On the principles and practice of interaction design, visualization and interpretability for ML systems, and how to understand and explain language models.
I really enjoyed this podcast.
Martin Wattenberg brings a unique perspective to all the topics discussed.
I appreciated his nuanced conception of language and the way meanings of words are grounded in representations of worlds.
I liked his additions to the current debate about methods of interaction with LLM. I have been a strong critic of chatbots, but he is right, text is still a very powerful medium. Perhaps the next step in interaction is finding new ways to manipulate text.
I also really like the moment when you very consciously spoke about AI genre text as "content that looks like fiction" vs. "content that looks like non-fiction." This resonated with an intuition I have had for a while now about AI and intentionality.
Some critics of AI continue to insist that it "lies" or "produces fiction." I personally think these statements flatten the complexity of an AI locution, and lead to unproductive lines of thinking about the larger cultural moment we live in. More on that later!!!