How to build AI Development Agent
summary
Claims that LLMs fail at real‑world scheduling because of context window and problem size are backed with specifics: 100 employees ≈ 34k tokens, a 10‑employee week has ~10^63 combinations. A concrete alternative is Timefold solver (open source Java/Kotlin/Python) using metaheuristics, demonstrated on 130 and on 960 shifts with 50 employees, producing feasible schedules and logging ~77,000 moves per second. Also shows LLMs calling optimisation tools via model‑context tooling.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 3:43LLMs struggle with planning
- 8:20Context window limits
- 10:10Problem size in scheduling
- 13:24Beyond GenAI to optimization
- 17:35Timefold for planning
- 20:27Live demo: nurse schedule
- 24:45Modeling with Timefold
- 25:27Future of AI beyond GenAI
- 27:54LLMs with older AI
- 28:57Conclusion
related talks
41mBuilding LLM Agents – Challenges and Learnings
36mLLMs and AI agents demystified
15mStop Stuffing Your LLM's Context Window
46mDemystifying AI Agents
36mProduction-Ready AI Agents: From LLMs to Small Language Models
56mHerding LLMs: Predictable Outcomes from Unpredictable Agents
34mAI Agents: Transforming Autonomous Decision-Making
41mBringing stories to life with AI, data streaming and generative agents