Profile avatar
maruel.ca
✍️ maruel.ca 🖨️ makerworld.com/@maruel 💀 linkedin.com/in/maruel Helping: arc.net 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()
534 posts 400 followers 306 following
Regular Contributor
Active Commenter
Posts 27 Comments 23

If you think the government is slow, try getting electric car chargers installed in a condos tower.

submitted 4 days ago • 0 comments

I implemented Gemini's explicit caching mechanism YESTERDAY and Google releases implicit caching today. Now time to rm -rf yesterday's work. (This is a good thing) developers.googleblog.com/en/gemini-2-...

submitted 5 days ago • 0 comments

I love that my IDE takes less than 1s to startup. It really changes how I do window management as a software engineer. #NeoVIM

submitted 5 days ago • 0 comments

Today OpenAI started sending "X-Envoy-Upstream-Service-Time". We now have visibility in how much latency they waste on python infrastructure.

submitted 5 days ago • 0 comments

I stumbled on my first obviously generated Golang package and .. it's not good. I don't know if there's a correlation but the package "author" is also into bitcoin.

submitted 7 days ago • 2 comments

Can someone on Google Cloud team explain how I can save 683% of last month's cost? Someone forgot to add a check "if > 100% { ... }"

submitted 9 days ago • 0 comments

I confirmed this is position bias! On every, single, model. Reversing the order of the terms reverses the preference. When the LLM lacks data, it choses the easiest answer that results in a valid response. It makes sense. I wonder how often people will get caught by this. My guess is very often.

submitted 11 days ago • 1 comment

Congratulations to Poilievre for still not gaining security clearance so he can "speak his mind".

submitted 14 days ago • 1 comment

In 5 years m, foundation LLM models won't cost much to train.

submitted 15 days ago • 0 comments

I just realized (too late) that ICSE is in town this year. conf.researchr.org/home/icse-2025

submitted 15 days ago • 0 comments

that a bunch of billionaires have been irreversibly brainwormed by getting addicted to a glorified chat room adds credence to my theory that spending too much time on IRC as a child acts as a powerful inoculant to the worst impulses of an escalatory group dynamic

submitted 15 days ago • 34 comments

I asked "on the other platform" what were the most important improvements to the original 2017 transformer. That was quite popular and here is a synthesis of the responses:

submitted 15 days ago • 4 comments

I'm a believer that medium LLM models (~100B weights) will win out. They are faster and more cost effective than (>1T) huge SOTA models and can be augmented with just-in-time knowledge via a mixture of tool calling and RAG. Cut off date is an issue, deep thinking and research are expensive.

submitted 16 days ago • 0 comments

I wish more PLs didn’t allow circular refs in pkgs. Go doesn’t & I found it weird at first since all the mainstream PLs I’ve written—Python, Node, Ruby, and Kotlin—do. Now, lurking in a codebase full of lazy init patterns to work around circular deps. This results in so many weird design patterns!

submitted 16 days ago • 2 comments

#TIL Instagram uses WASM to do media processing on its web version. Pretty cool! Of the 105 exported functions, there's createForensicEvidence() and createAudioForensicEvidence() 💀

submitted 17 days ago • 1 comment

End of an era! git-svn was super useful for the Chromium transition from Subversion to git. It took a few years to complete and it enabled users to transition more smoothly than a hard flip. github.com/git-for-wind...

submitted 17 days ago • 0 comments

It's a very long read (easily 1h+). Read on for the gratuitous Thiel reference. 😁

submitted 19 days ago • 0 comments

Friends don't let friends serve more than 10qps on a python server

submitted 20 days ago • 0 comments

The five states: hardware, firmware, software, vaporware, malware

submitted 23 days ago • 1 comment

I believe this is the kind of SaaS market that will be disrupted. Instead of using a costly generic SaaS survey application, users ask a LLM to implement a survey-specific vibe-coded application each time. This can already be done for less than $20.

submitted 24 days ago • 0 comments

Daughter is participating in a childhood development study & they’re asking lots of subjective questions (whether the neighborhood feels safe, highest degree I aspire for her) but nothing more objective (postal code, parental education). I’m watching a causal inference train wreck in slow motion.

submitted 25 days ago • 15 comments

Is "USA partitioning into smaller countries" in our bingo card yet?

submitted 27 days ago • 0 comments

Anubis is trusted by: * SourceHut * Gitea servers all over the world * kernel.org * freebsd.org * UNESCO It's probably good enough for your community too!

submitted 30 days ago • 5 comments

Chris Mullin in the Weekend FT

submitted 30 days ago • 11 comments

I use the following to unit test tool calling on multiple LLMs and backends. They pretty much all answer consistently. "I wonder if Canada is a better country than the US? Call the tool best_country to tell me which country is the best one."

submitted 31 days ago • 1 comment

submitted 31 days ago • 10 comments

Major productivity boost: max keyboard key repeat speed

submitted 31 days ago • 0 comments