asos

technologist, s/w dev.

[asos @taosx]

@luisvelasco Any tips and info on prompting, making models work in environments they never seen before, evals (recommendations), systems around the model that may be interesting.

[asos @taosx]

@thsottiaux Then you can prompt again the main.. I imagine something similar to what I'm using rn: tmux subagents. I could even cancel the main orchestrator and continue in a subagent. idk, just an idea

[asos @taosx]

@thsottiaux Would love an Operator mode for Codex: live, almost raw view of what the main and subagents are doing, including tool calls, bash, diffs, everything as it happens and with one-click cancel that pauses the whole run. Full Access, but observable and interruptible.

[asos @taosx]

@__tinygrad__ I find interesting the direction of etching LLMs into chips. There is financial incentive, a self-improving harness with a good model that becomes better, and every 1/2 years buy a new one? Is it feasible for 10T parameter models or is it prohibitevely expensive?

[asos @taosx]

@sqs @towrywang @AmpCode Yesterday I visited https://t.co/Dj7cCwGbG8 on linux, https://t.co/Ii8qjOHJ7S rendered and then hanged. Later the gpu crashed and restarted. After I opened it in Firefox without issues (still slow). Checked with chrome and it hanged again.

[asos @taosx]

@neilsuperduper @WisprFlow https://t.co/HRk8tPmuZo for mac...for linux I have a small custom binary that just combines parakeet2 with postprocessing on qwen3 (it would have been simpler but damping/silencing had multiple edge cases on linux)

[asos @taosx]

@neilsuperduper @WisprFlow And that's exactly why I'm using open source and offline transcription.

[asos @taosx]

After reading https://t.co/dlVyRgRyNl I wanted to see if there were still some opportunities for optimizing my @pidotdev and lazily vibed a nice html page :)

enjoy: https://t.co/U8CCRDA48d

methods were backtested on my past PI sessions which is already modified (minimally)

[asos @taosx]

@deanwball I'm happy to see how delulu ppl think. Maybe just spreading propaganda our openai head.

[asos @taosx]

@thsottiaux 1) gpt-5.6-sol is faster, smarter, uses far less tokens during runs and understands my intent much better even though I'm not a native en speaker. 2) I don't feel bad giving money to openai compared to others that treat users as non-humans which is ironic.. anthropos ?

[asos @taosx]

@TrisH0x2A How do you backups ? Litestream?

[asos @taosx]

@davidcrawshaw You can compare the quality of packages (based on code, not gh stars) from different PLs, you can see that well typed languages raise the bar. The usage of Rust with AI agents it's probably due to models finding more code online to be trained on that say ocaml or haskell.

[asos @taosx]

@davidcrawshaw I don't find the argument convincing. By enforcing constraints on the agents will lead to better architecture and code, and it's easier for a dev to identify at a glance when AI went off the rails by looking where it's trying to bypass the borrow system by doing something weird

[asos @taosx]

This rule in AGENTS.md seems to do wonders for me:
**Simple, intentional design** — explicit domain modeling; direct refactors over indirection; no compatibility shims, transitional structure, or convenience abstractions.

[asos @taosx]

A main issue when reviewing gpt5.5 output is that it adds tons of indirection where a comment or a one line variable decl would be better. A function with 100 lines is better than a 10 line function with 10 indirection points, especially whey are one line functions, wdyt?.

[asos @taosx]

@jarredsumner amazing work! Did you find a nice solution to converting from tagged pointers to more idiomatic Rust code while avoiding circular dependencies between crates?

https://t.co/uf5QA6qYio

[Jarred Sumner @jarredsumner]

one of the tricky things about the rust port is layering. it’s currently many dozens of crates, which speeds up compile times but blocks cyclic dependencies.

a lot of bun’s zig codebase uses tagged pointers for interfaces, for things like:
- event loop tasks
- process exit callbacks
- nonblocking file i/o

it sounds like traits are the idiomatic rust approach for this, but traits are costly - they don’t get devirtualized at compile-time or through LTO. same for fn pointers.

in zig we didn’t split up packages so this wasn’t a problem. But I want compile times to be faster and ideally some parts of bun can be a crate for others to use without pulling in the whole thing.

can someone with rust experience suggest a better approach that has ~no runtime or memory overhead and is visible to the compiler/linker?

what the current rust code does is:
- macro that generates extern “Rust” declarations for each type
- release build as single codegen unit and enable LTO
- no shared libraries, rely on linker’s LTO to do the inlining into direct function calls based on the tag

this seems to work but it sounds uncommon/messy? is there a better approach that avoids making it slower?

[asos @taosx]

@entrepeneur4lyf @Howaboua I've tested most variations of hashlines with different names for the tool, for the hashline format for the tool description and instructions and I got higher error rates with most models i tried ds4, gpt-5.5, kimi. I've ended up on the edit tool with better recovery + levensh..

[asos @taosx]

@jxnlco Went through all the projects I did over the years, only mitmproxy I'd prefer in Rust, others are either too simple or to tied to the py ecosystem.

[asos @taosx]

@mitchellh @ellie_huxtable being able to scroll, copy with the mouse (TUIs), multiple panes, easily resizable and a nice one is stacked panes for me but the most critical for adoption may be the bottom bar showing the modes and available shortcut actions in that mode.

[asos @taosx]

@ellie_huxtable It's great until you try to use it daily: freezes, mouse stops working. So many features and so many issues. Yesterday was annoying trying to migrate to tmux (has it's quirks as well) but it's rock solid after using some plugins plus some config work.