EHack

Project Description:

Authors (add links later)

• Alif
• Ryan
• Lucas
• Sylas

How to run:

Plan:

• We are going to use OpenAI Assistants API to create a virtual laboratory
• Connect it with a 3D interface with Needle Engine
• Automatically synthesize a team of agents to solve the problem
• Input problem, visualize the work of the virtual laboratory through the agents visualizing the chat bubbles and their work
• Have code interpreter agents that are running code, Principal Investigator agents to coordinate the activity

3D Interface:

• Needle Engine:
  • Basically allows us to do performant web rendering of 3D objects, built on top of three.js
  • We write in typescript, compiles to C#
• Have a programmatic interface for synthesis of scenes and agents working together
• Make it generic enough to work for many types of tasks
• Visualize all the backend events in the 3D interface

System Architecture for AI Swarm Planning:

• Shared task planning space

• Shared knowledge base
  • Shared file base for assistants
    • Need to upload all the files to the API (might be slow? Try to upload small things first.)
• Shared communication space
  • JSON message passing?
• Shared execution space
  • They will be able to run code, python packages
  • automatic chemistry research how to? (TODO: link python packages)
    • python outputs in a feedback loop
      • how to talk to other agents about progress?
• agents should have different roles and personalities (parameterized by OCEAN)
• the agents should be synthesized based on the problem at hand
  • an agent that synthesizes other agents (agent designer?)

Agent Designing Agent (ADA)

• This is a meta-agent that creates and manages other agents to tackle various sub-problems.

• Agent Synthesis:

  • ADA dynamically synthesizes other agents based on the specific problem-requirements.

• Task Analysis:

  • ADA breaks down the complex user task into tasks that individual agents, or groups of agents can solve.
  • Pseudoalgorithm:
    • Understand the problem.
    • Systematically reason about the problem.
    • Break down the problem into sub-problems.
    • Generates skill requirement lists for each sub-problem.
    • ADA will estimate the computational resources, time, data access required to solve each sub-problem.
    • ADA will generate (HIRE) Agent personalities that are best suited to solve each sub-problem, and will also define what they are supposed to be performant at, and generate metrics (KPIs) for managing the agents. (TODO: Agent Generation Function)
    • ADA will monitor performance of each agent, and re-synthesize (FIRE) agents as needed.

• Agent Customization:

  • Uses OpenAI Agents API (https://platform.openai.com/docs/assistants/overview) assistant = client.beta.assistants.create( name="Superalignment Researcher", instructions="You are an AI alignment researcher. Write and run code to align AI to human intentions.", tools=[{"type": "code_interpreter"}], model="gpt-4-1106-preview")

Running an agent simulation:

• 1. Input problem (input_text: str)
• 2. Decompose 1. into sub problems (sub_problems: List[str])
• 3. Synthesize agents to solve sub problems using an Agent Designing Agent (ADA) (agents: List[Agent])
• 4. The agents should communicate with each other, introducing themselves to each other and the user. (agent_introductions: List[str])
• 5. The agents should each have a "research journal" summarizing their insights into the task at hand.