2025-04-18
Chain of Thought
- Essentially the model has been trained on how to think and it goes through an explicit reasoning process before giving its final answer.
- The tradeoff being made is:
- Vastly improved problem solving capabilities for queries where there are verifiably correct answers.
- 100x or more slower than non-reasoning models since the LLM is producing all those internal reasoning tokens. A “fast” query with a reasoning model is 10s before you get any output. A slow query is minutes.
- 5x more expensive than non-reasoning models.
- This tradeoff is tunable. The longer you let the model reason, the more capable it is.
- Important The reasoning models do NOT improve the capabilities for subjective tasks. OpenAI even observed that people tend to prefer the text of gpt-4o vs o1. The improved capabilities are just for when there are verifiably correct answers.
Hot Pot and Poke
Newfounland viewing 536 flats, 97% occupancy, hey tea and zara.
Hot Pot was disappointing. Poke was ok-ish - at least the salmon was good, chicken not great.
Drank and had fun. Luckily not too drank.