AI Workflow & Vibe Researching

The presentation argues that for modern physics research, the key question is no longer whether to use AI, but how to use it effectively. It distinguishes two major roles: NN-based AI for principles (neural-network fitting, neural quantum states, DeepH) and LLM-based AI for workflow (planning, memory, tools, APIs, MCP, agent loops).

Core AI concepts

The talk introduces foundational AI concepts relevant to research:

• LLMs, transformers, neural networks — the architecture stack
• Tokens & context windows — the currency and memory of LLM interactions
• RAG (Retrieval-Augmented Generation) — grounding model responses in external knowledge
• Prompts & system prompts — how to instruct and constrain the model
• Tools, MCP (Model Context Protocol), APIs — connecting models to the world
• Agents & skills — composing capabilities into autonomous workflows
• Headless mode — running AI without a GUI for automation

Effective AI use depends on understanding how temporary context, saved memory, retrieval, and tooling interact.

Evolution of human–AI collaboration

The workflow shifts from “user directly writes code” toward “user manages agents that work across a codebase and multiple tasks.”

Proposed research-project structure

A practical layout for AI-assisted research:

• Main workspace — global state and configuration
• Task folders — active execution environments
• Recycled / archive folders — failed or deprecated work
• Dashboard files (main_dashboard.md, task_XX_dashboard.md) — track progress, results, and decisions
• System prompts (CLAUDE.md, setup prompts) — define project rules and workflow expectations

Core workflow principle

Break one difficult task into several easier steps.

LLM performance drops when a prompt is too broad or vague, but improves when tasks are decomposed and prompts are precise and concise. The research-coding spectrum runs from precise coding, to LLM workflow management, to vague prompting — the best results come from combining code-level precision with LLM-assisted iteration.

Toward further automation

Agents can implement tasks, check plans, evaluate results, run massive tests, and compare outputs. A possible research automation loop includes:

1. Browse arXiv for new papers
2. Grade and filter papers
3. Share results through Discord
4. Let the user approve plans and results before moving to the next task

Takeaway

“Vibe researching” is a structured approach to using LLMs as research workflow partners: organize memory, define prompts and rules, decompose tasks, use dashboards, test multiple agents, and keep the human in charge of judgment and approval.