One reason "vibe coding" doesn't work right now is the models will really fight you and try to add a lot of complexity.

It's still important to keep an overview of the system in your head with code quality heuristics. You don't need to type those in, but you do need to push back on edits.

This looks like a really nice local sandbox system: github.com/boxlite-ai/b...

This is interesting. The dumb approach is beating the complex approach. Foundational knowledge should be pushed into the top level (AGENTS.md) but aggressively compressed.

One of the core things agent driven development gives you is the ability to rapidly iterate on correct design.

Before "refactor this" was a decision you couldn't make on the fly (because of business impact), but now you can iterate on product and code in lockstep.

Nvidia dropped PersonaPlex-7B - a full duplex voice model. This is a really great sign of things to come huggingface.co/nvidia/perso...

I wish the agents would be more aggressive in pruning code... longer sessions can get out of hand unless you're constantly guarding against bloat.

TIL: git-note git-scm.com/docs/git-notes - git already has a built in note-taking system. Seems useful for agents?

Imagine an innovative EU that used the NATO funding gap to increase the research speed of asymmetric warfare:
* resistant communication platforms
* cooperative AI behaviors
* self-driving systems
* rapid scale up/down of industrial base and logistics