An LLM is basically a bunch of layers that transform data into other data
Some of those probably at some step encode some “knowledge” of sorts
The problem is that it will work whatever you throw at it, even if it is a nonexistent thing, and return the most sensible “knowledge” according to the training data
Semisol / semisol
npub122…cgrkj
2025-03-10 23:31:58
in reply to nevent1q…pwdk
Author Public Key
npub12262qa4uhw7u8gdwlgmntqtv7aye8vdcmvszkqwgs0zchel6mz7s6cgrkjPublished at
2025-03-10 23:31:58Event JSON
{
"id": "1833ef49e83340f6cf02e7d4eda0d20d0ab5ad4e23fe43f74e72298c91fd6910",
"pubkey": "52b4a076bcbbbdc3a1aefa3735816cf74993b1b8db202b01c883c58be7fad8bd",
"created_at": 1741649518,
"kind": 1,
"tags": [
[
"e",
"c801c18603870acaba2975aa27ccc493590f9cc1cae25e012f39cc2aa42aa366",
"wss://hbr.coracle.social/",
"root"
],
[
"e",
"5a71d83193ecc99cf42f2fe4ec281398ea24dce3f7d03b2a9b2cdaa58a2d11b1",
"",
"reply"
],
[
"p",
"b7cf9f42a796b091e843dce919d3ef4c0dc82e029452edf0bdbcdeb9ecb93e78"
],
[
"p",
"4b0bcab460adda31fad5a326fb0c04f6ec821fb24be85dbdc03c04cc0e12fc07"
]
],
"content": "An LLM is basically a bunch of layers that transform data into other data\n\nSome of those probably at some step encode some “knowledge” of sorts\n\nThe problem is that it will work whatever you throw at it, even if it is a nonexistent thing, and return the most sensible “knowledge” according to the training data",
"sig": "0f2ddf5b072553fe021a62ea1103153cb478c9f49875da818bceb4a0b5ca762e993b891dd1527d0957b911c0af5d3cb135585b56e0f369fe22ede795ec429a5e"
}