Jesse Baer 🔥 on Nostr: Something that confuses me about LLMs: the hidden prompts. And in general, prompts ...
Something that confuses me about LLMs: the hidden prompts. And in general, prompts that take the form "you are ___." Outside of cheesy hypnotism scenes, how often in the training data does A tell B who they are, and B proceed to act accordingly? Why does this work as well as it does?
Published at
2024-08-16 20:53:30Event JSON
{
"id": "f5fdc5aebe79dfc78100f69175986a657651bfc7e4c04af001dfaed5e085fd2a",
"pubkey": "faaf81ec8b66e5875dcb43e3b02a509ff4a36869da2ba49ed4bab1a1c5b0b328",
"created_at": 1723841610,
"kind": 1,
"tags": [
[
"proxy",
"https://mastodon.social/users/misc/statuses/112973683778272720",
"activitypub"
]
],
"content": "Something that confuses me about LLMs: the hidden prompts. And in general, prompts that take the form \"you are ___.\" Outside of cheesy hypnotism scenes, how often in the training data does A tell B who they are, and B proceed to act accordingly? Why does this work as well as it does?",
"sig": "128fe024c622610458b8f2723556d7b649b59598f7e75c384fb9b18739d4e570618c1058fb8dc6a573a12270274090dc5a03e4abc6523531dcec1cdab268a200"
}
