GPT-4 vs Starcraft II – Strategic Decision Making using Large Language Models
summary
Demo shows GPT-4 controlling StarCraft II through a Python API, producing 20-step build orders and switching between defend or attack strategies, with attack phases limited to five units and real-time games taking 30–40 minutes (accelerated mode pauses for reasoning). Prompt engineering uses chain-of-thought summaries; function-calling and simple agents also build a multi-page website via file functions. Results are mixed: some games lose, tensorboard traces resource/score gains, and the website copilot mostly succeeds but occasionally times out or produces brittle updates.
AI-written summary · about
chapters (10)
- 0:00Introduction and background
- 2:23StarCraft AI evolution
- 5:49GPT-4 integration architecture
- 9:23Demo and performance analysis
- 13:23Chain of thought reasoning
- 18:35Strategic reasoning and agents
- 23:41Agents and SQL interaction
- 29:18Function calling protocols
- 32:24Vibe project management demo
- 39:40AI safety and alignment
related talks
59mGPT-4 vs Starcraft II – Strategic Decision Making using Large Language Models
45mBringing the Power of AI to Your Application
45mHow to use GPT in your projects: A Case Study
46mConversations with ChatGPT: illusion or reality?
01h 28mFuture of Software Programming with AI Bots at Hand
56mSeparating fact from fiction in a world of AI fairytales
28mGPT or how I learned to stop worrying and love the generative AI
01h 01mKeynote: Transformers: The Rise of ChatGPT