3 ms·
The article has a flavor of trying to find a role for good old fashioned AI, with planning and search, in this new world. It points out that to some extent larg
by Animats 3y ago
The article has a flavor of trying to find a role for good old fashioned AI, with planning and search, in this new world. It points out that to some extent large language models can plan, but only for things where the steps required are roughly described on the Internet. It's basically splicing together bits of known sequences of steps. That can handle many cases. As usual, the problem is lack of evaluation - totally bogus plans can emerge unnoticed. This is the lack of a model problem.
The big lack remains manipulation in unstructured situations. Although that Google project mentioned a few days ago has some promise. They train on simple manipulation tasks with a force feedback teleoperator, and then use an LLM for the higher level task sequencing. But the simple manipulation tasks remain rather simple. Too simple to, say, assemble flat-pack furniture, let alone repair a car. LLMs just don't seem to be relevant to such non-language tasks.