In the realms of machine intelligence, I’m starting to question the applicability of ‘Thinking Fast and Slow’ concepts (System 1 & 2) in AI. Unlike human cognition, the time gap between AI’s direct inference (System 1) and multihop reasoning (System 2) seems rather small? 🤔
@WilliamWangNLP I think it's less about a time gap and more about a different mechanism: system 1 is autoregressive LLM inference, system 2 is when you pop out of that and do something else (e.g., tool use, other orchestration, etc.).
@WilliamWangNLP @nlpmattg Another aspect of System 2 is “checking your work before committing”, which intuitively feels more powerful than the reactive System 1.
@qi2peng2 @WilliamWangNLP @nlpmattg when is system 1 preferred over system 2 tho? or system 2 is always the way to go (ignoring cost ofc)