3 ms·

DeepSeek isn’t a side project or just a bunch of quants - these are part of the marketing that people keep repeating blindly for some reason. To build DeepSeek

DeepSeek isn’t a side project or just a bunch of quants - these are part of the marketing that people keep repeating blindly for some reason. To build DeepSeek probably requires at least a $1B+ budget. Between their alleged 50,000 H100 GPUs, expensive (and talented) staff, and the sheer cost of iterating across numerous training runs - it all adds up to far, far more than their highly dubious claim of $5.5M. Anyone spending that amount of money isn’t just doing a side project. The client facing aspect isn’t the problem here. This linked article is talking about the backend having vulnerabilities, not the client facing application. It’s about a database that is accessible from the internet, with no authentication, with unencrypted data sitting in it. High Flyer, the parent company of Deep Seek, already has a lot of backend experience, since that is a core part of the technologies they’ve built to operate the fund. If you’re a quantitative hedge fund, you aren’t just going to be lazy about your backend systems and data security. They have a lot of experience and capability to manage those backend systems just fine. I’m not saying other companies are perfect either. There’s a long list of American companies that violate user privacy, or have bad security that then gets exploited by (often Chinese or Russian) hackers. But encrypting data in a database seems really basic, and requiring authentication on a database also seems really basic. It would be one thing if exposure of sensitive info required some complicated approach. But this degree of failure raises lots of questions whether such companies can ever be trusted.

crummy 2y ago
You think they deliberately left their DB open to the internet, without a password? Why?
- blackeyeblitzar 2y ago
  No, I did not claim that it was purposeful. But they did leave their DB open to the internet without a password. And that seems really negligent.
  - matt-p 2y ago
    For an ops person yes, for a ML engineer (basically an academic) I'd be more surprised if it was secured to be honest.
diggan 2y ago
> Anyone spending that amount of money isn’t just doing a side project. You're reciting a bunch of absolute numbers, without any sort of context at all. $5M isn't the same for every company. For example, in 2020, it seems High Flyer spent a casual $27M on a supercomputer. They later replaced that with a $138M new computer. $5.5M sounds like something that could be like a side-project for a company like that, whose blood and sweat is literally money. > But this degree of failure raises lots of questions whether such companies can ever be trusted. This, I agree with though. I wouldn't trust sending my data over to them. Using their LLMs though, on my own hardware? Don't mind if I do, as long as it's better, I don't really mind what country it is imported from.
- sailingparrot 2y ago
  > that could be like a side-project for a company like that, whose blood and sweat is literally money. From the mouth of Liang Wenfeng, co-founder of both High Flyer and DeepSeek, 18 months ago: "Our large-model project is unrelated to our quant and financial activities. We’ve established an independent company called DeepSeek, to focus on this." https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-frontier https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-fr...
  - ipaddr 2y ago
    It's a side project called DeepSeek .
- octacat 2y ago
  5.5M is a single latest training run, if they would rent gpu-s in the cloud - from the paper.
nicce 2y ago
> To build DeepSeek probably requires at least a $1B+ budget. Between their alleged 50,000 H100 GPUs, expensive (and talented) staff, and the sheer cost of iterating across numerous training runs - it all adds up to far, far more than their highly dubious claim of $5.5M. This is not fair. Is OpenAI, for example, including the CEO paycheck for the model training costs?
- Filligree 2y ago
  There's a sliding scale. On one end is "Include the CEO's paycheck"; on the other is "include nothing except the price tag on the final, successful training run". Neither end is terribly useful. Unfortunately, the $5.5M number is for the latter.
harrall 2y ago
Some academic projects have a lot of funding and what they are researching is some top tier stuff. But the software? Absolute disaster. When people say DeepSeek is a side project, this is what I assume they mean. It's different when a bunch of software engineers make something with terrible security because it's their main job. With bunch of academics (and no offense to academics), software is not their main job. You could be working on teaching them how to use version control.
fulladder 2y ago
>To build DeepSeek probably requires at least a $1B+ budget. Zero evidence that the above statement is true, and weak evidence (authors' claims) that it is false. Have you read their papers even? https://arxiv.org/html/2412.19437v1#abstract https://arxiv.org/html/2412.19437v1#abstract https://arxiv.org/pdf/2501.12948 https://arxiv.org/pdf/2501.12948
- sailingparrot 2y ago
  Parent is (I assume) talking about the entire budget to get to DeepSpeek V3, not the cost of the final training run. This includes salary for ~130 ML people + rest of the staff, company is 2 years old. They have trained DeepSpeek V1, V2, R1, R1-Zero before finally training V3, as well as a bunch of other less known models. The final run of V3 is ~6M$ (at least officially...[1]), but that does not factor the cost of all the other failed runs, ablations etc. that always happen when developing a new model. You also can't get clusters of this size with a 3 weeks commitment just to do your training and then stop paying for it, there is always a multi-month (if not 1 year) commitment because of demand/supply. Or, if it's a private cluster they own it's already a $200M-300M+ investment just for the advertised 2000 GPUs for that run. I don't know if it really is $1B, but it certainly isn't below $100M. [1] I personally believe they used more GPUs than stated, but simply can't be forthcoming about this for obvious reason. I have of course not proof of that, my belief is just based on scaling laws we have seen so far + where the incentives are for stating the # of GPUs. But even if the 2k GPUs figure is accurate, it's still $100M+
- tyre 2y ago
  H100s can cost about $30k. There was an interview with a CEO in the space speculating that they have about 50,000 H100s. That's $1.5bn. Presumably they got volume discounts, though given the export bans they might have had to pay a premium on that discount to buy them secondhand. If it were H800s, that would be ~half the price, which is still high hundreds of millions for the chips alone. Is that true? No idea. But there isn't zero evidence.
cma 2y ago
> Between their alleged 50,000 H100 GPUs I'm sure you were just mislead by all the people including Anthropic's Dario parroting this claim, but even Dario already said he was wrong to say that and semi analysis already clarified it was a misunderstanding of their claim, which was 50,000 H series, not 50,000 H100.
- Filligree 2y ago
  H800s, right?

crummy 2y ago

You think they deliberately left their DB open to the internet, without a password? Why?

blackeyeblitzar 2y ago
No, I did not claim that it was purposeful. But they did leave their DB open to the internet without a password. And that seems really negligent.
- matt-p 2y ago
  For an ops person yes, for a ML engineer (basically an academic) I'd be more surprised if it was secured to be honest.

diggan 2y ago

> Anyone spending that amount of money isn’t just doing a side project. You're reciting a bunch of absolute numbers, without any sort of context at all. $5M isn't the same for every company. For example, in 2020, it seems High Flyer spent a casual $27M on a supercomputer. They later replaced that with a $138M new computer. $5.5M sounds like something that could be like a side-project for a company like that, whose blood and sweat is literally money. > But this degree of failure raises lots of questions whether such companies can ever be trusted. This, I agree with though. I wouldn't trust sending my data over to them. Using their LLMs though, on my own hardware? Don't mind if I do, as long as it's better, I don't really mind what country it is imported from.

sailingparrot 2y ago
> that could be like a side-project for a company like that, whose blood and sweat is literally money. From the mouth of Liang Wenfeng, co-founder of both High Flyer and DeepSeek, 18 months ago: "Our large-model project is unrelated to our quant and financial activities. We’ve established an independent company called DeepSeek, to focus on this." https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-frontier https://www.chinatalk.media/p/deepseek-from-hedge-fund-to-fr...
- ipaddr 2y ago
  It's a side project called DeepSeek .
octacat 2y ago
5.5M is a single latest training run, if they would rent gpu-s in the cloud - from the paper.

nicce 2y ago

> To build DeepSeek probably requires at least a $1B+ budget. Between their alleged 50,000 H100 GPUs, expensive (and talented) staff, and the sheer cost of iterating across numerous training runs - it all adds up to far, far more than their highly dubious claim of $5.5M. This is not fair. Is OpenAI, for example, including the CEO paycheck for the model training costs?

Filligree 2y ago
There's a sliding scale. On one end is "Include the CEO's paycheck"; on the other is "include nothing except the price tag on the final, successful training run". Neither end is terribly useful. Unfortunately, the $5.5M number is for the latter.

harrall 2y ago

Some academic projects have a lot of funding and what they are researching is some top tier stuff. But the software? Absolute disaster. When people say DeepSeek is a side project, this is what I assume they mean. It's different when a bunch of software engineers make something with terrible security because it's their main job. With bunch of academics (and no offense to academics), software is not their main job. You could be working on teaching them how to use version control.

fulladder 2y ago

>To build DeepSeek probably requires at least a $1B+ budget. Zero evidence that the above statement is true, and weak evidence (authors' claims) that it is false. Have you read their papers even? https://arxiv.org/html/2412.19437v1#abstract https://arxiv.org/html/2412.19437v1#abstract https://arxiv.org/pdf/2501.12948 https://arxiv.org/pdf/2501.12948

sailingparrot 2y ago
Parent is (I assume) talking about the entire budget to get to DeepSpeek V3, not the cost of the final training run. This includes salary for ~130 ML people + rest of the staff, company is 2 years old. They have trained DeepSpeek V1, V2, R1, R1-Zero before finally training V3, as well as a bunch of other less known models. The final run of V3 is ~6M$ (at least officially...[1]), but that does not factor the cost of all the other failed runs, ablations etc. that always happen when developing a new model. You also can't get clusters of this size with a 3 weeks commitment just to do your training and then stop paying for it, there is always a multi-month (if not 1 year) commitment because of demand/supply. Or, if it's a private cluster they own it's already a $200M-300M+ investment just for the advertised 2000 GPUs for that run. I don't know if it really is $1B, but it certainly isn't below $100M. [1] I personally believe they used more GPUs than stated, but simply can't be forthcoming about this for obvious reason. I have of course not proof of that, my belief is just based on scaling laws we have seen so far + where the incentives are for stating the # of GPUs. But even if the 2k GPUs figure is accurate, it's still $100M+
tyre 2y ago
H100s can cost about $30k. There was an interview with a CEO in the space speculating that they have about 50,000 H100s. That's $1.5bn. Presumably they got volume discounts, though given the export bans they might have had to pay a premium on that discount to buy them secondhand. If it were H800s, that would be ~half the price, which is still high hundreds of millions for the chips alone. Is that true? No idea. But there isn't zero evidence.

cma 2y ago

> Between their alleged 50,000 H100 GPUs I'm sure you were just mislead by all the people including Anthropic's Dario parroting this claim, but even Dario already said he was wrong to say that and semi analysis already clarified it was a misunderstanding of their claim, which was 50,000 H series, not 50,000 H100.

Filligree 2y ago
H800s, right?