DuckDB's growing support for embeddings, remote parquet reads, Python UDFs, and vector search makes it a surprisingly strong tool for modern similarity workflows.
I learned teaching techniques from IIT Madras, automated video highlights with OpusClip, and summarized DuckCon #6. I also found that DeepSeek R1 excels at text-to-CAD and gathered OpenAI’s latest prompting advice for reasoning models.
I explored DeepSeek R1 training, how AI models are absorbing app capabilities, and fixing Windows symlinks for Hugging Face. I also discovered DuckDB's built-in notebook UI, Gemini’s YouTube API, and Karpathy-inspired note-taking workflows.
I discovered DuckDB's ability to read JSON APIs and found an undocumented link for managing scheduled LinkedIn posts. I also wrote a JS snippet to export npm version history as Markdown and defined better HITL review workflows.
The newer OpenAI embedding models produce much lower cosine similarities than ada-002, so practical similarity thresholds need recalibration from around 85% to roughly 45%.
I investigated binary embeddings, empathic AI with Hume.ai, and audio splitting with Spleeter. I failed at building a Rust-based Parquet server and noted distinct corporate attendance patterns while hosting workshops for Gramener and Straive.