yev
410 posts

yev
@itisyev
building something new
surfing the interwebs Katılım Eylül 2013
871 Takip Edilen708 Takipçiler
Sabitlenmiş Tweet

@MaxDiffusionRL This is a repeat of Signal being for drugs and is now for privacy
English

@anothercohen request for startup: site that tracks all forbes 30u30 recipients and whether they've committed fraud yet or not
English


@ivanburazin I really wanna argue the way things play out is with some new network in which humans and AIs are interacting with the same platform
However there are tiers of tools whether you're consumer, prosumer, or professional so agent-strict apps make sense
English

Feel free to give it a read or have your agent summarize it instead
vers.sh/blog/headless-…
English

Not sure what you mean by "preempted" but I'd be inclined to say it runs where it needs to as it's invoked.
This isn't a JIT optimization using WASM project, it's instead a toolkit for two-way Elixir->WASM (compiling Elixir programs into WebAssembly) and WASM->Elixir (using WebAssembly programs in Elixir, so you can bring over Rust, C, or whichever language you like!)
English

@itisyev When executing WASM from Elixir, is WASM execution preempted every 4000 reductions?
English

There are plenty of flashy demos like Cursor making a browser or Anthropic making a C compiler but they often run into quick criticism of not being working solutions.
Rather than join the mob of "there's no real use here!" I wanted to find an open-ish (as in not something with many/recent examples) problem and corral coding agents at it.
A while ago I had fun writing interpreters in Elixir and saw an opportunity here to bring something to that community
English

@itisyev Just curious is there a real project/idea you working on, that push you to made it or just "why not" ?
English

@ChristophGockel Fair, but I don't suspect tech names from 20 years ago to be worried about a modern project
English

@goneflyin I recently ran into a 2M input tokens/minute limit but I didn't intelligently cache prompts which resulted in hitting that ceiling. I don't recall running into it with firebird so my suspicion is that the pi agents were caching tokens well
But, based on napkin math, at least 2k
English

@itisyev Dumb question, but … all in, what was the token cost for that 1bn tokens?
Just curious of a ballpark. It’s an interesting metric I haven’t quite seen before.
English

Orb is like Vecty (Go WASM frontend) for Elixir. I wanted something that'd topically fit well in the awesome list for WASM
In short, firebird covers both Elixir->WASM (refined for general cases as well as incorporating into Phoenix) as well as WASM->Elixir (so you can use WASM modules from inside an existing Elixir project, similar to Orb but generalized so you can use Rust, C, etc)
English

@itisyev Interesting!
How does this compare with Orb from @royalicing?
English

Since I didn't have "checkpoints" or steps for the models to validate themselves (although I suspect this just means including in the original prompt), I did step in a bit
1. Original two files results in close to 1 billion of tokens consumed (see pi-vers on GitHub)
2. I steered it to provision more testing/benchmarking agents after I configure branch protection on GitHub (to force guardrails)
3. It then runs off predominantly on its own and, to my surprise, agents would basically shut off when they had exhausted what could be done
4. I vibe a bit of clean up at the end so guides are readable etc
So, all in all, a *ton* of work that'd be tough without LLMs. Rather than sit on this for weeks I can blast it out in a weekend
English

@itisyev Billion tokens via coding agents to build an Elixir/WASM toolkit? That's the most 2026 dev story ever. The real question: how much of that was agent iteration vs human guidance? Would love to see the breakdown of agent vs human tokens.
English

@unifiedh @magpie_lang @PyTorch How does it compare to a limited subset of LLVM? Based on website example of Magpie
Wonder if something in that direction could unlock more compilation targets (ie WASM)
English

.@magpie_lang is an AI-native language.
Ported @PyTorch to Magpie & Rust. Now at 91% performance of optimized MPS kernels for BERT training, at less than 10% memory consumption.
Will release after some bug fixes.

Daniel Hong 🪄@unifiedh
I built the world's first programming language designed for LLMs and AI Agents. Introducing Magpie. Check it out yourself at: magpie-lang.com Details below.
English

@itisyev I'm actually using Rust -> WASM for a new AudioWorklet-hosted real-time language that is designed to pair well with SuperSonic - my recent port of the SuperCollider scsynth to WASM.
sonic-pi.net/supersonic/dem…
English

Sonic-Pi is getting a rewrite in Elixir (shoutout @samaaron ) and it's already evaluated Lua as the sandbox for untrusted code but...
May I suggest using WebAssembly with Elixir?
yev@itisyev
Introducing firebird! A modern Elixir/WebAssembly toolkit that was the outcome of billions of tokens via coding agents
English

I set up some coding agents to built a modern bridge between Elixir and WebAssembly
Viva la functional programming languages!
yev@itisyev
Introducing firebird! A modern Elixir/WebAssembly toolkit that was the outcome of billions of tokens via coding agents
English

Otherwise, if you scrolled all this way for the GitHub, here ya go github.com/hdresearch/fir…
English

If you're curious about what Elixir and WebAssembly are or why bridge the two together, you can read about it here yev.bar/firebird

English






