
AI at Cursor and LLM limitations

Scaling Slack and GitHub with Maude Lemaire
Topic 10 of 11
Published Nov 5, 2026 · Clip 1:02:22–1:13:00 (10:38)
Discussion of her Cursor work improving time to first token and her relationship with LLMs: AI struggles with performance investigation and instrumentation, generates useless mega-traces and misplaced sampling, and naively parallelizes rarely-used paths. She remains heavily hands-on—printing giant sequence diagrams, manually verifying everything—and uses mutation testing to force LLMs to write real edge-case tests.