Marc-Antoine Ruel

Marc-Antoine Ruel

@maruel.ca

✍️ maruel.ca 🍲 caic.xyz 🖨️ makerworld.com/@maruel 💀 linkedin.com/in/maruel Embedded, ML, Go, Perf: github.com/maruel periph.io fuchsia.dev Wrote Google Chrome's: - large parts of it's CI - small parts of it's sandbox - window.print()

560 followers366 following1.1K posts

Pinned

And yes, Epic Games Launcher is pre-installed on CI machines used by many Fortune 500 companies to build the binaries they ship to end users. If someone can compromise Epic, they get a turbo boosted SolarWinds-style supply chain attack at scale.

Both run on Azure. The Windows one runs the Windows Server Datacenter edition in Test Mode. Both use a local SSD for performance. Windows sucks at 2 things: opening files and creating processes. TL;DR: This is due to the way processes are created and the driver framework. Display looks like this:

[While extracting as much juice out of a esp32-s3 as possible] I'm not sure how much time it would have taken me to notice the difference. Thanks shitbag! (I now name all my AI agents shitbag, collectively known as shitbags.) ((Only Muse was amused by learning its name.))

" That trial found a substantial gap: PSRAM code/data placement cut H.264 time from 236 ms to 103 ms per frame, and delivery reached the firmware’s 5 fps limit with AAC and Wi-Fi active. " Dang I hadn't realized how slower its flash is while running code. I now have a webcam with AAC audio! yay!

git diffnt: git diff without tests git config --global alias.diffnt '!f() { git diff "$@" -- . ":(exclude)*_test.*" ":(exclude)test_*.*" ":(exclude)*.test.*" ":(exclude)*.spec.*" ":(exclude)*test/**" ":(exclude)*tests/**" ":(exclude)*__tests__/**" ":(exclude)*testdata/**" ":(exclude)*e2e/**"; }; f'

I'm really excited for this release. While not enabled yet in the UI, it unblocks the same flow that Swarming uses on the Chromium CI. For those who understand how the Chromium CI or Bazel's Remote Build Execution API works, you'll understand. Join discord.com/invite/2ec7e...

discord.comJoin the caic Discord Server!https://caic.xyz | 11 members

finally released caic v0.12.0, nearly a month since last release. There's just too much to list! Frankly it's been great to use it as I improved it: good keyboard-first navigation, subagent and UX improvements, and a long tail of polish items. github.com/caic-xyz/cai... Docs: docs.caic.xyz

finally released caic v0.12.0, nearly a month since last release. There's just too much to list! Frankly it's been great to use it as I improved it: good keyboard-first navigation, subagent and UX improvements, and a long tail of polish items. github.com/caic-xyz/cai... Docs: docs.caic.xyz

github.comRelease v0.12.0 · caic-xyz/caicThis is a hell of a big release. Features Native subagent activity from Claude Code, Codex, OpenCode, and Pi now renders as inline cards with foreground/background distinction, batch cards, paused...

Marc-Antoine Ruel reposted

DeepSeek V4.1 Flash tech report section on reward hacking. I wonder what repercussion signal means here, I don't see it mentioned anywhere else.

The left laptop costs 7000$+, the right one costs 1500$. I wish there was a way to work outside without bleeding money, as I prefer the one on the right.

Real-time audio LLMs process audio at 16kHz or 24kHz and generate audio at 24kHz so the only acceptable frequency to sample at is 48kHz.

apenwarr@apenwarr.ca

This weekend I learned why sending 44.1kHz 16-bit lossless music, on macOS, to a 44.1 kHz 24-bit DAC, causes audible (to me) distortion. And now I can never be happy again

my life is now generating 1500 LoCs diff in 2~3 prompt, and then doing tens of round trips to fix the structure of the code. Ultimately it's still faster than writing it myself but it's somewhat draining.

one gotcha with Go language server is that it fails with conditional compilation, so if you plug gopls in your coding agent and use conditional compilation (multi-OS, smoke test, etc), you're likely hurting yourself.

TFW when you edit a text in Opus 4.8, remove a LLM form, review it in GPT-5.5 and it adds *exactly the same formulation back*. Yay for training diversity.

Told my coding agents to never document code. I prompt it to document when it actually matters, i.e. there's subtleties that are not obvious from reading the code due to external factors. The code is so much cleaner and the agent is less confused as there's no documentation drift.

I'm leaning towards recreating the concept of task trees similar to how Swarming tasks fan out (CI system for Chromium), but this time native for agents. Ping me if it resonates with you.

super excited with caic v0.10.1. Slowly migrating the custom Voice agent tool call into a proper MCP server. As a side effect, you can integrate your **personal** caic instance into claude, codex, claude.ai or chatgpt.com (!)

one funny thing with wealthy people that are not status driven is that they will be like: " I'm wondering if I should get a Porche or a Hyundai... "