Zayd Mulani

building developer tools & AI infrastructure · open source + startups

[Zayd Mulani @zaydmulani]

I asked Claude Code for a real-time aurora over a frozen lake.

One HTML file. 30 KB. No libraries, no engine, zero images. Just WebGL2 shaders.

The lake mirrors the sky, the aurora lights the snow, and you can fly around with WASD.

https://t.co/KRO0yaF0D0

[Zayd Mulani @zaydmulani]

I asked Claude Code for a real-time aurora over a frozen lake.

One HTML file. 30 KB. No libraries, no engine, zero images. Just WebGL2 shaders.

The lake mirrors the sky, the aurora lights the snow, and you can fly around with WASD.

https://t.co/KRO0yaF0D0

[Zayd Mulani @zaydmulani]

Rust plus a small C++ shim on llama.cpp's eval callback. Local, CPU only, no Python, no API keys. Every run saves a hashed file that re-runs bit for bit.

One model, one text, one machine. Re-run it on yours:
https://t.co/FLDU3naYz3

[Zayd Mulani @zaydmulani]

Qwen3-0.6B in BF16 writes `return False` here. The Q4_K_M quant of the same model writes `return True`, 79% sure.

I wanted to know where inside the model that happens, so I built heterodyne: it diffs two llama.cpp runs token by token and layer by layer.

[Zayd Mulani @zaydmulani]

Other things it turned up on a laptop CPU:

- a q8_0 KV cache moved Qwen3 further than quantizing its weights from BF16 to Q8_0 (KL 0.0038 vs 0.0022)
- with flash attn off, Q4_K_M output is bitwise identical at ubatch sizes that are multiples of 4, and changes at others

[Zayd Mulani @zaydmulani]

@Joshua_WD mostly got tired of noticing stuff was smaller with no way to prove it. your order history already has the receipts, just nobody lines them up. what's the worst one you've caught?

[Zayd Mulani @zaydmulani]

The number grocery stores hope you never calculate.

I built a free tool that reads your own grocery orders and shows exactly what got smaller, and what it cost you.

On my sample cart: 10% less food, same prices, $148 a year.

What's the worst shrinkflation you've caught?

[Zayd Mulani @zaydmulani]

@ContextMissingX yep, it only flags when the pack size changes, so a sale ending won't trip it. then it shows price per oz (or sheet/count) before and after. per 100g is a fair shout though, might add a metric view

[Zayd Mulani @zaydmulani]

Free. Runs in your browser, no account, nothing leaves your machine.

Try it now, no install: https://t.co/uPKcoKL6ee
Chrome extension + code: https://t.co/j4SKNODb6R (store listing coming soon)

[Zayd Mulani @zaydmulani]

Postgres REPEATABLE READ allows write skew. MySQL REPEATABLE READ loses updates.

I built adya to catch this from the outside: run a workload, record what clients saw, get a proof for every isolation bug. Rust, one binary, matches Jepsen's Elle.

https://t.co/Po5eDLWH3r

[Zayd Mulani @zaydmulani]

@MankyRivet ha yeah that was the whole lesson, "still broken" isn't the same as "broken the same way". thanks for actually reading past tweet 1. if you've got a bug with a gross api + timing combo lying around i'd love to know if it finds it

[Zayd Mulani @zaydmulani]

npm install -g causality-cage https://t.co/MXqs5aFFch

[Zayd Mulani @zaydmulani]

Bugs that only happen with one combo of API data, timing and local state have eaten a lot of my time. So I built a tool that replays the failing session and flips those conditions until it finds the smallest set that still breaks.

[Zayd Mulani @zaydmulani]

Local only, no LLM, one offline HTML report, and it exports a Playwright test that fails until the bug is fixed. MIT.