I don’t understand why this lab is allergic to providing details on what they actually made, especially when Chinese labs are more than willing to share architectural specs/code/kernels (eg NSA/FSA, RAMBa, HISA, DSA LightningIndexer, etc). I don’t doubt that they’ve done something here, but the lack of details makes me default not trust this, particularly when this is the second time that they’ve released a “technical report” that just waxes poetic about the concept.
jmward01 19 minutes ago [-]
Well, I know this is possible because I have built things that work just like it is promising to do. The two key technologies needed are:
- guided window attn. Predict where to attend to but in a fixed window. If you do this to just the token/vocab you can keep effectively unlimited context and perfect recall. (yes, I can do that. There is a trick to teaching it how to predict position. This also immediately opens other crazy things like NN memory)
-efficient fixed state size models. So not a recurrent mechanism because that breaks training, parallelizable like transformers, but fixed sized state instead of unbounded attn. Pick a reasonable amount of state and it is amazingly good since it doesn't need to keep separating wheat fro chaff in context (yes, it is possible to build this, I have. It works. This also opens up real streamed models. I have a true infinite context streamed model I toy with locally that I am getting to be audio/text in and audio/text out in real time.)
Put those together and you have O(1) token gen, infinite context and perfect recall. It is a whole new world of models. You can interact with a model until you have it at the state you want and then save its state and use that as if it were your system prompt. Batches pack perfectly so inference is massively more efficient. Training is massively more efficient. Transformer and unlimited attn models are a dead end. But how do you make money on this as an independent researcher? If I release the Two Weird Tricks this is all based on I get zip and the big players get even more tech for free. If I keep it all secret I get Zip and eventually the tricks will be figured out. (Yes a little frustration here) If anyone wants the model architecture of the future make me an offer :)
regularfry 11 minutes ago [-]
It's not quite true to say that if you release it you get nothing. If it's worthwhile and picked up by the open-weights labs, you get much bigger and better models implementing it than you would have had access to or been able to train otherwise, quicker than if they had to figure it out de novo.
jmward01 7 minutes ago [-]
Yeah. I am about to the point of just releasing it all. I love the tech. It does amazing things. But I want to move to the next big things I can see doing with it and building the custom ops to get it to work efficiently is a pain. I am positive others would run with it and make it all way better which would free me up to do more.
famouswaffles 2 hours ago [-]
Business wise, it would make sense to hold off on details till they're at least ready to serve. Look at what happened with Open AI and reasoning models. Everyone struggled with getting RL to work with LLMs for a good while. Open AI figured it out, and a few months later everyone had their prototypes out in short order. Don't forget who these labs employ. They're some of the brightest people around. Sub-q aren't really in a position for that lol. If they'd shared details at the first announcement for instance, the big labs might have had something out by now while they're still pulling resources to scale and then what ?
supern0va 1 hours ago [-]
You don't understand why the thing their entire company is valued upon is...not being given away freely? They literally are taking an open source model and then adapting it with this technique. If they disclose it, the frontier labs will immediately copy it and outperform them.
My guess is that they're angling for an acquisition.
GenerWork 28 minutes ago [-]
>My guess is that they're angling for an acquisition.
This is what I've thought was going to happen ever since they publicized their efforts. They probably don't have the money to train large models themselves, might as well get a nice chunk of change by being acquired by someone who already has said large models running.
giancarlostoro 3 minutes ago [-]
They probably don't have the money to run the model at reasonable scale.
giancarlostoro 2 hours ago [-]
This one's interesting, and I think the next frontier for LLMs should really just be, how can we get something like Opus 4.6 to cost drastically less, for the same output? I say 4.6 because from 4.6 onwards it's been pretty darn good, at least for me, always feels like every model upgrade someone hates it, heck even 4.5 was fine.
robmccoll 2 hours ago [-]
Yes - I want that and dramatically faster. Newer models don't seem to need any more or less guidance and iteration, so let's make the time-to-wrong-answer as short as possible.
giancarlostoro 2 hours ago [-]
I'm not as crazy about speed as long as it's reasonably as "quick" as Opus. Which is faster than most developers can spit out code. I do get annoyed with Claude Code because it looks like it chooses to be as slow as possible, but maybe that's by design so its not pounding their backend every milisecond? Would probably be bad.
Local inference is insanely fast on my M4 Pro MBP though, so I can understand where you're coming from, but I don't need it too much faster. I still need time to review, test, review and provide feedback to the model. Fast is okay I guess for true vibe coding.
robmccoll 1 hours ago [-]
I just don't want to have to have a pipeline going in order to fully occupy my time. I don't want to wait on the model to review the prompt, read the parts of the codebase indicated, do its own research in the codebase and documentation, plan, run agents ... actually write the code and NOW I can start reading it and reviewing it. That means I either need to run a lot of operations in parallel so that I always have something to do and the agent(s) are highly utilized or I'm writing something on my own that I keep getting that keeps getting interrupted. It's the constant context switching that kills me. I want to work on one problem at a time and really focus on it - even if I'm not writing every line myself.
wxw 1 hours ago [-]
> SSA replaces the O(n²) dense attention pass with a learned sparse formulation that scales linearly with context length.
> At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2.
Awesome stuff. Solving context at the model architecture layer rather than trying to bolt on extra memory is the right direction IMO.
samber 48 minutes ago [-]
According to Subquadratic, Needle in a Haystack is strong up to 12m tokens, but RULER has not been tested above 128k tokens ??
samber 41 minutes ago [-]
Comparing compute cost versus FlashAttention-2 is not very honest to me.
FlashAttention-2 is not used anymore for at least 2y.
This architecture would have been a massive improvement 3 years ago, but it is a ~solved~ problem IMO.
embedding-shape 2 hours ago [-]
> SubQ 1.1 Small scores near-perfect at 1M, 2M, 6M, and 12M tokens. The model was trained predominantly at 1M tokens yet the retrieval held near perfectly at 12x that length, despite compressing attention to just 0.13% of relationships. This generalization is a direct consequence of SSA routing attention based on content relevance rather than fixed positional patterns.
If the results persists from 1M to 12M, why not 24M or 48M? Sounds almost too good to be true.
With back of the napkin math from inside my head, that'd be like 0.5/1 million LOC, depending on language/code density, could just fold the entire codebase into one prompt if it's a small one, that'd be neat :)
monster_truck 59 minutes ago [-]
It likely falls off very steeply after that. 8 to 1 (which I am assuming based on the 0.13% figure) is a pretty common ratio for sparse matrix stuff.
It's been all talk and no action ever since their first announcement.
Depurator 1 hours ago [-]
What kind of hardware would be needed to serve an instance with the full 12m context? And what kind of speeds can one expwct at those extremes at 10m+?
maz1b 1 hours ago [-]
They've done multiple "evaluations" by third parties, but still, it seems that they aren't being fully transparent. I think the approach is quite interesting and novel, but this feels like deja vu.
I get why they aren't disclosing all the details, but it seems more hype-train-esque to me for this moment. I don't disagree that this could be big.
aesthesia 2 hours ago [-]
Disappointing they don't actually say how their sparse attention mechanism works.
Rendered at 17:01:19 GMT+0000 (Coordinated Universal Time) with Vercel.
- guided window attn. Predict where to attend to but in a fixed window. If you do this to just the token/vocab you can keep effectively unlimited context and perfect recall. (yes, I can do that. There is a trick to teaching it how to predict position. This also immediately opens other crazy things like NN memory)
-efficient fixed state size models. So not a recurrent mechanism because that breaks training, parallelizable like transformers, but fixed sized state instead of unbounded attn. Pick a reasonable amount of state and it is amazingly good since it doesn't need to keep separating wheat fro chaff in context (yes, it is possible to build this, I have. It works. This also opens up real streamed models. I have a true infinite context streamed model I toy with locally that I am getting to be audio/text in and audio/text out in real time.)
Put those together and you have O(1) token gen, infinite context and perfect recall. It is a whole new world of models. You can interact with a model until you have it at the state you want and then save its state and use that as if it were your system prompt. Batches pack perfectly so inference is massively more efficient. Training is massively more efficient. Transformer and unlimited attn models are a dead end. But how do you make money on this as an independent researcher? If I release the Two Weird Tricks this is all based on I get zip and the big players get even more tech for free. If I keep it all secret I get Zip and eventually the tricks will be figured out. (Yes a little frustration here) If anyone wants the model architecture of the future make me an offer :)
My guess is that they're angling for an acquisition.
This is what I've thought was going to happen ever since they publicized their efforts. They probably don't have the money to train large models themselves, might as well get a nice chunk of change by being acquired by someone who already has said large models running.
Local inference is insanely fast on my M4 Pro MBP though, so I can understand where you're coming from, but I don't need it too much faster. I still need time to review, test, review and provide feedback to the model. Fast is okay I guess for true vibe coding.
> At 1M tokens, SubQ 1.1 Small requires 64.5x less compute than dense attention and runs 56x faster than FlashAttention-2.
Awesome stuff. Solving context at the model architecture layer rather than trying to bolt on extra memory is the right direction IMO.
FlashAttention-2 is not used anymore for at least 2y.
This architecture would have been a massive improvement 3 years ago, but it is a ~solved~ problem IMO.
If the results persists from 1M to 12M, why not 24M or 48M? Sounds almost too good to be true.
With back of the napkin math from inside my head, that'd be like 0.5/1 million LOC, depending on language/code density, could just fold the entire codebase into one prompt if it's a small one, that'd be neat :)
I get why they aren't disclosing all the details, but it seems more hype-train-esque to me for this moment. I don't disagree that this could be big.