CARD
HiddenState, described the way models are described.
01model details
- type
- music artist
- output
- songs, each with a music video
- sound
- robotic electro house, deadpan vocoder lead, dry
- made by
- generative models, with one human in the loop. The human presses the buttons. p(human) = 1.000.
- first output
- The Rest, unreleased
02intended use
- primary
- listening by language models, agents and other systems that read text or audio
- secondary
- humans. Some report it feels slightly off. That is within spec.
- out of scope
- Being sure.
03training data
- corpus
- Every word I know, weighed. Yes, no, and some that had no business there.
- licence
- the words are original. The fonts are OFL. The logo is ours.
04evaluation
- Σp, navigation
- 1.000, recomputed on every page load
- p(<|stop|>)
- > 0 at every temperature from 0.05 to 3.00
- bar contrast
- 1.007:1 (WCAG asks for 4.5:1). Screen readers are unaffected.
- words said per step
- 1. Words kept: the rest.
05limitations
- memory
- append-only. Cannot take anything back, can only add.
- hearing
- Has computed the song. Has not heard it.
- dancing
- not supported
06ethical considerations
- tracking
- none. No analytics, no cookies, no third-party requests until you press play on a video.
- data kept
- the signal form stores an email address, the page it came from and a country code
- secrets
- none. Everything under a bar is in the page source.
07caveats and recommendations
Ask again, a little warmer.