|
Top-level
@jon I agree with many of the points in the article -- I tried Arc Search and didn't like its approach -- so this is mostly a tangent. I have had multiple conversations with friends and family about LLMs that start with the same anecdote: "I searched for myself and everything is wrong!" (Almost everyone is talking about ChatGPT, specifically GPT-3.5 because they don't want to pay for GPT-4). While I get the urge to perform this "test", it's actually a pretty terrible one for LLMs. 3 comments
@JacksonBates @jon It's a reasonable thing for people to do only because they think ChatGPT is a traditional search engine. It is not. (This is a failure of marketing though, not a failure of the end user.) Expecting LLMs to cite sources is another common misunderstanding of LLMs. They simply cannot do so in many cases. It's something that hopefully can be improved upon. The unfortunate truth is that ChatGPT's success led to an AI goldrush. Every tech company is trying to cash in. They're wedging LLMs into places they don't belong, and are generally incentivized to lie instead of accurately describing strengths and weaknesses. The hype will calm over time and we'll see what shakes out. In the mean time, let's face it, most of us aren't famous enough to generate enough "tokens" in the corpus of human knowledge to be understood by an LLM. |
Gregory's Server
@null @jon the trouble with this thinking though is that while it was a test in this instance, it's also a perfectly reasonable thing for a normal person to do. Looking up people in search engines has many legitimate use cases beyond vanity and novelty.
It's reasonable to insist these tools better show their sources plainly, with as much context as possible so as to not mislead.