zellyn on Nostr: npub10fmsw…zx0qg Folks have played with it a bit. So far the most interesting thing ...
npub10fmsw3ukhmqnnsjmalfekl6j4cucxvfl50he9dmqcna36a99nf6s8zx0qg (npub10fm…x0qg) Folks have played with it a bit. So far the most interesting thing I've seen is one that makes up responses as you go, while keeping the format of int-fiction.
Remember, they're trained to complete text with the likely next word, so they remain consistent only when that's the most likely response based on their training data.
The increasing context length should allow them to "remember" more of the transcript (and thus stay more consistent), but it's fundamentally not how they work.
Published at
2023-11-07 15:52:34Event JSON
{
"id": "f008fb72c0fd248d057f2aa1952c08ae154b336d8239f7488f7d09995dddfc88",
"pubkey": "20598784bd7e3d0ab164602cb86dc5b8664ceb509cfa6fc22e1a77065d7481fc",
"created_at": 1699372354,
"kind": 1,
"tags": [
[
"p",
"7a77074796bec139c25befd39b7f52ae3983313fa3ef92b760c4fb1d74a59a75",
"wss://relay.mostr.pub"
],
[
"p",
"3ad806433f1d09d2643dd8bf4162890270449539839aab7b9da215e1d5681bac",
"wss://relay.mostr.pub"
],
[
"e",
"8c96b8de4b2c56952735aef6b1575eba1f2ae301d63a641ebf806082f059efa8",
"wss://relay.mostr.pub",
"reply"
],
[
"proxy",
"https://hachyderm.io/users/zellyn/statuses/111370066597051764",
"activitypub"
]
],
"content": "nostr:npub10fmsw3ukhmqnnsjmalfekl6j4cucxvfl50he9dmqcna36a99nf6s8zx0qg Folks have played with it a bit. So far the most interesting thing I've seen is one that makes up responses as you go, while keeping the format of int-fiction.\n\nRemember, they're trained to complete text with the likely next word, so they remain consistent only when that's the most likely response based on their training data.\n\nThe increasing context length should allow them to \"remember\" more of the transcript (and thus stay more consistent), but it's fundamentally not how they work.",
"sig": "32c8af98884962e694a76cc0ad071129b7871db6c2014a5c20009c6b431d1339dfe6579679962adb04ce99aaa629cd851e5c629b6d4181d7eae7f363c1997d99"
}