TeMPOraL on Nostr: Okay, so apparently, the better the #LLMs get at reasoning*, the worse they get at ...
Okay, so apparently, the better the #LLMs get at reasoning*, the worse they get at writing jokes. Now, did anyone tried to identify the actual vector in the latent space corresponding to humor? Wonder what would happen if one did, and then cranked it up to eleven when running inference.
(I'd try it myself, but am not sure which tools and which models support this without too much work.)
#LLM #AI #DeepSeek #ChatGPT #o1
* - I.e. they get better at dissecting the metaphorical frog.
Published at
2025-01-20 22:18:26Event JSON
{
"id": "ea2f7aa5f840c68d44e87e5aee998712ab9e0c1ae860c7fc10e57e55d61acbd4",
"pubkey": "8d23f08fe393c94cadd64bbb3aa52f9afd9d3da542a6f5eaee2e2e18dc7dd26a",
"created_at": 1737411506,
"kind": 1,
"tags": [
[
"t",
"llms"
],
[
"t",
"llm"
],
[
"t",
"ai"
],
[
"t",
"deepseek"
],
[
"t",
"chatgpt"
],
[
"t",
"o1"
],
[
"proxy",
"https://fosstodon.org/users/temporal/statuses/113863000487792253",
"activitypub"
]
],
"content": "Okay, so apparently, the better the #LLMs get at reasoning*, the worse they get at writing jokes. Now, did anyone tried to identify the actual vector in the latent space corresponding to humor? Wonder what would happen if one did, and then cranked it up to eleven when running inference.\n\n(I'd try it myself, but am not sure which tools and which models support this without too much work.)\n\n#LLM #AI #DeepSeek #ChatGPT #o1 \n\n* - I.e. they get better at dissecting the metaphorical frog.",
"sig": "25a257587bfdbf04c60487af52c1b9123f39eb05fb4167a3428543b198ad99f2ac1ea7939e85443f4f53c7293fffcb9e6729c565597543fcfb87b3dbc0f58191"
}
