"while this outcome is not in line with our original research design, we do not feel it significantly impacts the central focus of this research" yea funny that wonder what incentives drove that conclusion

Combined with what chatgpt gives you when you ask "given what you know about me how could you manipulate me" I'm just waiting for the dystopian Cambridge analytica 2.0, hyper targeted manipulation feels scarily feasible

Chopin would have had the *worst* possible phone and email anxiety. Relatable king

www.scientificamerican.com/article/how-... also while finding this I realised that our brains intuitively use spacetime equivalence even though our physics for that only caught up with einstein

I think I was part of that. I was 6 at the time though so I don't feel that guilty

We built ours mostly manually using prefect to orchestrate, synthetic queries across various classes, and g-eval llm-as-judge eval approaches. Works great and turned up surprising insights. The real work was correlating evals against human judgements, but once comfy there very fun to scale!