For a stochastic parrot to count to one trillion, it must explicitly train on the entire sequence of numbers. And to count backwards, it must also train on the reverse sequence of numbers. To do math, it must train on every permutation of every equation that it hopes to solve. There is essentially an infinite number of permutations that we expect grade school kids to solve.
No one is training LLMs like this: the hidden layers are somehow generalizing answers, not stochastically reciting them
ynniv /
npub12a…x6fdh
2025-04-27 15:26:19
in reply to nevent1q…hepj
Author Public Key
npub12akj8hpakgzk6gygf9rzlm343nulpue3pgkx8jmvyeayh86cfrus4x6fdhPublished at
2025-04-27 15:26:19Event JSON
{
"id": "d0b5083a65a25ec4480827d68f5416de35af4b634cf055dc01d39e349658d00a",
"pubkey": "576d23dc3db2056d208849462fee358cf9f0f3310a2c63cb6c267a4b9f5848f9",
"created_at": 1745767579,
"kind": 1,
"tags": [
[
"e",
"a6fff95bf29f5e10d2440c8f8e2e8f6d14cef160df81da92df300401e6c7413d",
"wss://nostr.bitcoiner.social/",
"root"
],
[
"e",
"d0da6034cfd615f97cefc583c6540defbf50248e77bda24ee3163fb423810dda",
"",
"reply"
],
[
"p",
"efd4d4a38bcd0004d31031a17972a6a6a3b32fe6f3953958c47d6325f2b8d106"
]
],
"content": "For a stochastic parrot to count to one trillion, it must explicitly train on the entire sequence of numbers. And to count backwards, it must also train on the reverse sequence of numbers. To do math, it must train on every permutation of every equation that it hopes to solve. There is essentially an infinite number of permutations that we expect grade school kids to solve. \n\nNo one is training LLMs like this: the hidden layers are somehow generalizing answers, not stochastically reciting them",
"sig": "3d1880d3948fd9e8beb5ea6d8aaff81ce72c2cf02014ed39aad1edd95f514aae1a08453511c1a083f7dcdfb3730407968d893c7dd578ecba1c7a9438112f5ccf"
}