14 comments

[ 2.6 ms ] story [ 35.8 ms ] thread
Looks like Deep Sparse Attention can help with code (structured and long-file reasoning)
awesome that sparse attention used in real world setting
Happy to see Chinese OSS models keep getting better and cheaper. It also comes with a 50% API price drop for an already cheap model, now at:

$0.28/M Input ($0.028/M cache hit) > $0.42/M Output

wow...gigantic reduction in cost while holding the benchmarks mostly steady. Impressive.
The 2nd order effect that not a lot of people talk about is price: the fact that model scaling at this pace also correlates with price is amazing.

I think this is just as important to distribution of AI as model intelligence is.

AFAIK there are no fundamental "laws" that prevent price from continuing to fall, at least correlated with Moore's law (or whatever the current AI/Nvidia chip development cycle is called right now)- each new generation of hardware is significantly faster/cheaper than the next- so will we see a ChatGPT-5 model at half the price in a year? (yes I know that thinking models cost more, but just on a per-token basis)

We are heading into the future of very low-cost AI inference. It's a good thing, and expected.
You guys rock! I'm very curious how will this perform against real word data, where small nuance matters. Also have you tested it beyond 128K context window?
Interesting that models still evolve fast enough that dedicated model-specific hardware isn't a big contender right now. We're still seeing major scaling gains on mostly generic platforms.
What happened to Meta Open weights models? Lately I keep hearing more of Deepseek than LAAMA?
One huge problem with these "cheap" models is that they happen to be more expensive in the typical agent workflow if the provider does not support caching.

Input and output costs are peanuts compared to the order of magnitude(or more) amount of tokens that hit the cache.

At that point you might as well use GPT-5. It will be the same price or cheaper, and more capable.

Not sure if I get it correctly:

They trained a thing to learn mimicking the full attention distribution but only filtering the top-k (k=2048) most important attention tokens so that when the context window increases, the compute does not go up linearly but constantly for the attention->[query,key] process (it does grow up linearly in the graph because you still need to roughly scan the entire context window (which an "indexer" will do), but just very roughly here in order to speed up things, which is O(L) here).

Prices fall, benchmarks remain stable. Maybe in the future, LLM will spend most of its money on electricity.