3 ms·
LLMs don't have any clue what the concept "research the question before answering" means. Maybe if you asked it a series of leading questions to get some of the
by daveguy 3y ago
LLMs don't have any clue what the concept "research the question before answering" means. Maybe if you asked it a series of leading questions to get some of the "research" in the context window first. Then you might as well just search. Otherwise it'll just confidently sound like it researched without any bearing on the answer.
LLMs are sloppy, and we shouldn't have to do "sociology" experiments to find out whether or not an LLM might give the right answer slightly more often if we trick it into adding more context to the answer.
- Workaccount2 3y agoWhile something like GPT4 cannot "do more research before answering a question", it is well established that telling an LLM it's an expert in something and asking it to spell out the steps in it's decision making process does make it more accurate in it's responses.
- daveguy 3y agoThat is allowing the LLM to place minor steps into its own context window where a more difficult result is more likely to follow as the next series of words. That is using our intelligence of how answers are determined to manipulate the output. It doesn't say anything about the "intelligence" of an LLM. Telling it that it's an expert does not have it somehow change the rote algorithm or have it "pretend" anything. It's just not that complicated.
- Closi 3y agoGPT-4 absolutely can do more research. Adding this to a prompt will encourage GPT-4 to do web searches in the background to augment its answer.
- Closi 3y ago> LLMs don't have any clue what the concept "research the question before answering" means. GPT-4 in ChatGPT will absolutely do research if you ask it (in the practical sense that if you say “research first” like in my original post, it will search the web and interpret/synthesise the results, and then use the findings to inform its answer).