Why LLMs Can’t Solve Complex Planning Problems

Описание: LLMs and optimization are not the same thing (planning, scheduling, NP-hard search, and why ChatGPT-style models fail without solvers like CP-SAT, CPLEX, or MiniZinc).

One of the biggest misconceptions in AI right now is that Large Language Models (ChatGPT, Claude, Gemini) can solve real optimization and planning problems. They can talk about solutions. That’s not the same as finding one.

Planning problems (employee rostering, routing, timetabling, job-shop scheduling) live in a different universe from next-token prediction. They are NP-hard, combinatorial, and require search and constraints, not linguistic guesswork.

You’ll learn:
→ Why LLMs fail at planning tasks (and often confidently return invalid schedules)
→ Why “reasoning models” still collapse under combinatorial explosion
→ What tools are actually built for optimization (mathematical programming, constraint solvers, meta-heuristics)
→ Where CPLEX, MiniZinc, Timefold (and similar tools) fit in the real stack
→ How to build hybrid systems where the LLM structures the problem and the solver does the heavy lifting

This is for software engineers, ML engineers, and AI builders who want to ship systems that work in production (and avoid the hype trap of using language models as generic problem solvers).

👉 Subscribe for more:    / @cesarsotovalero  

⏰ TIMESTAMPS
00:00 Intro
01:50 What a planning problem really is
04:29 What traditional optimization does differently
06:45 Why LLMs break down on planning tasks
13:31 The Hybrid Approach (LLM + Solver)
15:25 When to use which tool

📺 RESOURCES
→ Related video (1):    • How AI Is Killing the Ad-Funded Internet’s...  
→ Related video (2):    • How GPTs are Shaping our Perception of Truth  
→ Related video (3):    • The AI Secret: Why Predictive Machine Lear...  

#ArtificialIntelligence #LLM #Optimization #OperationsResearch #ConstraintProgramming #Scheduling #SystemDesign #SoftwareEngineering