Curtis "Ovid" Poe on Nostr: However, there have been several promising new LLM approaches which are virtually ...
However, there have been several promising new LLM approaches which are virtually identical, but they use integers for multiplication, or even trinary numbers (-1, 0, 1) and addition. This reduces the cost of training and using by an order of magnitude or more. So far, they seem to be almost as accurate as other LLMs, but scaling them up is still expensive.
Obviously, the above is a gross over-simplification, but I hope it gives you an idea of what's going on. 6/6
Published at
2024-11-14 12:24:00Event JSON
{
"id": "9e8ee8a3ba7c8d7bb36fe7a7fc58ab30d760df372c9bd49fef69b2b072d41b6a",
"pubkey": "a7426b90eef0ab497ad455a51e4896a6ad0a3ddee0f8a91e488341cd4841e4f3",
"created_at": 1731587040,
"kind": 1,
"tags": [
[
"e",
"331ddcac24da11494752b432cfcc4ca7fcedecb076554c2af4c9648b19a2ec3d",
"wss://relay.mostr.pub",
"reply"
],
[
"proxy",
"https://fosstodon.org/users/ovid/statuses/113481288290162750",
"activitypub"
]
],
"content": "However, there have been several promising new LLM approaches which are virtually identical, but they use integers for multiplication, or even trinary numbers (-1, 0, 1) and addition. This reduces the cost of training and using by an order of magnitude or more. So far, they seem to be almost as accurate as other LLMs, but scaling them up is still expensive.\n\nObviously, the above is a gross over-simplification, but I hope it gives you an idea of what's going on. 6/6",
"sig": "a2407cdd2858ed7f7116802b93477543feb272ccfcdcf9a24e1e932374349f139fa5402ae387f79fb24192fea602a459fb02d5619b20c498f6ee32ca4ec850cc"
}
