Epistemic vs. Aleatory Risk
A discussion unfolds on the divergence between epistemic and aleatory risks, highlighting how RLHF can disrupt the correlation between truth and model output. Insights emerge from experiments using uncensored models, revealing that pre-training may not be as crucial as previously thought. The conversation also touches on the intriguing possibility of training models from scratch to assess the importance of underlying architecture versus pre-trained capabilities.In this clip
From this podcast
Machine Learning Street Talk (MLST)
Daniel Franzen & Jan Disselhoff - ARC Prize 2024 winners
Related Questions