I'd also like to see a list of similarly-simple techniques for extracting rules where ML researchers could automatically try them all. In this case, the N-gram rules would be the starting point. For what predictions failed, they'd try to throw in the other techniques. Eventually most or all of the predictions should be captured by one or more simple rules. Some might be compound rules mixing techniques.

I think there will also be benefits to that both in interpretability and hardware acceleration. In time, maybe cheaper pretraining of useful models.

They literally can! The exact speculative method is supported on vLLM using `speculative_model="[ngram]"`[1]

1: https://docs.vllm.ai/en/latest/features/spec_decode.html#spe...

  Current prompt
  "Article: (CNN)French striker Bafetimbi Gomis, who has a history of [...]
  Summary: French stri"

  Prompt lookup algorithm
  1. Get last few tokens from prompt -"French stri"
  2. Search for "French stri" in prompt
  3. Match found - return next k tokens after match as candidate completion -"ker Bafetimbi Gomis, who has"

  Candidate tokens
  "ker Bafetimbi Gomis, who has"

I've always thought that LLMs are still just statistical machines and that their output is very similar to the superpermutation problem, though not exactly.

I just like to think of it as a high dimensional view of the relationships between various words and that the output is the result of continuing the path taken through that high dimensional space, where each point's probability of selection changes with each token in the sequence.

Unfortunately there's no thought or logic really going on there in the simplest cases as far as I can understand it. Though for more complex models/different architectures anything that fundamentally changes the way that the model explores a path through space like that could be implementing thought/logic I suppose.

It's why they need to outsource mathematics for the most part.

"How does this have 74 points and only one comment?"

I think one cause is hobbyists upvoting submissions that might be valuable to people in a specific field. We understand just enough to think it could be important but defer to subject matter experts on the rest. That's why I upvoted it.

These are different people:

https://arxiv.org/search/cs?searchtype=author&query=Nguyen,+...

Wikipedia mentions that up to ~40% of the Vietnamese population (~40,000,000 people) carries the name Nguyen:

https://en.wikipedia.org/wiki/Nguyen

For the paper itself, as someone working in the field, I find it interesting enough to consider reading at some point (I do not read that many analysis papers recently, but this one looks better than most). As for your accusation about it claiming that large language models are simply n-gram models, read the abstract until you realise that your accusation is very much unfair to the work.

> Thr author submitted like 10 papers this May alone. Is that weird?

Chances are, you just assumed all the search results for 'Nguyen, T' refer to the same author.