The next LLM Agents will be Python and not JSON-based! 🐍 CodeAct proposes a new framework to generate executable Python code instead of JSON for more challenging controll flows. It is not only more accurate but also reduces the number of actions for complex tasks. 🤯 LLM agents are typically prompted to produce actions by generating JSON, which limits and restricts flexibility (e.g., using multiple tools or loops). CodeAct uses executable Python code to consolidate LLM agents' actions into a unified “action space”. 💡 Implementation 1️⃣ Provide a system prompt that explains the CodeAct framework, including a few shot examples and the tools (Python function signature or docstring). 2️⃣ User prompts LLM with request or task 3️⃣ LLM generates a Python code as the action, based on the provided “tools” 4️⃣ Execute the code with a Python interpreter. (be cautious here) 5️⃣ Captures the output or error messages and provides this information to the LLM 6️⃣ LLM then decides to modify its previous code or generate the response for the user Steps 3-6 are repeated until the LLM decides to submit a final response or reaches a predefined number of interactions. Insights 🚀 Outperforms existing method with up to 20% higher success rates ⚡ Requires up to 30% fewer actions to solve tasks 🤖 GPT-3.5 CodeAct matches GPT-4 JSON performance 🛠️ Use of error messages allows self-debugging and improves performance 📚 Created CodeActInstruct dataset, consisting of 7,000+ multi-turn interactions 💬 Including datasets like OpenOrca and ShareGPT to keep general language understanding 🧩 Use of Python code allows complex control flows with loops or conditions Paper: huggingface.co/papers/2402.0… Dataset: huggingface.co/datasets/xing…

We built a web app that lets you fly a spaceship through a 3D constellation of music - powered by our Lyria RealTime model. 🎶 Space DJ is an interactive visualization where every star represents a different music genre. As you explore, your path is translated into prompts for the API, creating a continuously evolving soundtrack. ↓