Lightning talks - session 1
summary
Eight rapid demos and claims across tools papers and prototypes include concrete takeaways not fluff: Codex suggests buggy GitHub like code when training data includes bad repos, Llama models route translations through English intermediates from an EPFL June 8 2024 paper, an Anthropic notion of monosemantic boolean flags for interpretable features, a deployed real time soil moisture predictor using remote sensing plus ML that compares to ground truth, a Polars defensive checks package, a JupyterLite GitHub Actions deployment that actually builds.
AI-written summary · about
chapters (11)
- 0:00Opening
- 2:38Why large language models sometimes fail
- 8:04Soil moisture with ML
- 12:09Missing data risks
- 16:53STEM inspiration
- 20:52Defensive data analysis
- 23:21Data science in glass
- 31:15Open source science tools
- 33:44JupyterLite in 3 min
- 39:06Queues and philosophy
- 43:43AI and future work
