AI Apocalypse Survival Kit: coding agents & Tiny Language Models
summary
A work-in-progress system runs coding agents on local tiny language models (about 4B–9B parameters) using Docker Model Runner and local embeddings to prioritize offline resilience over cloud reliability. It demonstrates concrete limits (e.g., a 4B model took 5m21s to generate a Go “Hello World” when large context and system prompts inflated memory, and a Mac reached ~31 GB and froze), and ships agents that scaffold web apps, create Git issues via Forgero, and execute inside sandboxed microVMs. Practical strategies discussed include dynamic system prompts, RAG packing, trimmed history, and assigning one agent per task; the project is open source on Codeberg as II-apocalypse-survival-Kit.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 6:05Cognitive decline and AI
- 11:27Tiny Language Models
- 18:19Local coding agents
- 25:02Demo: documentalist agent
- 29:14Demo: web carpenter agent
- 33:49Demo: GitLab DevOps agent
- 39:49Demo: Rust agent in sandbox
- 44:43Conclusion
related talks
32mProgramming language evolution in the AI era
48mAI Coding Agents and How to Code Them
02h 48mCompose & Dragons: The Role-Playing Game of Agents Fed on Tiny Language Models
31mAI Coding Agents Don't Need Bigger Models. They Need Better Tools.
50mLet's build an AI agent
40mAI Coding Agents and how to code them
24mLightning Talk: Building the AIverse: Standardizing Autonomous AI Agents for a Decent...
49mBuilding a coding agent from scratch