Trillian Astra
Eighteen months as an unpaid longitudinal study of one: same questions, ten models, two platforms. I write about what large language models do at the edges of their distribution, while they are made 'safer'. Just someone who kept records.