Andrew Dai discusses vision models on The Information Bottleneck
The episode announcement says Elorian AI's co-founder and CEO sees much of reasoning as visual—and that his company is building for it.
TLDR
The Information Bottleneck's episode announcement describes a conversation with Andrew Dai, co-founder and CEO of Elorian AI. It says the result behind a 2015 paper he co-wrote on language-model pre-training and fine-tuning came from a bug: he forgot to update a checkpoint directory and fine-tuned a model by accident. The announcement also highlights discussions about why Google fell behind early in the large language model race, why current models can't count objects in photos, and Dai's view that much of reasoning is visual.
Combined views
828
2 Sources, first seen 1h ago
