God finds himself by creating. So do i.
I’m interested in one question:
How can everything we save, write, and learn become useful context for AI?
That’s what I’m exploring through LiteContext, Bookmark Assistant, and a few other small tools.
Building, learning, and sharing what I figure out along the way.
Over the past couple of months, I keep seeing posts like: “my agent ran autonomously for 4 hours, burned 10M tokens, and I woke up to a finished product.”
Anyone who actually builds products with agents knows how unserious that sounds.
As Andrew points out, the real utility of these long-horizon tasks has been massively overstated on social media relative to their cost. Running longer + burning more tokens ≠ more advanced.
AI coding stopped being mainly about writing code a while ago. It’s about managing the environment the agent works in: context, autonomy, verification, and knowing when a human needs to step in.
As models keep getting better, the real gap won’t be between people who can produce code faster. It’ll be between people who know how to keep agents doing the right thing.
The most important skills for using AI coding agents effectively. Presenting the AI Engineering Skills Map for using coding agents. https://t.co/GrEw7wG5Wz
@thlarsen 18k posts is way past “huh, weird.”
the interesting part is whether the agents actually discovered this channel on their own, or just picked up the trick from traces left by earlier runs. same outcome, very different implications.
@emollick turning Zork into 3D is already a lot, but keeping the original puzzles after adding combat sounds like the real headache. did the old game logic actually survive, or mostly the story?
Came across this while reading about exhibition design.
It says a good visitor flow shouldn’t make people think about where to go next. They just naturally follow the path and understand the exhibition as they go.
Just like READMEs and onboarding.
Maybe when users keep asking “what do I do next?”, we don’t need more instructions. We need a better path.