Under the Influence
Work in Progress
Language models are creatures of habit. If you ask a LLM to paint a field of flowers again and again and you'll get a very similar ouput each time.
To get more divergence and variance, I'm experimenting with a models activations. Activations, the inner state of a LLM can be nudged in meaningful directions as it plans and paints. Can one model have more than one default, each one recognisable and chosen on purpose? And if it can, does that make it a better creative collaborator?
[Write-up coming soon]
My contributions
Design
Development
Research
The team
Surya
