Three subtle examples of data leakage

Adam Optimizer Causes Privileged Basis in Transformer LM Residual Stream

Contra papers claiming superhuman AI forecasting

OpenAI shows 'Strawberry' to feds, races to launch it

We don't know how bad most things are nor precisely how they're bad

Superbabies: Putting the pieces together

Claude 3.5 Sonnet Reproduces BIG-Bench Canary String

I would have shit in that alley, too

Transformers Represent Belief State Geometry in Their Residual Stream

Autism as the Kolmogorov complexity phenotype