Reply to thread

Message

<blockquote data-quote="Dausuul" data-source="post: 9496811" data-attributes="member: 58197">I've spent the last couple weeks working on a project that uses an LLM to extract and classify data. It's been a fascinating process. &quot;Prompt engineering&quot; sounds pretentious as hell, but it's a real skill -- particularly when you need the bot to provide output in a form that can be parsed by a traditional computer program. Many chatbot quirks are really just human quirks reflected back at us. The way you phrase a question can have a big impact on the answer. If you load it down with too many fiddly conditions and rules, it starts to forget some. You get better results by forcing the bot to consider arguments for both sides. And providing context is essential, as is thinking carefully about exactly what you're trying to find out.One really &quot;non-human&quot; thing I've noticed is that you need to give it permission to say &quot;I don't know.&quot; If I ask it to explain how a quantum probability recalculator (some technobabble I just made up) works, it will spout off an answer... unless I add, &quot;If you don't know what that is, say UNKNOWN.&quot; Then it says &quot;UNKNOWN.&quot; Some humans will admit they don't know an answer, some will try to BS their way through, but very few BSers will stop when told to.</blockquote>

[QUOTE="Dausuul, post: 9496811, member: 58197"] I've spent the last couple weeks working on a project that uses an LLM to extract and classify data. It's been a fascinating process. "Prompt engineering" sounds pretentious as hell, but it's a real skill -- particularly when you need the bot to provide output in a form that can be parsed by a traditional computer program. Many chatbot quirks are really just human quirks reflected back at us. The way you phrase a question can have a big impact on the answer. If you load it down with too many fiddly conditions and rules, it starts to forget some. You get better results by forcing the bot to consider arguments for both sides. And providing context is essential, as is thinking carefully about exactly what you're trying to find out. One really "non-human" thing I've noticed is that you need to give it permission to say "I don't know." If I ask it to explain how a quantum probability recalculator (some technobabble I just made up) works, it will spout off an answer... unless I add, "If you don't know what that is, say UNKNOWN." Then it says "UNKNOWN." Some humans will admit they don't know an answer, some will try to BS their way through, but very few BSers will stop when told to. [/QUOTE]

Verification