🐬 AI-by-hand: Multi-head Latent Attention, RoPE, and MoE in Deepseek.
-
Updated
Feb 17, 2025
🐬 AI-by-hand: Multi-head Latent Attention, RoPE, and MoE in Deepseek.
Pure Python implementation of a ReAct agent loop, vector routing, memory tracking, and cost analytics from first principles. Mirrors a 6-page "AI by Hand" worksheet with handwritten scans, programmatically verifying sequential reasoning tokens, tool matching via explicit cosine similarity, and context explosion costs without frameworks.
To associate your repository with the ai-by-hand topic, visit your repo's landing page and select "manage topics."