Seonglae ChoGenerative Agent
How can we scale this work from sequential to parallel is going to be the next jump
- planning AI scheming
- memory
- tool interface
Specifically
- Tool learning
- System Calls
- Socket
- Headless Browser access to pretend human
- CLI access
- Curl
- GET - Search
- POST - Communication
- Permanent Storage like vector database
- Context Selection based on vector similarity
- Long-term memory by weight update
- Vanishing to prevent manage capacity based on LLM decision or dynamic sparsity
- Self-replication, Self-restart
- Self-initialization prompt, Self-editing
- Asynchronous communication port
- Interrupt
- …
Text to Task Generative Model
- Photoshop like AI software
AI Agent Notion
AI Agent Usages
Tips (AI Service, AI Agent)
- Split into micro-agents with single roles. (Model-based agents still don't perform well due to insufficient data accumulation)
- Keep only frequently used tools. (terminal, Language Server Protocol)
AI Agency
Google whitepaper
Workflow orchestration (Prompting Optimization)
Workflow-based vs. Model-based
Agent Software
- Parity: Agents should be able to perform any action through tools that is possible through the UI.
- Granularity: Tools should be atomic primitives without judgment. The agent makes the decisions.
- Composability: New features are composed through prompts alone without code changes.
- Emergent Capability: Agents should be able to solve unplanned requests by composing available tools.
- Improvement over TimePerformance: improves without deployment through context accumulation and prompt refinement alone.
