flâneur

What's the difference -- (physics of) AI, physics, math and interpretability | Ziming Liu

kindxiaoming.github.io · 1,460 words · saved by 1 readers

In a previous blog post, I discussed how to conduct Physics of AI research. Some friends later asked: Can AI and physics really be fully analogous? And what is the difference between Physics of AI and interpretability? This article will discuss two points: First, what’s the difference between physics and (physics of) AI? If we count from the time of Newton, physics has taken about 400 years to develop to its current state. Compared with the development of AI, this is extremely slow. This slowness comes from both practical constraints and philosophical constraints. Practical constraints — physics research has strong directionality Both directions are constrained by experimental and observational bottlenecks. It takes a long time to build microscopes and telescopes in the physical world. Philosophical constraints We do not actually know how the “creator” (physical laws) truly runs the universe. We can only infer laws from phenomena. All models are wrong, but some are useful. Even the mos

In a previous blog post , I discussed how to conduct Physics of AI research. Some friends later asked: Can AI and physics really be fully analogous? And what is the difference between Physics of AI and interpretability? This article will discuss two points: AI and physics are not fully analogous , but that does not prevent us from borrowing methodologies from physics. In fact, Physics of AI is technically a simpler game than physics . Its main obstacles lie in publication culture (as discussed in a previous blog post ), not in the intrinsic difficulty of the subject. (As I define) Physics of A

related reading