For all the discussion of "prompt engineering" and "finetuning", I think the most interesting biasing structure for modern AI that has flown somewhat under the mainstream discussion is the ability to directly constrain the output space through e.g. grammars for llms and control nets for image generation.
It's weird to see people deploy the raw output of large scale generative statistical models when there are pretty powerful tools just sitting there that allow more finegrained application.
Sarah Jamie Lewis /
npub14m…vp58n
2024-06-03 17:44:13
Author Public Key
npub14mfj9wu5ujvu6rxj8w62dvkvqj7myqc6kz0upz3kuw3wx4dz9vgstvp58nPublished at
2024-06-03 17:44:13Event JSON
{
"id": "cd3eeaf45f9792b94117547ebb73caa5981e6ad20e561ea9a2bb0890b04ce064",
"pubkey": "aed322bb94e499cd0cd23bb4a6b2cc04bdb2031ab09fc08a36e3a2e355a22b11",
"created_at": 1717436653,
"kind": 1,
"tags": [
[
"proxy",
"https://mastodon.social/users/sarahjamielewis/statuses/112553928538763850",
"activitypub"
]
],
"content": "For all the discussion of \"prompt engineering\" and \"finetuning\", I think the most interesting biasing structure for modern AI that has flown somewhat under the mainstream discussion is the ability to directly constrain the output space through e.g. grammars for llms and control nets for image generation.\n\nIt's weird to see people deploy the raw output of large scale generative statistical models when there are pretty powerful tools just sitting there that allow more finegrained application.",
"sig": "e10fb35bf1999aee437c9c5b1486b5ce1be5ba2ff29074ed1ac164a20c5bcaf92e3a6663dac2d33d334943fe2933d4358d650d717539b533334b8c3a79fce23e"
}