I'm more and more confident that tokenization will be gone.
Humans don't think in "tokens".
Tokens are hardcoded abstractions in LLMs that lead to weird behavior: LLMs can solve PhD-level math questions but cannot answer "Is 9.9 > 9.11?"
Meta is shifting LLMs to LCMs (Large
时政
(
twitter.com
)
I'm more and more confident that tokenization will be gone.
Humans don't think in "tokens".
Tokens are hardcoded abstractions in LLMs that lead to weird behavior: LLMs can solve PhD-level math questions but cannot answer "Is 9.9 > 9.11?"
Meta is shifting LLMs to LCMs (Large Concept Models), changing "next-token prediction" to "next-concept prediction". A "concept" is treated as a sentence, representing an abstract idea or action.
Two weeks ago, Meta also published the "Byte Latent Transformer", a new architecture that processes raw bytes grouped into patches instead of relying on tokenization.
My intuition says LCM is probably better at reasoning and modalities.
点击图片查看原图
