Google AI Studio The fastest path from prompt to production with Gemini

You've got to try Gemini for an extra ego boost (e.g. via aistudio.google.com). I'm not sure I've ever asked a follow-up that wasn't insightful or spot-on or something along those lines.

Travelling to Dublin for EuroLLVM today - looks like a great line-up of talks. Hopefully see some of you there!

You made me curious if anyone is still using MooTools. As the last release is now over 10 years ago...probably not? mootools.net/blog

Looking at the estimated figures of compute spend for frontier labs vs what those following behind with open weight models. It seems like logically there must be a lot of compute and engineering cost on failed experiments. I'd imagine you have much less room for that at Ai2? Same for Apertus etc.

Sharing the risk and reward for more aggressive exploration of model architectures and so on would also make logical sense, but it seems harder to get a consortium of organisations aligned on what risks to take, and ensure they remain bought in even if things are unproductive for a while.

It feels like a lot might depend on how the dynamics of model training evolve with respect to risk and need to innovate. A consortium model where people pay in x million USD and have some strong idea of what they'd get out seems an easier sell, especially if government funding is involved.

[shandbox] Move to using getino and the pid:inode format · muxup/medley@06766fc This protects against pid wraparound, but requires a very recent util-linux (v2.42).

Moving one of my scripts to using getino and pid:inode github.com/muxup/medley... and then adding a fallback using process start time github.com/muxup/medley...

The new util-linux release has a 'getino' utility to retrieve the inode number associated with a PID, and various other utils are updated to accept pid:inode format. This gives a pretty handy no-effort solution to PID wraparound www.kernel.org/pub/linux/ut...

This must've been one of the most challenging things I've worked on in a while. Thanks Bloomberg for sponsoring my work on this, and everyone who helped shipping a fix in time!

A picture of Igalians hanging out at WASM.io

That's a wrap on WASM.io Barcelona! Saúl Cabrera presented "Beyond Performance: 5 Years of Improving JavaScript on WebAssembly" — half a decade of deep work making JS faster inside Wasm runtimes. Great to connect with the WebAssembly community.

2026.wasm.io/sessions/bey...

#WebAssembly #WasmIo

Big news! Igalia (@igalia.com) is partnering with @eurosky.social to develop the next generation of the open social web. Let's stop waiting for others to fix our problems and start fixing them ourselves. We can just do things. More: www.eurosky.tech/s/Eurosky-Ig...

Our booth at Embedded World, Hall 4. White booth walls with "igalia is open source" branding in colorful text and "web, standards, for everyone" labels. Left wall has a monitor and the Moonforge demo station with hardware diagrams. Back walls display logos of open-source projects Igalia contributes to, including Servo, Vulkan, Chromium, WebKit, RISC-V, GStreamer, Wayland, and more. White meeting table with chairs in the foreground.

Tomorrow at Embedded World! Hall 4, Booth 4-443.

Demoing Moonforge — live OTA updates on RPi5 via RAUC — plus a Zephyr BLE app built entirely on upstream code, and WPE WebKit on embedded hardware.

Talks: Tue 14:45 Yocto best practices, Thu 15:00 web engines for embedded.

#EmbeddedWorld #ew26

"Start grinding now so you can be a level 57 prompt mage" indeed seems silly. Though it seems perfectly plausible that if applying AI tooling in a way they changed your development process, it might draw on skills you don't currently use so much. Most of the discourse is pretty imprecise.

Minipost: Additional figures for per-query energy consumption of LLMs Per-query energy consumption figures based on recent Lambda benchmarks

Some additional datapoints on per-query energy consumption of LLMs across a selection of newer models, thanks to figures from Lambda's model cards muxup.com/2026q1/minip...

Minipost: Additional figures for per-query energy consumption of LLMs Per-query energy consumption figures based on recent Lambda benchmarks

Some additional datapoints on per-query energy consumption of LLMs across a selection of newer models, thanks to figures from Lambda's model cards muxup.com/2026q1/minip...

shandbox A simple shared sandbox using unshare+nsenter.

There are many tools for unprivileged sandboxing on Linux. You should probably go use one of them. But I wrote shandbox to scratch my itch muxup.com/shandbox

/home/$user/sandbox shows up as /home/sandbox within the shared sandbox, which otherwise can only access explicitly mapped files/dirs.

Unique per-process sandboxing provides a tighther security boundary, but it's also darn convenient having this separate but shared low privileged space.

Preventing agents from reading/writing files you don't want them to access is one obvious use case, but there are others.

shandbox A simple shared sandbox using unshare+nsenter.

There are many tools for unprivileged sandboxing on Linux. You should probably go use one of them. But I wrote shandbox to scratch my itch muxup.com/shandbox

/home/$user/sandbox shows up as /home/sandbox within the shared sandbox, which otherwise can only access explicitly mapped files/dirs.

Igalia's Response to the European Commission on the Open Digital Ecosystem Strategy | Igalia Igalia is an open source consultancy specialised in the development of innovative projects and solutions. Our engineers have expertise in a wide range of technological areas, including browsers and cl...

The EC is putting together an initiative to develop their strategy for open digital ecosystems. Given its potential to push forward the development and funding of open-source software in Europe and beyond, Igalia submitted a response with some suggestions. Read more: www.igalia.com/2026/02/03/I...

Wrote a blog post about the work I've been doing at @igalia.com to implement the Temporal proposal in JavaScriptCore:

blogs.igalia.com/compilers/20...

One of the nice parts of #llvm is that often times you'll find yourself needing to do some sort of non-trivial analysis, but usually there's already a pass for it.

Here's how you can reuse a block frequency analysis to make a chess engine 7% faster on #riscv: lukelau.me/2026/01/26/c...

New blog post on the journey of the new --build-sea flag and how SEA injection works

joyeecheung.github.io/blog/2026/01...

Gödel, Escher, Bachelorette

Thanks for the helpful auto-complete, Gmail.

Graph showing the percentage of respondents deploying Go to various processor architectures. x86-64 85%, arm64 53%, x86 25%, Arm 16%, RISC-V 2%, S390x 1%, and others also at 1%

2% of Golang 2025 survey respondents are deploying their Go software to RISC-V. Take that, s390x! go.dev/blog/survey2...

I'm going to be putting my life savings into an inverse wingo ETF as soon as such a thing exists

If you're thinking of applying to PLISS, you've got three days left! pliss.org/2026/registr...

Graph showing the percentage of respondents deploying Go to various processor architectures. x86-64 85%, arm64 53%, x86 25%, Arm 16%, RISC-V 2%, S390x 1%, and others also at 1%

2% of Golang 2025 survey respondents are deploying their Go software to RISC-V. Take that, s390x! go.dev/blog/survey2...

Plus there's things like Anthropic models hosted on Google Vertex and Amazon Bedrock. In both cases they'd probably take a bit of a hit to keep customers with them...but there are surely limits.

Per-query energy consumption of LLMs Can we reasonably use the InferenceMAX benchmark dataset to get a Wh per query figure?

Nice post! You might also be interested in my attempt to get some figures on inference energy usage, but coming from the perspective of the concrete data available in the InferenceMAX benchmarks. muxup.com/2026q1/per-q... Many many provisos and limitations of course

This release contains a bunch of PRs I recently submitted to mark features I contributed to as stable/release candidate. Here is a thread about them 🧵: