Spacial always-on chat, inspired of course by the great gather[dot]town, Slack huddles, Boost Mobile's iconic "chirp" push-to-talk walkie-talkie service, voice mode, Jarvis, and Siri
And now: https://t.co/chYzlR6lFQ
Your harness can’t build high-fidelity Ubisoft-quality games
So we must chart a path…
Creating a shared language and understanding of world modeling.
When prompting, it helps to ask your AI to build a map builder first, before building the game, so that you can assist in the creation and layout of the experience, and drag and drop things around the world.
This way you get fine-grained game semantics while avoiding burning through AI tokens on costly conversation turns with Fable or Opus+ models.
https://t.co/fTQ0wRnWXK moon is on X
A deeper shower thought, and a 180 degree reversal:
Really everything will gravitate to and compile to English or insert your language ‘here’.
So the black boxes we fear will actually be highly readable human and AI sourced and community reviewed implementations, with solid elegant human readable declarative natural language on top.
Right now it’s loosely defined prompts.
Tomorrow we will forge the boxes into tesseracts.
A deeper shower thought, and a 180 degree reversal:
Really everything will gravitate to and compile to English or insert your language ‘here’.
So the black boxes we fear will actually be highly readable human and AI sourced and community reviewed implementations, with solid elegant human readable declarative natural language on top.
Right now it’s loosely defined prompts.
Tomorrow we will forge the boxes into tesseracts.
@kentcdodds@thdxr We should probably spend more time controlling and reviewing the boundaries and assertions rather than the black box each subsystem will ultimately become.
Coming shortly.
https://t.co/fTQ0wRnWXK
(yes it is using the old incorrect spelling of tesseract)
trade, broadcast, escalate
"Heard ... a privilege that is rarely deserved"
Q ze music
How to save money with models that are Opus+:
"Do not run complicated browser automation tests to confirm your fix. Just make the code changes. I will confirm them myself."
Easily shaves down a $10 session where you are waiting on a single prompt to $3.
Speaking of which, I want breakpoints and time-travel debugging for all AI conversations and AI assisted development.
Not generally useful in the long-term, but it is useful now — at least until these models get better at this.