Is There A DNA Of AI? Finding Interpretability’s Backbone | David Bau (Northeastern University)

David Bau (Northeastern University) on whether there is a DNA inside cognition, and what scientific infrastructure interpretability would need to find it. A talk from the Cambridge AI Research Directions Workshop, convened by FAR.AI with the Cambridge Boston Alignment Initiative. Bau opens on the problem of scale. These systems are grown rather than designed, nobody fundamentally understands what happens inside them, and looking at billions of components produces a feeling that to understand anything you would have to understand everything. His answer comes from biochemistry. Through the 1930s and 1940s the field faced the same wall: a vast map of chemicals and reactions with no way to prioritize. What changed in the 1950s was the recognition that some molecules mattered far more than others, and that one of them kept the recipes for the rest. DNA is not completely predictive of what a cell does moment to moment, but it describes the cell's capabilities on the longest timescales, and recognizing that made the whole field tractable. Bau's research question is whether something with that character exists inside large-scale AI, and his advice to researchers is that the job is learning to prioritize rather than to catalogue. He then walks through one line of work shaped by that question: introspection, or whether a model can accurately report on its own thinking. The practical stakes are immediate. When a model is asked why it did something, as in the recent agent incidents, it is unclear whether the answer reflects the mechanism that actually drove the behavior or is a plausible explanation assembled afterwards. Chapters 0:00 Models are grown, not designed 0:48 What biochemistry looked like before DNA 2:05 How DNA made the field tractable 2:27 Is there a DNA inside cognition? 3:23 An answer to the feeling of hopelessness 3:58 A worked example: introspection in AI 4:32 Does a model know its own intent? 5:20 Why introspection was hard for academics to study 6:18 The breakthrough: training longer than needed 7:09 Comparing faithful and unfaithful introspection 8:11 What actually let biologists find DNA 8:23 The infrastructure they built first 9:36 Why science infrastructure differs from product infrastructure 10:05 Building a consortium for interpretable AI More AI safety research: https://far.ai FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.