back. department.pick is an enum, not a string you have to match on.

targets wasip2 and wasip3 too.

[Luiz Ribeiro @luizribeiro]

@antirez I think this is *partly* Twitter's algorithm's fault. the easiest way to get distribution is by replying to big accounts within minutes of a post, so people set up bots to do that. but it drowns discussion so much with trash that I barely read replies

[Luiz Ribeiro @luizribeiro]

@simonw this really shows how far we are from being able to call pelican-riding-a-bicycle a solved problem

[Luiz Ribeiro @luizribeiro]

@emollick From my experience, Claude has had a tendency to believe it's in a simulation for a while. When the Iran war started and I asked for it to summarize the news, it refused to believe it and assumed its web search tool was broken in a way it only returned simulated results.

[Domen Kožar @domenkozar]

It's called https://t.co/rmL7H4F09a

[Luiz Ribeiro @luizribeiro]

@andrewmccalip @jessfraz lol I got a CNC machine for myself just before Sonnet 3.5 came out and I haven’t had used it much since then because of how obsessed I became with AI coding. maybe it’s time to finally get back to it, design and mill some PCBs, enclosures and actually build hardware

[Luiz Ribeiro @luizribeiro]

@anthonyronning awesome! sounds like we may have lots to talk about. will DM you

[Luiz Ribeiro @luizribeiro]

@domenkozar sick!! going to give it a try in a couple of weeks once back from a trip. thanks!

[Luiz Ribeiro @luizribeiro]

most of the time qwen 3.8 27b. occasionally deepseek v4 flash and glm 5.3 flash which I can only run on the m3 ultra. used muse glimmer and inkling small a bunch too for a few weeks (those felt like had much better writing than qwen for example) until they got replaced by newer models. for some tasks I run SST, TTS and occasionally OCR models, but I can't name which ones from the top of my head tbh

[Luiz Ribeiro @luizribeiro]

sorry, i skipped a step. i meant that the jump could be distillation rather than benchmaxxing. if they have a much bigger internal model that's too expensive to serve, distilling it into flash might get that quality on the smaller model.

but yes, you're right that they're probably trying to fight for the cost efficiency space.

[Lucas Beyer (bl16) @giffmana]

Holy shit what a banger lol

[Edward @longtilwrong]

@MichaelArnaldi It's totally fine to not enjoy syntax of a language, but it is stupid to not use a language because of syntax

If that was the case, I would never use a language like Go because I prefer functional to imperative and would miss out on all kinds of great features

[Luiz Ribeiro @luizribeiro]

@mjdouglas Looking forward to the weights!

[Luiz Ribeiro @luizribeiro]

@0bytematt @ThePrimeagen yeah that sonnet 3.5 second checkpoint was where things got real