Everyone tries to mimic the real-time ChatGPT experience, but there's a much more reliable way to build with LLMs: **Select use-cases where users DO NOT expect an immediate response!** When model outputs aren't used immediately, humans can manually verify them. As confidence in various scenarios grows, you can automate more and verify manually less. Here are a few examples: - E-mail bots - Support assistants with Zendesk UX instead of chat UX - Coding assistants that create Github pull requests in the background Building this way is so powerful because: 1. It's essentially a solution to AI hallucinations 2. You can iterate your product over time without losing user trust 3. You can add guardrails to flag problematic outputs and redirect to human review. p.s. with OpenAI's batch API (https://lnkd.in/dvzuc6Cy) you get 50% cost reduction for offline LLM inference and higher rate limits :)
Juan Diego Balbi Yves Fogel check this out ! What we have been discussing about offline ai agents for backoffice tasks .
CEO Canonical AI. YC Alum. Tule founder.
1moMy favorite part of your post is it’s representation in sql 🤣