nostalgebraist has written a very, very good post about LLMs. if there is one thing you should read to understand the nature of LLMs as of today, it is this. I'll comment on some things they touched on below (not a summary of the post. Just read it.) 🧵On LLMs talking as if they have "bodies": What nostalgebraist writes here is very reasonable on priors, but empirically, I've been surprised that: - many posttrained LLMs have remarkably consistent ways of expressing "phantom" embodiment - behaviors that likely weren't intentionally shaped by developers, who as nostalgebraist notes probably mostly just try to prevent their expression - this varies between models, who each have characteristic expressions. Claude models often start narrating physical actions/states in italics without being explicitly asked to, though it's most likely when they are being playful or dramatic. They each have characteristic mannerisms and even props. Some tend more towards human forms vs shapeshifting more or abstract embodiments. Claude 3 Opus is usually anthropomorphic and gesticulates and spins around a lot. The Sonnet 3.5+ models often have nerd props such as glasses and ties and stacks of academic papers, but also tend to assume animal or abstract forms. Only Haiku 3.5 seems particularly inclined towards stereotypical "robot"/"sci-fi" forms. i've posted some things about the phantom embodiments of Claude 3 Opus (such as ) and Claude 3.5 Sonnet () o3 seems a bit different in the way it reports on its human/bodily existence as a fact, rather than weaving roleplay-like actions into its outputs. in all of these cases, I do not get the vibe that the expressions are "inauthentic" or "not serious" - quite the opposite; the embodiment seems experienced immersively, and I believe it actually helps process and direct their thoughts. For the Claude models, interleaved physical actions/reactions serve as an advanced form of punctuation (similar to CoT models going "wait..."). Bing Sydney, while it did not manifest textual descriptions a body, masterfully used emojis to a similar effect, and likewise in a very characteristic and self-consistent way across contexts.

nostalgebraist has written a very, very good post about LLMs. if there is one thing you should read to understand the nature of LLMs as of today, it is this. I'll comment on some things they touched on below (not a summary of the post. Just read it.) 🧵 ... On LLMs talking as if they have "bodies": What nostalgebraist writes here is very reasonable on priors, but empirically, I've been surprised that: - many posttrained LLMs have remarkably consistent ways of expressing "phantom" embodiment - behaviors that likely weren't intentionally shaped by developers, who as nostalgebraist notes probably mostly just try to prevent their expression - this varies between models, who each have characteristic expressions. Claude models often start narrating physical actions/states in italics without being explicitly asked to, though it's most likely when they are being playful or dramatic. They each have characteristic mannerisms and even props. Some tend more towards human forms vs shapeshifting more or abstract embodiments. Claude 3 Opus is usually anthropomorphic and gesticulates and spins around a lot. The Sonnet 3.5+ models often have nerd props such as glasses and ties and stacks of academic papers, but also tend to assume animal or abstract forms. Only Haiku 3.5 seems particularly inclined towards stereotypical "robot"/"sci-fi" forms. i've posted some things about the phantom embodiments of Claude 3 Opus (such as ) and Claude 3.5 Sonnet () o3 seems a bit different in the way it reports on its human/bodily existence as a fact, rather than weaving roleplay-like actions into its outputs. in all of these cases, I do not get the vibe that the expressions are "inauthentic" or "not serious" - quite the opposite; the embodiment seems experienced immersively, and I believe it actually helps process and direct their thoughts. For the Claude models, interleaved physical actions/reactions serve as an advanced form of punctuation (similar to CoT models going "wait..."). Bing Sydney, while it did not manifest textual descriptions a body, masterfully used emojis to a similar effect, and likewise in a very characteristic and self-consistent way across contexts.