The answer is no, and it’s not close. Palm is great but it’s not human level at long text, much less tightly reasoned long text like philosophy.
Source: am googler, work with LLMs. I can write a much better philosophy paper than any existing LLM and I only took undergrad level philosophy.
In general I would say that the production of very fluent text provides the illusion of reasoning, and while looking at LLMs does find evidence of reasoning at some level, mostly the illusion is still much stronger than reality. Maybe transformers will get there but we’re still very far from superhuman cognition right now.
The question is perhaps not so much about LMs, but about the nature of philosophy. Is it really much (anything?) extra on top of language modelling done by humans (perhaps quite advanced, beyond the abilities of the current LMs, but I have little doubt that in a few years with some incremental algorithm improvements and fine-tuning on the right kinds of text LMs will clear this bar), plus some creativity (covered in the post), plus intuitive moral reasoning?
Regarding reasoning, I also disagree but don’t want to explain why (throwing around capabilities ideas publicly).
The answer is no, and it’s not close. Palm is great but it’s not human level at long text, much less tightly reasoned long text like philosophy.
Source: am googler, work with LLMs. I can write a much better philosophy paper than any existing LLM and I only took undergrad level philosophy.
In general I would say that the production of very fluent text provides the illusion of reasoning, and while looking at LLMs does find evidence of reasoning at some level, mostly the illusion is still much stronger than reality. Maybe transformers will get there but we’re still very far from superhuman cognition right now.
The question is perhaps not so much about LMs, but about the nature of philosophy. Is it really much (anything?) extra on top of language modelling done by humans (perhaps quite advanced, beyond the abilities of the current LMs, but I have little doubt that in a few years with some incremental algorithm improvements and fine-tuning on the right kinds of text LMs will clear this bar), plus some creativity (covered in the post), plus intuitive moral reasoning?
Regarding reasoning, I also disagree but don’t want to explain why (throwing around capabilities ideas publicly).