Reasoning Models (o1, DeepSeek R1)
Until recently, models just predicted the next word instantly. Reasoning models (like OpenAI's o1 or DeepSeek R1) pause to "think" before answering. They generate a hidden chain of thought—breaking the problem down, checking their own work, and fixing mistakes—before outputting the final answer.
This shifts compute from training to inference. Instead of needing a massively smarter base model, you just give the model more time to think. It's the AI equivalent of System 1 (fast, instinctive) vs System 2 (slow, deliberate) thinking.
Ask a standard model (like GPT-4o or Claude 3.5 Sonnet) a tricky logic puzzle. Then ask a reasoning model (o1 or R1) the exact same puzzle. Notice how the reasoning model outputs a "thought process" block that takes 10-20 seconds before giving the right answer.
What actually makes a "reasoning model" like o1 or DeepSeek R1 different from a standard model?