Spring Jo 🥚 
🍀Apple did the research; LLMs cannot do formal reasoning. Results change by as much as 10% if something as basic as the names change.
https://garymarcus.substack.com/p/llms-dont-do-formal-reasoning-and
Not too surprising, but a big weight like Apple standing behind this should shift sentiment more and more
yopp@ShadowJonathan it’s really weird that some people are pushing LLM as something that can reason, while its architecture is Key-Value storage with sophisticated probabilistic query and value encoding mechanisms.
LLMs just don’t have enough layers for anything besides queries, so it can’t have any relational capabilities that allow to make multi step decisions.
Also tokenization hides a lot of structure of the language from encoding process, which adds additional source of errors.
I’m sure we can build something that can reason at some point, but it requires very diffirent and more complex architecture.
Nova🐧✨@alex @ShadowJonathan it's definitely not key-value storage, it's more latent space interpolation than anything... tokenization isn't that bad either, it's mostly just that there's no depth like you mentioned!
@technobaboo @ShadowJonathan but isn’t latent space interpolation just basically maps some input to some output? Input embeddings always return the same next most probable token? Sure, underlying structure is more complex with all this encodings, decoding and attention layers, but on a surface you just putting some query in and getting some value out from the constant storage. And we have no idea how to reliably write or read from that storage, but that’s entirely diffirent problem :D
@alex @ShadowJonathan it's definitely a mapping function for sure but key value storage implies a direct mapping via table when it's way less linear than that