LLMs that think: Demystifying Reasoning Models
summary
New LLMs claim to “think before replying” by producing a hidden reasoning monologue, calling tools like map lookups and calculators during that phase, then emitting a concise answer. They let developers set adjustable reasoning effort so low effort gives quick guesses (e.g., single-number drive time) while high effort runs more checks and tool calls to produce ranges (e.g., 15–20 hours) and catch mistakes. Reasoning is still the model talking to itself not a symbolic solver, and the output is an auditable reasoning trace that trades latency and cost for reliability.
AI-written summary · about
chapters (12)
- 0:00Reasoning models intro
- 1:06Distinguishing reasoning models
- 3:36Reasoning vs AI agents
- 4:38Chain of thought vs reasoning
- 5:43LLM model taxonomy
- 7:29Reasoning output example
- 9:55How reasoning models work
- 12:24Practical Lisbon to Paris
- 17:42Limitations of high-effort reasoning
- 18:45Reasoning vs non-reasoning models
- 20:30New scaling laws
- 22:38Q&A
related talks
27mExploring the Mechanisms of LLM Reasoning Capabilities
29mThe Evolution of LLMs: From Imitation to Reasoning
28mSymbolic Reasoning in telecom oriented LLMs
31mBeyond LLMs and Maturing RAG use cases
16mBeyond LLMs - The Next Frontier of Machine Reasoning
39mUnlocking reasoning in Large language models
32mWhat can we expect of LLMs as software engineers?
19mLLMs reasoning and agentic capabilities over time