your translation is correct, I would say Chinese labs may be able to figure out the current capability of latest frontier models in 3-6 months, but then Anthropic and OpenAI may have already been ASI in that time already. China's main problem is still lack of (good) chips, and that is a hardware issue that is unlikely to be solved for a while. and more effort for efficiency means less effort for actual capability improvements. we have already seen what anthropic can do if they focus on efficiency with opus 5.5
Short term they might have less compute, but long term they will most definitely have more compute. They don't have to worry about energy, they don't have to worry about people blocking them building data centers the only thing stopping them is there no Chinese manufacturer that can produce a chip as good as Nvidia but I would bet that solved in a year or so.
They don't have to worry about people blocking DC construction in the US either. All new AI datacenters are designated "dual-use" so the federal government is already letting local councils know to fuck off.
they can also put them in better places, vs trying to arbitrage expensive electricity costs and various US corruption thats built more around paying people off than getting things done
Yeah, but that brings up another problem that China doesn't have. Their federal government is always on the same page. The US federal government changes every two years and it doesn't seem like one side is going to be allowing building of AI data centers anymore and in fact might just ban AI in general.
It's hard enough to build them in the US that multiple companies are unironically spending 10s of millions of dollars to try to build them in space
Which cynics could say is marketing hype, but I tend to believe it's extremely difficult to build anything land/energy intensive in the modern United States
Well we do know the results. They’ve been at the frontier constantly, including developing the entire field and set of capabilities. In the beginning every OpenAI paper was basically a landmark.
They were at the frontier early since they allegedly “stole” the internets data first. Now they’re just lame ducks trying to ensure they stay at the top by legislation and ensuring open source doesn’t get its say. Also OpenAI trying to adopt any other product (legal tooling, healthcare, latex replacement, etc…).
They were at the frontier because they invented many of the techniques. Google already had the internet’s data well beforehand. Anthropic and OpenAI are still at the frontier today, independent of legislation.
Oh no I agree, they were among the first to be popular in the llm frontier with their techniques. Today they are only holding on to the frontier because of that lead.
They want the legislature to keep their lead from other competitors.
I'm actually the most innovative, I've written thousands of papers advancing the state of human knowledge. They're just in my basement and I don't show anyone.
if it's agentic stuff, they likely aren't hammering a core constantly and they will maybe sit idle quite often between model requests, so it makes sense. I do wonder how much memory they allocate to each one though.
it's just very efficient use of shared cores that is required to make these kinds of workloads cost efficient
A place I worked back around 2020 was running a Grafana instance per customer that got embedded on the web dashboard. We had 110 pods per GKE (kubernetes on gcp) 4 CPU node because that was a network imposed pod limit at the time. The nodes were usually idle--could have shoved a lot more on if not for the IP limit.
I think around that time Grafana changed their license tho so you couldn't host OSS Grafana as part of your service.
Even without that particularly special case, high-energy physics collaborations have long since broken the idea of authors. There are now many collaborations with hundreds of authors publishing regularly and quite a few that are into the thousands.
It just describes the platform they built for scheduling workloads, and running those workloads. After a second thought it is not that impressive and probably doesn't deserve any kind of hype. It's the same kind of setup AWS is running for Lambda, as well as anyone else basically running SLURM clusters out there.
Still giving them credit because creating such as scheduler/platform from scratch is quite complex, and I know that they probably struggled a lot to get it right.
I'd say it's cool that they're openly writing about it and how they run their workloads. shows a nice window into how these things are actually deployed at scale.
it also shows how much density you can get easily from a single core if you wanted to replicate this
That's not quite right. A recurring trend in ML is figuring out how to get an elastic interface for the the particular quirks of ML workloads (source: I worked on an open source one years ago for inference). Recently, there's been a lot of interest in doing this for agent workloads. Google recently released something called ax that is similar in spirit. The core of it is this: https://github.com/agent-substrate/substrate
At a high level, agent sandboxes have peculiar needs. Agents are really bursty, but also long lived. You need low latency suspend/resume calls. Checkpointing and recovery have some particular considerations. And naive approaches are often really wasteful, but over optimizing without harming durability, isolation, or consistency in performance can get tricky.
This is the new hot infra topic for agent swarms, for the time being. Whether that's impressive or not is up to the reader I guess, but it's a pretty involved project nonetheless.
The person simplified the paper with such a misunderstanding that I thought 10x before rebutting. They clearly didn’t read anything of the paper. Thanks for the clarification.
Is it possible they're doing a research lab "socialism style" and everyone gets equal credit for just being a part of the lab, regardless of actual input into the specific papers? If they're innovating in computer, maybe they're not so afraid of innovating in social/academic structures as well?
Sorry, didn't mean it like "yelling", I'm quite a fan of socialism and don't see it as any negative. I didn't think anything in my comment was negative.
It's negative in the sense it suggest some authors didn't contribute (or contribute little0 to the research, and were perhaps just warm bodies that happened to be present at a certain location
I don't understand why half the comments here are about the author list, this is very common practice in e.g. large-scale physics experiments and biology, and every new GPT release from OpenAI equally had papers with tons of authors
It seems every DeepSeek paper/patent has a huge number of authors, and this one is no exception. They couldn't even fit everyone on the page, there are 31 others not shown. This could be an asset protection strategy (i.e., human assets). Imagine if there were only 3 authors. Those authors may get hired away by competitors. If you list every employee on every paper then competitors don't know who to lure away.
Corresponding author is the head of the group, usually the one who coordinates the study, the one who knows everyone else. Is the "if you have any question about anything in this paper, contact me" person, even years after publication.
I miss the YOLO (CNN computer vision model) days where one dude can publish a paper, completely disregard academic conventions, and yet push the field forward by leap and bounds.
I wonder if they are signalling that if they can do this for training, then they can create an style agent swarm to hack anyone with 380k concurrent agents.
"Within one scale unit, the platform spans nearly 160 CPU nodes with 30K cores and ∼250 TB of DRAM. It manages petabytes of layers and images. On a typical day, a single scale unit serves about 3 M sandbox instances, with peak concurrency reaching ∼380K and a creation rate exceeding 5,000 instances per second."
Impressive numbers!
Whoever would have thought (in prior years) that in 2026 AI Agents (not people or corporations, at least not directly) seem to be (or seem to be rapidly becoming) the biggest consumers of cloud computing resources...
12 sandboxes per code is insane, I wonder how many of these sandboxes are idle at a time. Depending on the tasks assigned the resource requirements are different. Compare an agent doing pdf conversion and one responding to a simple question. One is cpu bound the other is mostly network wait.
This is an interesting problem from infra perspective since you cannot predict the workload. On a bigger scale you may get away with forecasts.
Im waiting for tech that elastically allocates cpu/mem without restarting a container.
Is there a lab more innovative than DeepSeek? Imagine if they had the same compute resources that Anthropic and OpenAI have.
Food for thought: Constraints are the source of creativity.
Yeah if they had the resources of an OpenAI or anthropic they’d be OpenAI or Anthropic. Scrappy underdogs have to be nimble and innovative.
They are as “underdog” as Linux is to Windows.
Not really, they are underdogs in the true sense of the word.
Possibly relevant: 突破技术壁垒, "break the technical barricade" -> overcome an obstacle through innovation (in this case, sub SOTA GPUs at least).
Native Chinese speakers to confirm....
your translation is correct, I would say Chinese labs may be able to figure out the current capability of latest frontier models in 3-6 months, but then Anthropic and OpenAI may have already been ASI in that time already. China's main problem is still lack of (good) chips, and that is a hardware issue that is unlikely to be solved for a while. and more effort for efficiency means less effort for actual capability improvements. we have already seen what anthropic can do if they focus on efficiency with opus 5.5
Short term they might have less compute, but long term they will most definitely have more compute. They don't have to worry about energy, they don't have to worry about people blocking them building data centers the only thing stopping them is there no Chinese manufacturer that can produce a chip as good as Nvidia but I would bet that solved in a year or so.
They don't have to worry about people blocking DC construction in the US either. All new AI datacenters are designated "dual-use" so the federal government is already letting local councils know to fuck off.
they can also put them in better places, vs trying to arbitrage expensive electricity costs and various US corruption thats built more around paying people off than getting things done
Yeah, but that brings up another problem that China doesn't have. Their federal government is always on the same page. The US federal government changes every two years and it doesn't seem like one side is going to be allowing building of AI data centers anymore and in fact might just ban AI in general.
It's hard enough to build them in the US that multiple companies are unironically spending 10s of millions of dollars to try to build them in space
Which cynics could say is marketing hype, but I tend to believe it's extremely difficult to build anything land/energy intensive in the modern United States
Just because other companies don't write papers about what they do doesn't mean they aren't innovative...
Sure, but we’ll never know what they do or if it is innovate because we won’t know what they do.
Well we do know the results. They’ve been at the frontier constantly, including developing the entire field and set of capabilities. In the beginning every OpenAI paper was basically a landmark.
They were at the frontier early since they allegedly “stole” the internets data first. Now they’re just lame ducks trying to ensure they stay at the top by legislation and ensuring open source doesn’t get its say. Also OpenAI trying to adopt any other product (legal tooling, healthcare, latex replacement, etc…).
They were at the frontier because they invented many of the techniques. Google already had the internet’s data well beforehand. Anthropic and OpenAI are still at the frontier today, independent of legislation.
Oh no I agree, they were among the first to be popular in the llm frontier with their techniques. Today they are only holding on to the frontier because of that lead.
They want the legislature to keep their lead from other competitors.
I'm actually the most innovative, I've written thousands of papers advancing the state of human knowledge. They're just in my basement and I don't show anyone.
"Necessity is the mother of invention."
Not sure they'd be the same without the constraints.
This is literally just a scheduler over firecracker man what
380.000 concurrent sandboxes on 160 Epyc based server nodes. Crazy stuff
that's only 12 sandboxes per core
Um, how is that not impressive
if it's agentic stuff, they likely aren't hammering a core constantly and they will maybe sit idle quite often between model requests, so it makes sense. I do wonder how much memory they allocate to each one though.
it's just very efficient use of shared cores that is required to make these kinds of workloads cost efficient
If they sit mostly idle, you can swap out a lot of the memory to SSD, I guess.
A place I worked back around 2020 was running a Grafana instance per customer that got embedded on the web dashboard. We had 110 pods per GKE (kubernetes on gcp) 4 CPU node because that was a network imposed pod limit at the time. The nodes were usually idle--could have shoved a lot more on if not for the IP limit.
I think around that time Grafana changed their license tho so you couldn't host OSS Grafana as part of your service.
how is it impressive?
Will admit that I haven’t read it yet, just saw the crazy number of authors and think this may compete for one of the papers with the most authors.
This one [0] about the Higgs boson is hard to beat. Pages 25-32 are just full of around 5150 „authors“, pages 33-37 are their affilliations.
[0] https://arxiv.org/abs/1207.7214
Even without that particularly special case, high-energy physics collaborations have long since broken the idea of authors. There are now many collaborations with hundreds of authors publishing regularly and quite a few that are into the thousands.
I like it, represents science's general 'standing on the shoulders of our precedents' far more realistic then the 'genius solo' PR mythos.
Where is the balance between this and citations? Authors implies said person provided additions to get the work done.
With projects like the LHC, there really are thousands of people involved in "getting the work done."
I think the number is ~2930 authors according to the CDS [1] which is comparable to the corresponding CMS paper which has about ~2900 authors [2]
[1] https://cds.cern.ch/record/1471031
[2] https://impact.ornl.gov/en/publications/observation-of-a-new...
how do you even keep up with the amount of research coming out these days
So.. serverless ?
Still very impressive! Love how it's done!
Can you give a brief for what this is and why it is impressive?
It just describes the platform they built for scheduling workloads, and running those workloads. After a second thought it is not that impressive and probably doesn't deserve any kind of hype. It's the same kind of setup AWS is running for Lambda, as well as anyone else basically running SLURM clusters out there.
Still giving them credit because creating such as scheduler/platform from scratch is quite complex, and I know that they probably struggled a lot to get it right.
I'd say it's cool that they're openly writing about it and how they run their workloads. shows a nice window into how these things are actually deployed at scale.
it also shows how much density you can get easily from a single core if you wanted to replicate this
That's not quite right. A recurring trend in ML is figuring out how to get an elastic interface for the the particular quirks of ML workloads (source: I worked on an open source one years ago for inference). Recently, there's been a lot of interest in doing this for agent workloads. Google recently released something called ax that is similar in spirit. The core of it is this: https://github.com/agent-substrate/substrate
At a high level, agent sandboxes have peculiar needs. Agents are really bursty, but also long lived. You need low latency suspend/resume calls. Checkpointing and recovery have some particular considerations. And naive approaches are often really wasteful, but over optimizing without harming durability, isolation, or consistency in performance can get tricky.
This is the new hot infra topic for agent swarms, for the time being. Whether that's impressive or not is up to the reader I guess, but it's a pretty involved project nonetheless.
The person simplified the paper with such a misunderstanding that I thought 10x before rebutting. They clearly didn’t read anything of the paper. Thanks for the clarification.
That number of authors though
The topic isn't as interesting as how 131 authors communicated to get this out.
Is it possible they're doing a research lab "socialism style" and everyone gets equal credit for just being a part of the lab, regardless of actual input into the specific papers? If they're innovating in computer, maybe they're not so afraid of innovating in social/academic structures as well?
This is common practice in Biology labs in the west I think
Maybe there are even reasonable justifications for it, instead of yelling about "socialism".
Sorry, didn't mean it like "yelling", I'm quite a fan of socialism and don't see it as any negative. I didn't think anything in my comment was negative.
It's negative in the sense it suggest some authors didn't contribute (or contribute little0 to the research, and were perhaps just warm bodies that happened to be present at a certain location
The west does it too in experimental papers. This is not a socialism thing.
This is your brain on propaganda.
I don't understand why half the comments here are about the author list, this is very common practice in e.g. large-scale physics experiments and biology, and every new GPT release from OpenAI equally had papers with tons of authors
Agreed, I'm very confused as well. It's like no-one here has been paying attention to research papers.
Or they are just rushing to say anything, and it's much easier to comment on that than the content of the paper.
Some people don’t want to recognize the important work that went into this.
Is this like agent substrate?
It seems every DeepSeek paper/patent has a huge number of authors, and this one is no exception. They couldn't even fit everyone on the page, there are 31 others not shown. This could be an asset protection strategy (i.e., human assets). Imagine if there were only 3 authors. Those authors may get hired away by competitors. If you list every employee on every paper then competitors don't know who to lure away.
Just look for the “corresponding author”
The unfortunate intern who had to upload to arxiv?
Corresponding author is the head of the group, usually the one who coordinates the study, the one who knows everyone else. Is the "if you have any question about anything in this paper, contact me" person, even years after publication.
click the link, or read the PDF.
all authors are listed. there is no conspiracy to hide authorship.
arxiv simply want to keep the page short not too long.
Compare to https://arxiv.org/abs/2407.21783
Lmao with the conspiracy. Large scale experimental research is always like this, many papers in experimental physics have pages of authors.
Competitors will try to touch everyone on the list.
Yeah what possible company can try to lure 100 people... Just for reference, Linkedin has 17,000+ full time employees.
Maybe aquihire like Nvidia with groq?
just click the link and it will show the others, this seems to be a limitation/UI feature of arxiv. The paper itself contains the full list
I think they just meant that it's so many other people that they don't reasonably fit in the UI
Why is this the top comment? Many of the comments, as well as this one, have no relation to content and only mention a triviality
It’s top comment because it’s anti-Chinese.
I miss the YOLO (CNN computer vision model) days where one dude can publish a paper, completely disregard academic conventions, and yet push the field forward by leap and bounds.
I wonder if they are signalling that if they can do this for training, then they can create an style agent swarm to hack anyone with 380k concurrent agents.
Appears to be similar to what Google is building with ax https://github.com/google/ax
Impressive numbers!
Whoever would have thought (in prior years) that in 2026 AI Agents (not people or corporations, at least not directly) seem to be (or seem to be rapidly becoming) the biggest consumers of cloud computing resources...
Anyway, a very interesting paper and environment!
12 sandboxes per code is insane, I wonder how many of these sandboxes are idle at a time. Depending on the tasks assigned the resource requirements are different. Compare an agent doing pdf conversion and one responding to a simple question. One is cpu bound the other is mostly network wait.
This is an interesting problem from infra perspective since you cannot predict the workload. On a bigger scale you may get away with forecasts.
Im waiting for tech that elastically allocates cpu/mem without restarting a container.