The day in conversation for 2026-10-22

The discussion: Tom: I’m still stuck on how much weight we should give to the name prediction work, specifically that eBISG approach and how they're trying to estimate race probabilities using embeddings instead of just census data. Jane: I think that’s a massive claim because it addresses a fundamental bias in LLM outputs right at the input stage. If we can effectively mitigate those biases early on with text embeddings, it fundamentally changes what kind of "real-world complexity" the models are capable of handling later. It moves us from fixing hallucination downstream to fixing representation upstream. Lu: But isn’t that a niche application, Tom? Estimating race probabilities based on names—it's very specific. I’d argue the more impactful work right now is still the infrastructure stuff, like how BASC improves confidence calibration. If we can reliably calibrate LLM self-consistency, that makes everything else—whether it’s image generation or 3D reasoning—more trustworthy, regardless of the domain. Meng: Trustworthiness is key, Lu, but if the underlying