<antirez>
6 hours ago
- Critics who previously dismissed LLMs as fundamentally flawed now claim that reasoning models like OpenAI o1 and DeepSeek R1 are not 'just LLMs,' but this is false.
- DeepSeek R1 is a pure decoder-only autoregressive model using the same next-token prediction criticized earlier, with no explicit symbolic reasoning.
- R1 Zero achieved strong reasoning capabilities through reinforcement learning and chain-of-thought generation without any supervised fine-tuning.
- The S1 paper shows that as few as 1,000 examples can enable complex reasoning, indicating that pre-training already encodes the necessary representations.
- Reasoning models are fundamentally LLMs, and those who declared LLMs a dead end were wrong; attempts to rewrite history are unacceptable.