The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Is It Time to Rethink LLM Pre-Training? with Aditi Raghunathan - #747
159 snips
Sep 16, 2025 In this discussion, Aditi Raghunathan, an assistant professor at Carnegie Mellon University, tackles the limitations of large language models (LLMs). She presents insights from her award-winning paper on enhancing creativity beyond next-token prediction. Aditi introduces the innovative 'Roll the dice' method to foster randomness and 'Look before you leap' for deeper thought processes. The conversation also covers the paradox of 'catastrophic overtraining' and her pursuit of more controllable models through concepts like 'memorization sinks.' Her research aims to reshape our understanding of AI adaptability.
AI Snips
Chapters
Transcript
Episode notes
Unlearning Fails Without Architectural Bias
- Unlearning via post-hoc fine-tuning or neuron removal often fails because training doesn't naturally disentangle facts.
- Aditi shows there's no guarantee knowledge localizes to specific neurons under standard training biases.
Use Memorization Sinks For Targeted Unlearning
- Design models to isolate document-specific information into designated 'memorization' neurons during pretraining to enable targeted removal.
- At test time, drop those memorization neurons to erase document-specific data while preserving shared knowledge.
Random Subspaces Make Sinks Scalable
- Memorization sinks use random orthogonal subspaces so many document-specific slots fit without huge model bloat.
- Moderate increases in model size plus orthogonal subspaces encourage disentanglement empirically.