LLMs can't optimize schedules, but AI can!
summary
Claim: LLMs fail at real scheduling because of context window limits and sheer combinatorial scale — a 10‑employee week can have 10^63 combinations — so token‑based next‑token sampling breaks. Practical solution uses mathematical optimization and metaheuristics: model shifts and constraints in Java/Kotlin/Python, run a solver like Timefold solver that generates an initial solution then iterates (demo shows ~77,000 moves per second), handles hundreds of shifts and tens of employees, and produces feasible, explainable schedules.
AI-written summary · about
chapters (12)
- 0:00Welcome
- 1:13LLMs aren’t the solution
- 4:05LLMs fail at planning
- 6:09Context window limits scheduling
- 9:46Problem size blocks LLMs
- 13:24Old-school AI for planning
- 17:35Timefold for optimization
- 20:27Demo: Timefold nurse schedule
- 24:45Modeling constraints in Timefold
- 25:27Future of AI beyond LLMs
- 27:54LLMs with older AI tools
- 28:57Conclusion
related talks
36mLLMs and AI agents demystified
33mDesigning LLM Applications for Massive Datasets
10mThe Impact of LLMs on the Circular Economy and Beyond
15mStop Stuffing Your LLM's Context Window
32mWhat can we expect of LLMs as software engineers?
01h 00mFrom LLM Theory to Practical Agentic Implementations
10mRethinking collaboration: LLMs, teams, and cognitive load
38mNobody Pages the LLM: Engineering Rigour for Vibe Coding