Fast KV Compaction via Attention Matching
arxiv.orgConsidering the insanity of the AI arms race going on now, and the incredible sums of money be thrown at any slight advantage, is there any reason to believe that any meaningful AI breakthrough would be openly published for anyone to leverage?
These folks are MIT, so citations are valuable to them. Citations convert into prestige, academic career progression, or a favorable exit from academia into industry.
Also, I don't see why you couldn't patent this if you wanted to monetize it.
> Also, I don't see why you couldn't patent this if you wanted to monetize it.
We all just saw the prior art published for the public. That will preclude patenting this work. Further reduction to practice is required.
(I am not a lawyer).
Yes there is. Lots of researchers are more interested in making a contribution to societal flourishing than in making incredible sums of money. That’s why there’s still lots of top AI researchers in academia.
I do sometimes wonder -- if the transformers paper wasn't published, what would the industry be like? Would the same ideas have been put together in almost the same way weeks or months later somewhere else?
I would say yes.
The reality is that the money being thrown = the time of humans. I guess compute as well, but in terms of people doing innovation - openly published things are the same thing, minus the money.
The inventor's grace period under first to file changes still gives them/their university a year to file if they publish openly.
I know the frontier “labs” are holding back publications.
I don’t think it will last among researchers who think beyond production LLMs
Superficially it sounds like this could create a bit more of a move toward doing compaction on some continuous basis, or compacting in batches once you hit the context limit, rather than starting fresh with a summary and system prompt..
Feels like high fidelity, fast compaction could be a path to “solving” long context.
This looks promising. I've added it to my reading list.
This is big for long-horizon tasks
None of the compaction accuracies look impressive.
I think matching or exceeding the original cache at 20% compacted size is fairly impressive.
The original cache had 70% accuracy, and the alternatives were only worse.
It sounds like you looked at figure 1 but not figure 3.