started a mini site / side project this weekend.
It's just things that makes me think, or I like. No visuals just links to places.
https://t.co/CInXNZDUL4
Itβs going to happen, again. The AI mockups of what could be will end up being more impressive than the actual thing.
AI Brand Psychosis is a thing that brands have to deal with now.
I've continued working on my little m5 device that acts as a push to talk / speaker to gemini realtime. I'm using it more and more. I love that it starts up in five seconds, that it turns off when I tell it to turn off, can actually set and manage multiple timers, create and retrieve files, and search / read the web so quickly. This isn't an openclaw / agent harness, or at least is a very basic one, just a CloudFlare worker calling Gemini and using the worker apis. Responses are basically instant. Really really fun way to use technology. I use the system, want things from the system, vibe them in with codex mobile, have it push and ota update, and get the new firmware on my tiny device
prompt = intent
intent => function
function => structured json tree
leaf of tree === user facing UI decision
fold back you have a permissioned scope with user input.
What if you could unfold prompts into UI inspired by Folding Constructs.
intentβplan.trip(copenhagen-syd-festival)
A prompt becomes an intent.
The intent unfolds into a tree of UI primitives you can fill in.
The filled context decomposes into agent briefs β each itself an intent that can be unfolded again.
Recursion all the way down.
A user types a prompt.
An LLM unfolds it into a tree: sub-flows in the middle, fixed UI primitives at the leaves (tags, sliders, pickers, text fields, buttons).
@isaaccyn okay so my hot take is don't even try to have a system. the more you engineer at the beginning the more annoying it gets over time. the simpler the better.
One of the craziest things that I don't understand really why but is fucking amazing is that somehow because of pretext from @_chenglou your search is just in hyperdrive.