Demo
- Camel agent can use video as input, generate marketing material (social media copy)
- Can also use Camel RAG to connect to various sources of info for a chatbot
- Can operate a laptop and a mobile together
Finding the scaling law of Agents
- Marvin Minsky defined agents as mindless processes, which can be joined in 'societies' to lead to true intelligence (The Society of Mind, 1986)
- One of the earlier agents was called WebGPT (OpenAI, 2021) used an agent to browse the web
- Lilian Weng write up (Lil'Log, 2023) outlined the idea of LLM as agent, with key differences: language as input, language as output, state and action are expressed as natural language, generalisability, ...
- Of course many aspects still missing: benchmarking, embodiment, ...
- LLM based communicative agents: CAMEL, generative agents, GPT-Bargaining, ...
- Role-playing framework (Camel, NeurIPS 2023)
- 📄 Li et al. (2023) CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society (arX⠶2303.17760v2 [cs⠶AI])
- Roles and communication patterns must be clarified in prompts to different agents (inception prompting)
- CAMEL agents: AI society with 50 assistant roles, 50 user roles and 10 tasks; code, 20 programming languages, 50 domains and 50 tasks
- One of the uses is agent-driven data generation
- Many challenges in cooperative role-playing: role flipping, the assistant repeats the question, flakey reply (assistant replies "I will do it"), infinite conversation
- Used by MosaicML MPT and Nous-Hermes
Q&A
- Q: Could you use a 'manager' agent?
-
A: This is an idea we proposed, of a 'critic in the loop' (critic agent), do critic tree search
-
Q: Is Critic Tree Search like MCTS?
- A: It's inspired by it but no complicated simulation.