Deep seek papers are a must to read for anyone who wants to understand how to make LLMs operate at hyper scale. All western labs hide their best results, or at most release summaries that are about as meaningful as the answers Cleo used to give on stack exchange: https://math.stackexchange.com/questions/562694/integral-int...
I have a suspicion with how quiet all the major players got after the two weeks after deepseek R1 was released that they were reading and implementing everything in the papers that came with it as fast as humanly possible.
None of the major players have ever been quiet. DeepSeek enjoyed about a week or two's worth of press before its spotlight was stolent from the next great model. It never held the top spot, ever, mind you. So I don't understand why you think major players had to say anything about it, when the model was neither first, second or third in real world capability, and why they would have to say anything about it when DeepSeek service processes maybe an 1/8 of what OpenAI, Google or Claude in any given span of time.
I applaud their open efforts. But being "altruistic" and being best are two different things.
DeepSeek's contributions to training efficiency improvements were as, if not more, important than the models themselves. A lot of the worry people had about DeepSeek was related to people questioning the moat of the big AI players, since DeepSeek was able to train a competitive model with so much less compute.
Their innovations in training efficiency were almost guaranteed to have been heavily considered by the big AI labs. For example, Dario Amodei talks about the efficiency improvements being the real important contribution of DeepSeek V3 here: https://www.darioamodei.com/post/on-deepseek-and-export-cont...
> DeepSeek's team did this via some genuine and impressive innovations, mostly focused on engineering efficiency. There were particularly innovative improvements in the management of an aspect called the "Key-Value cache", and in enabling a method called "mixture of experts" to be pushed further than it had before.
It crashed the market because retail investors and perhaps non-retail as well had a great deal in overconfidence with the ability of the USA to maintain a lead thanks to the chip gap. High Flyer's innovations allowed them to scale and show that is not the case. This major event then likely spurred on many others. It was a mini 'sputnik moment'
Genuinely many times it seems most people need to find reasons to assume the best about DeepSeek and China in order to confirm their prior bias that “America bad” and “Capital is evil”. The reality is grey and fuzzy, with neither side landing on truth yet
How would people use deepseek to think "Capital is evil?" It was from a private hedge fund named "High Flyer," not a state university project or something.
Yes, exactly. How the heck? It makes no sense to me either, but you can certainly find plenty of laymen/not-in-the-know folks making those kinds of comments, often in non-technical spaces. Often the worst parts of the internet where discourse is non-existent. Human psychology allows for us to hold many contradictory positions all at once. Ideologies are the lens through which we view the world and it distorts our perception.
… wait did you just seriously tell SamA that he’s an asshole because of copyright issues… while praising Chinese labs who couldn’t give a rat fuck and won’t follow the same laws? Or pay creators? Physician, heal thyself
Sam's an asshole for a lot of reasons, a ridiculous commons grab of intellectual property draped in threadbare rhetoric about human welfare (get those developing nation eyeballs SCANNED people!) being just one of them.
Watching the Chinese labs kick the shit out of better funded US enclaves of TESCREAL psychopathy in the public fucking domain is gravy.
I don't care that their internal calculus or that of the PRC is to Cloud Strife Limit Break a bunch of "shareholder value" in the form of a bloated NVIDIA cap feeding frenzy by bloated "public benefit corporations" with a bunch of creepy ties to Thiel et al: they're publishing papers, code and weights. So they're hoovering up of the commons has something of value going back into the commons.
So yeah, fuck Sam and its going to be fun watching OpenAI and Anthropic pivot ever more towards trying to outlaw competition than they already have. Amodei already sounds like Donald Rumsfeld on Taiwan hawkishness, this is not the positioning of someone who loves their product roadmap.
It turns out that a zillion ScaleAI and SurgeAI turks don't have economics any better than paying NVIDIA to run 85% net earnings for CapEx that's obsolete by the time its racked and powered.
... You did not speak to the key point at all and went on some massive rambling incoherent political commentary. I feel this comment is unworthy of the thread.
Native Sparse Attention matters. Your commentary is beneath this paper.
I remember on february Deepseek's <think> caused a moderately sized market crash. They didn't just go silent, almost every vendor implemented their own version of thinking models while blaming Deepseek for stealing their tech/training on their models. It was rather pathetic to watch.
OAI and others were already on their way there or released the models. How did you manage to convince yourself that High Flyer did it first ? And that everyone else copied from them post-hoc? You’ve created a new chain of causality that simply does not match neutral reality
Do not try to signal intelligence by being sardonic or intentionally being obtuse. Actively avoiding the point someone is making rather than confronting it head on is beneath you.
ChatGPT o1 was made generally available in December 2024
DeepSeek r1 open weights were released in January 2025
My response wasn't intended to signal my intelligence as much as to show the overall lack of serious thought in your reply since you seem to be ignoring the fact that there was a 15%+ overall drop in the stock market, much higher in tech sector. But let's spell it out for you:
o-1 was mimicking the newest "explain your solution step by step" prompts which were proven to be more effective at the time.
ds-v1 came up with an actual chain of thought, imitating meandering and self doubt which sometimes went on for a while creating entertaining loops and introducing a new class of halting problem and this became the de-facto standard. They also revolutionized the industry by programming cards directly via PTX.
Then all of huggingface implemented the paper and we got q4 versions that "thought".
Hope that jolted your memory without killing Hitler.
For the first time, it introduced native sparse attention into the full training process, achieving up to 11× inference speedup while maintaining model performance.
> Despite being sparse, NSA surpasses Full Attention baseline on average across general benchmarks, long-context tasks, and reasoning evaluation.
Isn't it very notable that the latency improvement didn't have a performance loss? I'm not super familiar with all the technical aspects, but that seems like it should be one of the main focuses of the paper.
The performance maintenance (or even improvement) isn't surprising - sparse attention can reduce noise by focusing only on relevant tokens. Traditional full attention dilutes focus by attending to everything equally, while NSA's pruning approach mimics how humans selectively process information.
I am always skeptical of RNN approaches but this paper is just sparsifying the input, it is not compressing any size input to a fixed memory. I am hopeful maybe this is a big break. 11x inference speedup with no degradation from an algorithmic improvement. Is it really that good? almost too good to be true. Adoption in the next 6 months will tell us the truth.
I'd say award for best title is a tie between: "Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems"; "Finding Needles in Images: Can Multi-modal LLMs Locate Fine Details?"; and "Steering off Course: Reliability Challenges in Steering Language Models."
Yea I agree. It’s sad to find so much of the comments are focused on reinventing reality and jingoism instead of scientific discussion on the merits and technicals. I’ll return tomorrow and hope for better comments.
Deep seek papers are a must to read for anyone who wants to understand how to make LLMs operate at hyper scale. All western labs hide their best results, or at most release summaries that are about as meaningful as the answers Cleo used to give on stack exchange: https://math.stackexchange.com/questions/562694/integral-int...
I have a suspicion with how quiet all the major players got after the two weeks after deepseek R1 was released that they were reading and implementing everything in the papers that came with it as fast as humanly possible.
None of the major players have ever been quiet. DeepSeek enjoyed about a week or two's worth of press before its spotlight was stolent from the next great model. It never held the top spot, ever, mind you. So I don't understand why you think major players had to say anything about it, when the model was neither first, second or third in real world capability, and why they would have to say anything about it when DeepSeek service processes maybe an 1/8 of what OpenAI, Google or Claude in any given span of time.
I applaud their open efforts. But being "altruistic" and being best are two different things.
DeepSeek's contributions to training efficiency improvements were as, if not more, important than the models themselves. A lot of the worry people had about DeepSeek was related to people questioning the moat of the big AI players, since DeepSeek was able to train a competitive model with so much less compute.
Their innovations in training efficiency were almost guaranteed to have been heavily considered by the big AI labs. For example, Dario Amodei talks about the efficiency improvements being the real important contribution of DeepSeek V3 here: https://www.darioamodei.com/post/on-deepseek-and-export-cont...
> DeepSeek's team did this via some genuine and impressive innovations, mostly focused on engineering efficiency. There were particularly innovative improvements in the management of an aspect called the "Key-Value cache", and in enabling a method called "mixture of experts" to be pushed further than it had before.
Almost all of High Flyers achievements have more to do with scaling the process but when scaling is all you need, it’s darn effective
It crashed the market because retail investors and perhaps non-retail as well had a great deal in overconfidence with the ability of the USA to maintain a lead thanks to the chip gap. High Flyer's innovations allowed them to scale and show that is not the case. This major event then likely spurred on many others. It was a mini 'sputnik moment'
Genuinely many times it seems most people need to find reasons to assume the best about DeepSeek and China in order to confirm their prior bias that “America bad” and “Capital is evil”. The reality is grey and fuzzy, with neither side landing on truth yet
How would people use deepseek to think "Capital is evil?" It was from a private hedge fund named "High Flyer," not a state university project or something.
Yes, exactly. How the heck? It makes no sense to me either, but you can certainly find plenty of laymen/not-in-the-know folks making those kinds of comments, often in non-technical spaces. Often the worst parts of the internet where discourse is non-existent. Human psychology allows for us to hold many contradictory positions all at once. Ideologies are the lens through which we view the world and it distorts our perception.
MLA is just one example of a best-in-class technique from Hangzhou that's seen wide adoption in US prestige labs.
And the saltiness of US labs about DeepSeek is well-known. "O3, explain model distillation like I'm five."
No Sam, explain intellectual property rights to the judge in the NYT test case asshole.
… wait did you just seriously tell SamA that he’s an asshole because of copyright issues… while praising Chinese labs who couldn’t give a rat fuck and won’t follow the same laws? Or pay creators? Physician, heal thyself
Sam's an asshole for a lot of reasons, a ridiculous commons grab of intellectual property draped in threadbare rhetoric about human welfare (get those developing nation eyeballs SCANNED people!) being just one of them.
Watching the Chinese labs kick the shit out of better funded US enclaves of TESCREAL psychopathy in the public fucking domain is gravy.
I don't care that their internal calculus or that of the PRC is to Cloud Strife Limit Break a bunch of "shareholder value" in the form of a bloated NVIDIA cap feeding frenzy by bloated "public benefit corporations" with a bunch of creepy ties to Thiel et al: they're publishing papers, code and weights. So they're hoovering up of the commons has something of value going back into the commons.
So yeah, fuck Sam and its going to be fun watching OpenAI and Anthropic pivot ever more towards trying to outlaw competition than they already have. Amodei already sounds like Donald Rumsfeld on Taiwan hawkishness, this is not the positioning of someone who loves their product roadmap.
It turns out that a zillion ScaleAI and SurgeAI turks don't have economics any better than paying NVIDIA to run 85% net earnings for CapEx that's obsolete by the time its racked and powered.
... You did not speak to the key point at all and went on some massive rambling incoherent political commentary. I feel this comment is unworthy of the thread.
Native Sparse Attention matters. Your commentary is beneath this paper.
I remember on february Deepseek's <think> caused a moderately sized market crash. They didn't just go silent, almost every vendor implemented their own version of thinking models while blaming Deepseek for stealing their tech/training on their models. It was rather pathetic to watch.
OAI and others were already on their way there or released the models. How did you manage to convince yourself that High Flyer did it first ? And that everyone else copied from them post-hoc? You’ve created a new chain of causality that simply does not match neutral reality
Yeah I confess I rewrote history and crashed the stock market. Then ran out of juice just as I was about to kill Hitler.
Do not try to signal intelligence by being sardonic or intentionally being obtuse. Actively avoiding the point someone is making rather than confronting it head on is beneath you.
ChatGPT o1 was made generally available in December 2024 DeepSeek r1 open weights were released in January 2025
My response wasn't intended to signal my intelligence as much as to show the overall lack of serious thought in your reply since you seem to be ignoring the fact that there was a 15%+ overall drop in the stock market, much higher in tech sector. But let's spell it out for you:
o-1 was mimicking the newest "explain your solution step by step" prompts which were proven to be more effective at the time.
ds-v1 came up with an actual chain of thought, imitating meandering and self doubt which sometimes went on for a while creating entertaining loops and introducing a new class of halting problem and this became the de-facto standard. They also revolutionized the industry by programming cards directly via PTX.
Then all of huggingface implemented the paper and we got q4 versions that "thought".
Hope that jolted your memory without killing Hitler.
For the first time, it introduced native sparse attention into the full training process, achieving up to 11× inference speedup while maintaining model performance.
> Despite being sparse, NSA surpasses Full Attention baseline on average across general benchmarks, long-context tasks, and reasoning evaluation.
Isn't it very notable that the latency improvement didn't have a performance loss? I'm not super familiar with all the technical aspects, but that seems like it should be one of the main focuses of the paper.
The performance maintenance (or even improvement) isn't surprising - sparse attention can reduce noise by focusing only on relevant tokens. Traditional full attention dilutes focus by attending to everything equally, while NSA's pruning approach mimics how humans selectively process information.
Yes that’s what makes it so interesting and novel you nailed it
I am always skeptical of RNN approaches but this paper is just sparsifying the input, it is not compressing any size input to a fixed memory. I am hopeful maybe this is a big break. 11x inference speedup with no degradation from an algorithmic improvement. Is it really that good? almost too good to be true. Adoption in the next 6 months will tell us the truth.
DeepSeek and the Sparse Attention Revolution: How a Research Paper is Redefining AI Efficiency
https://deep.liveblog365.com/en/index-en.html?post=50
I'd say award for best title is a tie between: "Dehumanizing Machines: Mitigating Anthropomorphic Behaviors in Text Generation Systems"; "Finding Needles in Images: Can Multi-modal LLMs Locate Fine Details?"; and "Steering off Course: Reliability Challenges in Steering Language Models."
Well deserved
Yea I agree. It’s sad to find so much of the comments are focused on reinventing reality and jingoism instead of scientific discussion on the merits and technicals. I’ll return tomorrow and hope for better comments.
Title: Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
The awards page for ACL seems to disagree with this editorialized title: https://2025.aclweb.org/program/awards/
The ACL webpage has not been updated yet. Here are the announcement slides: https://cspaper.org/topic/116/record-breaking-acl-2025-crown...
The page that the person you’re replying to does have this so it may not be updated, or they were looking in the wrong place originally, or both:
> Industry Track Awards
> Best Paper
> Speed Without Sacrifice: Fine-Tuning Language Models with Medusa and Knowledge Distillation in Travel Applications
> Daniel Zagyva, Emmanouil Stergiadis, Laurens van der Maas, Aleksandra Dokic, Eran Fainman, Ilya Gusev, Moran Beladev
Per TFA, the paper we’re looking for is this one:
> Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
> Jingyang Yuan, Huazuo Gao, Damai Dai, Junyu Luo, Liang Zhao, Zhengyan Zhang, Zhenda Xie, Y. X. Wei, Lean Wang, Zhiping Xiao, Yuqing Wang, Chong Ruan, Ming Zhang, Wenfeng Liang, Wangding Zeng
I’m not finding it by author on the page you linked but I think it’s this reference by title:
> DeepSeek × PKU × UW — Native Sparse Attention: Hardware-Aligned and Natively Trainable Sparse Attention
I did find it on this page:
https://2025.aclweb.org/program/main_papers/
Link to the published paper rather than the preprint (update link?):
https://aclanthology.org/2025.acl-long.1126