I’m a non-specialist who’s interested in both the System 1/System 2 cognitive systems distinction from Kahneman et al. (particularly in relation to language learning in my case) and AI, so this may be a naïve question: is Anthropic’s J-space something like System 2 for an LLM? The fact that ablating J-space leaves fluent, automatic language generation largely intact but impairs multi-step reasoning struck me as a parallel to the System 1/System 2 distinction.
Anthropic’s paper also describes J-space as supporting “deliberate reasoning” alongside more automatic processing elsewhere in the network.
I’m curious whether Anthropic or anyone else is doing equivalent work on the System-1-like side: what does the cognitive/representational space underlying automatic processing in an AI actually look like?
I’d love to hear what people who know much more about this / in the industry think. I’m not at all in the industry, just someone exploring a personal hypothesis as to whether, for language learning, we need to train System one and System two differently (comprehensible input for language acquisition to train System one, traditional grammar and syntax for language learning to train System two.) The Anthropic article struck me as super interesting (following up a reference in an Insider interview on the Economist.)
Hi all, first post here.
I’m a non-specialist who’s interested in both the System 1/System 2 cognitive systems distinction from Kahneman et al. (particularly in relation to language learning in my case) and AI, so this may be a naïve question: is Anthropic’s J-space something like System 2 for an LLM? The fact that ablating J-space leaves fluent, automatic language generation largely intact but impairs multi-step reasoning struck me as a parallel to the System 1/System 2 distinction.
Anthropic’s paper also describes J-space as supporting “deliberate reasoning” alongside more automatic processing elsewhere in the network.
I’m curious whether Anthropic or anyone else is doing equivalent work on the System-1-like side: what does the cognitive/representational space underlying automatic processing in an AI actually look like?
I’d love to hear what people who know much more about this / in the industry think. I’m not at all in the industry, just someone exploring a personal hypothesis as to whether, for language learning, we need to train System one and System two differently (comprehensible input for language acquisition to train System one, traditional grammar and syntax for language learning to train System two.) The Anthropic article struck me as super interesting (following up a reference in an Insider interview on the Economist.)
Cheers, Chad