### 💬 My Take on Codex-tools (LLM-based dev tools) - Similar to JavaScript → cool/strange/simple at the beginning - Evolution of tools/practices for defining context/prompts/LLMs --- > 3 months usage of passive use at SCB ### 1.1 Day-to-Day: - Generating tests - Explaining code - Simple refactoring - Code generation - Reading errors/extracting values --- ### 🔍 Observations - Slow generation - Settings disabled by admin (model, scopes) --- ## 2. Copilot J+History Jun 2021: GitHub (Microsoft) released Copilot based on Codex (OpenAI GPT-3) - Uses FIM (fill-in-the-middle) - new LLM approach for capuring context - GPT-3.5 Turbo - first success - Introduced system to block risky patterns/exploits --- ### 🔧 Settings 🔗 https://github.com/settings/copilot 1. Subscription: - Individual - Business - Enterprise (learns codebase, COBOL, scans PRs for the best practices) 2. Models: - GPT-4.0, D3 - Anthropic Claude 3.7 - Gemini 2.0 Flash --- ## 2.1 Internals > NLP, ML 1. Input Preprocess – prepare for model context and prompt (code related) → `processed_prompt` 2. LM Analysis – neural network analyzes `processed_prompt` (large body of text) 3. Generation – LM generates code/suggestion 4. Output Format – indent/highlight --- ## 3. Stats - **AMD 2023**: Specific HDL Language – generated code better aligned with standards. - **Shopify**: Out of 2,000 developers, 75% used Copilot, with 26% accepting suggestions. - **Accenture**: Out of 500 developers, 35% accepted suggestions. - **General Satisfaction**: 75% of users are satisfied with Copilot. Competitors: - Amazon CodeWhisperer - Cursor VSCode Fork - Claude Code (recent impressive results) --- ## 4. Features: 1. __Completion__ → VSCode (ghost text) 2. __Smart Actions__: - __Explain__ - __Fix__ - __GenTest__ - __GenDocs__ 3. Cmd+I → change/refactor 3.2 Cmd+I in __terminal__ 3.3 QuickChat 4. __Commits__ summary - auto-generated - __PR__ summary → GitHub interface disabled - GitHub VSCode_workspaces_workflow → cloud-hosted dev.env