Discussion about this post

User's avatar
John C's avatar
4mEdited

One nitpick about Noah's repeated example... LLMs generally can't do arithmetic any better than humans can do it in their heads.

The reason... LLMs learn hierarchical correlations between tokens that are present in large datasets like natural language, or structures in real world images. That means that if I try to teach an LLM the old white pages, or a long list of arithmetic data, there aren't really good patterns in them for it to latch on to. Modern LLMs are not trained to 'memorize' random data.

When you ask a chatbot an arithmetic problem, it punches it into a calculator and reads you the answer. The LLM 'helper' code that does that has improved in 4 years.

Conversely, since MATH is a complex hierarchical system of interacting symbols... LLMs are very good at THAT.

rahul razdan's avatar

Interesting perspective. two follow-up thoughts.

1) Intelligence: It is not clear that this is a generic thing. AI is a new tool which is good at a specific sort of intelligence. Whenever there is a new tool, there is a quick harvesting of solutions and it appears that is what we are seeing right now.

2) AI Safety: The key interface is virtual to physical. Regulations trying to "slow" down AI are not likely to work. However, governance of physical objects is well understood and needs to be extended to AI.

Ready for more?