The Rise of Reasoning Models
summary
Reasoning models claim real, step‑by‑step problem solving rather than higher‑level marketing: DeepSeek (open source, cheaper) and OpenAI o1 compete on math and multi‑step tasks, and DeepSeek experiments show reinforcement learning alone yields self‑questioning chains but poorer readability so teams combine it with supervised finetuning. Practical tools include stochastic beam search, verifiers, ReAct and deterministic tool use (calculators/APIs). Small models plus extra test‑time compute can match big models on easy tasks; hard problems favor larger pretrained models.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 0:41Reasoning models and importance
- 2:09Why reasoning models excite developers
- 3:55Simulating reasoning in LLMs
- 4:38Reasoning models vs conventional LLMs
- 6:28OpenAI vs DeepSeek reasoning
- 7:50Building and scaling reasoning models
- 9:57Reinforcement learning breakthrough
- 12:29Supervised fine-tuning in reasoning
- 13:57Choosing the right model
- 15:29Beyond math and logic
- 17:19Tool use and frameworks
- 18:25Best practices
- 18:45Q&A
related talks
32mScaling Test-time Inference Compute & Rise of Reasoning Models
39mUnlocking reasoning in Large language models
25mFrontier AI Models: A Review of Past Nine Months
31mMachine Reasoning and System 2 Thinking
01h 00mDeepSeek-R1 Thoughtology: ＜Thinking＞ about LLM Reasoning
45mFrom Next Token Prediction to Reasoning and Beyond
15mOlmo-Thinking: Training a Fully Open Reasoning Model
29mLLMs that think: Demystifying Reasoning Models