I built a TypeScript framework for AI agents that replaces shelling out and parsing stdout. Agents discover and call typed functions with structured inputs and outputs. Permissions are first-class — every operation is allow, ask, or deny — so you get granular control instead of "allow bash: yes/no." LLM-generated code runs in a SES sandbox with no access to fetch, require, or process.

I'm curious how this performs on everyone's machines? Worked great for me on macOS but not as well on Linux (supposedly there are still WebGPU quirks?)

I was surprised to see the drastic performance difference between gpt-oss and GPT-4. Makes me wonder if gpt-oss just re-used the GPT-5 tokenizer...

I got 11/28, feeling pretty proud of myself honestly

I wrote about how humans are developing new linguistic patterns to distinguish themselves from AI-generated content. I predict the emergence of rapidly-changing new dialects which I call "Human English."

Why we should combine JS build tool config files.

Quick tutorial on how to use Accelerate to run inference on LLMs in parallel

Haha true, and I'd definitely agree that Chromium's borders are one of the least important parts of the software.

However, just because something is less important than something else doesn't mean it's not worth changing.

I doubt most businesses would leave lots of ugly graffiti on the side of their bike shed, even though actual business operations were much more important.

Bike shedding refers to being so distracted by unimportant details that important business isn't taken care of. It's possible to give the shed a good coat of paint once in a while while still taking care of more important matters.

but I'm also one of the people who likes borders the way they used to be, with visual indication of where you can click with your mouse and I also use focus-follows-mouse, so I like to light up the border of the active window. So, to me, this isn't bike shedding, it's bike shredding.

One of the most enjoyable articles I've read in a long time.