The 100,000 Whys of AI: Exploring Deep Questions

In a substack post, lcamtuf (a well-known computer security researcher) explores the nature of AI by asking 'why' repeatedly, similar to a child's curiosity. The post digs into questions like: Why do LLMs produce coherent text? Why do we attribute understanding to them? Why might alignment be hard? He touches on scaling laws, emergent abilities, and the difficulty of interpreting internal representations. lcamtuf argues that our anthropomorphic tendencies lead us to overestimate AI's capabilities. He also discusses the 'replication crisis' in AI research, where published results often fail to reproduce. The post is a thoughtful meditation on the limits of current AI.