you see this, wyd?

The AC is broken. It's been turbo freezing and I've not felt this focused in a while. Cold plunge health gurus are onto something

R to @connor_beem: The universe always delivers. I currently focusing on growth for the agency I'm building with @edrick_dch - following back.

R to @edrick_dch: Though I'm not into it now, I can understand how valuable this is. Thanks for the share. 🙌

R to @edrick_dch: Great use case of a startup leveraging their data to scale customer support. But not bulletproof

Just realized Framer is now using AI to answer user questions on their forums

R to @shawnsqvl: I gotchu 😎 lmk when you get here

R to @edrick_dch: hook me up with your papaya plug, i wil b there in oct :O

R to @edrick_dch: 🧠 Chat model: Compatible with the new chat message APIs (gpt-3.5, gpt-4, ...) 💬 Chat history: Context is passed to future iterations 💪 Multi-args tool: Can use tools which take in more than 1 argument (gmail, ...) These agents follow the same Thought-Action-Observation loop

👨‍🦲 Lost hair while searching for the differences between @LangChainAI's "react-description-" AgentTypes… Save your hairline any further trouble. Here it is visualized. Some more context below 👇

R to @alonsosilva: it's worth the wait, even though 1 day feels like a week in this space 😅

R to @edrick_dch: 😭

R to @edrick_dch: This is maybe the direction of agents we’re most excited about

R to @edrick_dch: about model selection for agents, I noticed gpt-3.5-turbo: painful to watch, but the spirit is there gpt-4: clean thoughts + execution, barely any errors anyone else? curious what/how people are using

Testing plan-and-execute agents using @LangChainAI The leverage these agents can provide, even rn, is actually surprising As models get better, businesses that won't have their own agents running 24/7 will just be making it harder for themselves

gm bangkok $0.50 freshly cut papaya is all you need

R to @edrick_dch: That's it! If you're looking for more Generative AI content, you can find me on YouTube: https://www.youtube.com/@edrickdch I'll also start posting to the newsletter: https://practical-ai-builder.beehiiv.com/subscribe

R to @edrick_dch: 15/ And finally... 🤖 Agents - Lots of potential to become a central piece of the LLM app architecture - Autonomous agents are still experimental and not reliable yet

R to @edrick_dch: 14/ 🏨 App hosting (everything other than the model) - Most use Vercel - End-to-end hosting for LLM apps: Steamship - Model + Python code hosting: Anyscale, Modal

R to @edrick_dch: 13/ 🧰 Operational tooling - Caching: Most use Redis to improve app resp time and cost - Log, track, eval LLM output: Weights & Biases (ported from trad ML), MLflow or PromptLayer and Helicone (purpose-built for LLM) - Eval LLM output: Guardrails - Detect prompt injection: Rebuff