aarnphm/tinymorph#79

https://www.lesswrong.com/posts/3CZF3x8FX9rv65Brp/mech-interp-lacks-good-paradigms

synthesize unknown features for models to understand

follow up with Mishra about “machine-faithful”

accessibility:

  • tooltip
  • waria
  • keyboard nav (modal editor)
    • search: cmd + K

Determine how to use tonality, different ways it can be expressed.

  • suggestion on authors’ style
  • frequency of word choices:
    • business: formal jargon
    • writers: (personal emotions) (academic theoretical work)

Protocol: on-top of files.

problems.

  • increase efficiency
  • if this engineer is given the tasks to utilize these models (interface) they would build?
  • AR models: compression of our understanding through text (reason): it acts as if it is an assistant
  • we don’t do rewrite

machine-faithful: alignment issues

data poisioning during pre-training (we can’t control this) Llama 3.2 features ablation

stakeholders.

creative writing professional doesn’t care

Users: someone that doesn’t defined themselves as writers.

  • writers are for their own interfaces and rules (no AI)

  • writers block

    • ideas curations
    • defend their arguments
    • inspirations (resources) Search
  • engineers

    • who writes
    • language game / how to structure sentences

students: pedagaogical implications (editor to help users to write essays? Probably not) no

Crucial

We don’t rewrite text (synthesize based on users past essays) tinymorph