Personalization, Data, and the Chaos of Fine-Tuning with Fred Sala (UW-Madison / Snorkel AI)

Fred Sala of UW-Madison and Snorkel AI discusses how to make large language models genuinely personal. He explains why prompting preferences does not scale, why even LoRA adapters may be too costly per user, and why activation-steering methods such as ReFT could be a practical middle ground. The episode also covers self-distillation for continual learning, the difficulty of building realistic personas for evaluation, the continuing importance of weak supervision, and why data remains undervalued relative to architecture.