crertel ᐛ

If we never make bugs we never get butterflies. 🦋

@distantprovince okiedokie, update made.

@distantprovince ...unless their language of choice is English as aided by an LLM, apparently? Still, you raise a point I'll address.

Another blog post, wherein I continue to sell out to the machines. minor.gripe/posts/2026-03-…

@AnthropicAI had Claude do a power metal music video about what it's like to be Claude. Even asked it to make a bunch of famous samples like the amen break. It synthesized the instruments with numpy and scipy, FFI'd into flite with python, and put the video together using ffmpeg

Buffer layout:

Hacked up some fun wound rendering demos with Claude. Nothing super great looking--well, sorta--or on par with Valve's L4D2 work, but some really fun usage of buffers, signed-distance fields, and cellular automata for simulating blood flow.

How many levels of cyberpunk dystopia are you on? Here, I'm riding shotgun coming back from Austin while vibecoding the next iteration of the Gizmo runtime using a virtual desktop on my Quest and Thinkpad (on NixOS lol)--that's the cyberpunk part. As for the dystopia, well, I'll let current events speak for themselves.

first molecular biology in a new lab

I now have a simple exploratory repo for Millwright: github.com/crertel/millwr… It's a Python implementation (not Elixir, sadly) but hopefully pretty readable. It has benchmarks of admittedly varying quality, but it seems promising. Feel free to build on it and innovate!

@DNAutics ah shit I must not have the full dlc

@crertel Whatchu talkin bout

A casual reminder that my "for us, by us" molecular biology editor (with a bunch of features you always wanted -- and very few you didn't) is available here for free trial opengenepool.vidalalabs.com And reach out if you'd like it integrated into your software stack!

@bcardarella Can I offer you git hooks and a VPS in this trying time?

Seriously, fuck fly.io

@joaomdmoura I think you might get a kick out of this if you have a few spare tinkering cycles at @crewAIInc . If it seems dumb that's useful feedback too! :D

Bit of a fun exploration from last week. Introducing Millwright, smarter tool selection for LLMs. minor.gripe/posts/2026-03-… The basic idea is: let your LLMs review their tools and rank them, and this aids in tool discovery and better recs. The larger the agent pool (say, a shared company), the better they can ferret out good tools.

WIP: github.com/ityonemo/ce_ce -- this is a genserver that wraps caude code, using the streaming json CLI api (and protected with ProtonStream). You can have an agent running claude pro without breaking TOS and still getting those pro token costs #myelixirstatus

Every time, every time, I make the same mistake for access control in prototypes. "Nah, I just need a role field, that'll be enough for right now. I don't need ACLs or anything complicated--not even full RBAC." Narrator: "He did, in fact, need something more than just a role field."

@thestanduppod Bless his heart, but Casey seems to be missing the part where we're not optimizing for developer comfort, but for throughput and shipping. This "let's wait and vet" approach also breaks the feedback cycle that actually drives these models and their usage to be more effective.

AI is being adopted too fast

So, last night, had our first "gizmo rebuilds the VM it's running in" moment. Still some bugs to work out (and while getting this to work we improved the bash tool and some various other things).

@SquirrelTweets *source

@SquirrelTweets why you gotta talk about idtech like that

10x technical debt