Legal Copilot for Work Code and Documents
summary
Talk explains how Copilot is used legally on closed, NDA-bound projects at VKontakte, describing practical constraints and why external AI services pose risks for proprietary code. It reviews alternative approaches (on-premises or restricted tools), the speaker’s personal workflow with AI assistants for frontend development, and specific limitations that prevent generating the presented text by a neural model.
AI-written summary · about
chapters (30)
- 0:00Introduction
- 1:06Speaker background
- 3:57Search vs neural networks
- 5:00ChatGPT in dev
- 7:10Risks of external models
- 8:35Local model solution
- 10:20Choosing local model
- 13:11Transformer basics
- 14:56Quantization impact
- 15:58Model size & hardware
- 17:02Text & instruction models
- 18:08Hardware recs
- 19:12Editor setup
- 20:17CodeGeeks 4 9B demo
- 22:24Large model issues
- 23:27Corporate neural nets
- 24:34Cadestral 22B demo
- 27:05Local code limits
- 28:11Text processing
- 29:18Gemma 3 4B features
- 30:23Project & doc analysis
- 33:09Manager use cases
- 34:53Gemma 3 4B for notes
- 35:15Local model tasks
- 36:37Prompting principles
- 37:19Public model safety
- 38:45Corporate gateways
- 39:28Need for neural nets
- 40:10Personal experience
- 40:55Q&A
related talks
32mBeyond the Hype: A Pragmatic Look at How Developers Actually Use AI Tools
36mME, MYSELF AND AI
34mUsing AI to build an AI Java app
45mThe era of the AI Copilot
46mAI Integration for Devs: ChatGPT and Copilot in Your Projects
14mon Collaboration, Copilot & the Future of Work
25mImproving developer happiness with AI
09mCopilot App Deep Dive: Intelligence, Context & Agents