This is a link post. New Anthropic research (tweet, blog post, paper): We investigate whether large language models can introspect on their internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model's activations, and measuring the influence of these manipulations on the model's self-reported states. We f...
All content for LessWrong (Curated & Popular) is the property of LessWrong and is served directly from their servers
with no modification, redirects, or rehosting. The podcast is not affiliated with or endorsed by Podjoint in any way.
This is a link post. New Anthropic research (tweet, blog post, paper): We investigate whether large language models can introspect on their internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model's activations, and measuring the influence of these manipulations on the model's self-reported states. We f...
TL;DR: AI progress and the recognition of associated risks are painful to think about. This cognitive dissonance acts as fertile ground in the memetic landscape, a high-energy state that will be exploited by novel ideologies. We can anticipate cultural evolution will find viable successionist ideologies: memeplexes that resolve this tension by framing the replacement of humanity by AI not as a catastrophe, but as some combination of desirable, heroic, or inevitable outcome. This post mostly ...
LessWrong (Curated & Popular)
This is a link post. New Anthropic research (tweet, blog post, paper): We investigate whether large language models can introspect on their internal states. It is difficult to answer this question through conversation alone, as genuine introspection cannot be distinguished from confabulations. Here, we address this challenge by injecting representations of known concepts into a model's activations, and measuring the influence of these manipulations on the model's self-reported states. We f...
