3 ms·
I'm using LLMs at my day job and claude v2 getting close to gpt-4's intelligence. It does use a different prompting style and has a huge 100,000 context window
by foxhop 3y ago
I'm using LLMs at my day job and claude v2 getting close to gpt-4's intelligence. It does use a different prompting style and has a huge 100,000 context window so it expects multishot prompting to train the LLM just-in-time.
- lambdaba 3y agoI've recently subscribed too. The huge context window is a game-changer and I can't wait for API access. Unfortunately it hallucinates a lot. It's amazing for extracting information out of large documents, altough doing so made me realize how limited current models are in terms of detecting "subtleties" that would be obvious to a human.
- lambdaba 3y agoBTW what do you mean by "different prompting style"?
- foxhop 3y agohttps://docs.anthropic.com/claude/docs/constructing-a-prompt https://docs.anthropic.com/claude/docs/constructing-a-prompt
- foxhop 3y agoHuman: I have a message and I need to categorize it into the right bucket. Can you help? General Inquiry Feedback & Suggestions Technical Issues Billing & Payments Appointments & Scheduling Others return only JSON like {"category": "Technical Issues"} Assistant: Sure using just JSON! Human: The app freezes every time I launch it. Assistant: {"category": "Technical Issues"} Human: I think your platform could use a night mode feature. Assistant: {"category": "Feedback & Suggestions"} Human: How do I reschedule my consultation? Assistant: {"category": "Appointments & Scheduling"} Human: I don't like the new open layout. Assistant: {"category": "Feedback & Suggestions"}