Notes on DeepSeek
twitter.com75 points by vinhnx 3 hours ago
75 points by vinhnx 3 hours ago
Post appears to have been removed, I caught a copy of it: https://pastebin.com/rcAqEFG1
I assume it will get reposted at some point.
thanks. this really isnt that long, might as well paste in full here since OP deleted.
Notes on DeepSeek:
We visited the company HQ last Tuesday. It was founded in 2023 by Liang Wenfeng and operated out of his hedge fund, High-Flyer, until somewhat recently. The company released their R1 model in January 2025, so it was interesting to see what they’ve been doing
The company is located in an unmarked, 12-story building in Hangzhou. There is no DeepSeek branding visible from the street or lobby. I asked why this is, and the team demurred and said, “Well, there are many companies in this building, and we are not special.” They want to keep a low profile.
We met with their Head of Data and Head of Infrastructure. The company only has 300 employees. They are at least an order-of-magnitude smaller than Anthropic, and don’t care to scale further just yet. Their Head of Infrastructure, in particular, was young; maybe 30 years old and apparently one of the best AI buildout and energy experts in the country. (We briefly walked through the labs, and everybody seemed young. There was a lot of discussion; it felt like an exciting and energetic place.)
Lots of competition is coming from Alibaba (Qwen), ByteDance, and Moonshot (Kimi). People in China seem to mostly use Kimi or Deepseek. Young people use VPNs to access Claude, though Anthropic has blockers around usage in China and make it difficult. Poaching between groups is common, just like in the U.S. DeepSeek has a reputation as being really smart and “cool,” maybe similar to Anthropic. Big labs are mostly in Beijing, near Tsinghua and Peking University, with Hangzhou as the main exception (DeepSeek and Alibaba/Qwen are there).
The DeepSeek team reads western AI writers. They listen to Dwarkesh and read Gwern. The people we met with said they had never met with any employees from Anthropic. They were not at all concerned with some kind of hostile / AGI takeover scenario. They kept bringing up job loss (which is already high amongst youth in China) as their main concern. When we asked if they do red teaming on their models, they said no. In China, AI models are not regulated directly; the government instead has restrictions on how those models can be used in software, services, etc.
As a whole, China seems to treat AI as just another technology, rather than as some kind of singularity moment. National attention is still on basic needs and infrastructure buildouts, and on providing more medicines for people. The “dreams of singularity" seem like a luxury or distant consideration.
We asked the DeepSeek team: “What has the highlight been so far? What are your plans for an exit?” And they said that their highlight and great achievement was R1. They did not gesticulate at a future model or vision, but rather seemed proudest of what they’ve already done. They are content for now to remain ~6 months behind U.S. companies while maintaining a lower profile and team size.
I don't get the part of "AI models are not regulated directly, the government instead has restrictions on how those models can be used in software, services". Is it not the same thing? When I chat with DeepSeek about any (Chinese) political/social issue, it immediately begins aligning with the party's line or just cut off the conversation abruptly.
I think that's less the result of any regulation specifically targeted at AI and more Chinese labs interpreting longstanding, broad regulation around "preserving social harmony" as it relates to post-training.
They listen to Dwarkesh
oh jesus, that guy and his absolute baloney, empty interviews.... sigh.
Not sure what I read, but sounded like a lunch meeting description; felt void of actual information, with the restaurant replaced by the office. I am in China and can tell it is either Kimi, DeepSeek or Claude (proxied or actually deepseek/fake). The bigger push for the general public died down a lot since last year; kids were pushed to use AI for homework, now it is disallowed and frowned upon. In short mixed messaging.
With government billions fund pushed for AI build out, fast pace integration on large scale and sweeping national education reform for AI, I don't think it can be called "died down".
[0] https://www.reuters.com/world/china/china-prepares-295-billi...
[1] https://www.globalneighbours.org/en/articles/china-unveils-n...
[2] https://english.www.gov.cn/news/202606/10/content_WS6a296017...
> kids were pushed to use AI for homework, now it is disallowed and frowned upon. In short mixed messaging.
in the early 2000s in california universities you'd get marked down for citing wikipedia. so the good souls told everyone "see the number in brackets[2] after what you're trying to cite the article for? just click that then click the archive.org or whatever link there, then cite that."
Now? i think wiki is considered a valid source? or has it flopped back to being "unreliable"?
It's not that it's unreliable, it's just lazy research. Wikipedia, like all encyclopedias, is a tertiary source, but ideally your essay should be a mix of primary and secondary sources, while Wikipedia discourages original research and prefers only secondary sources. Wikipedia itself recommends against citing it as research[0] for this reason.
[0]: https://en.wikipedia.org/wiki/Wikipedia:Citing_Wikipedia
Laziness should never be the issue.
The issue is that Wikipedia can be wrong and you’d only know that by going to the source (or lack thereof), or checking other sources.
I think Wikipedia's still considered unreliable, but the question that should be asked is whether the author even read the source in "the number in brackets" to ensure that it's even backed properly.
Just like how people should use AI for research, I guess.
Were things like "300 employees" and descriptions of the deliberately low key hdq out there before? That counts as actual information to me.
It’s a puff piece written by someone who didn’t know (or didn’t care) they were being managed.
From the notes, they seem humble and empathic.
We're lucky to have China imposing competiton to the western AI megacorps.
If it wasn't for China, I would probably have to spend $100/mo on AI instead of $10 like I do currently while using DeepSeek and MiMo (opencode Go plan).
And while I could do so comfortably, I feel for those who can't. It must feel incredibly isolating to only watch others have access to expensive models to leverage their careers.
I hope SoTA AI becomes an universal right because it will contribute to too much income disparity otherwise.
> We're lucky to have China imposing competiton to the western AI megacorps.
The second they get a hold of the market, Chinese Big Tech will be as bad or worse than US Big Tech.
We're lucky to have DeepSeek.
In every market China dominates, Chinese products are still inexpensive. Solar panels, batteries, EVs, drones,..
Because they are subsidized by the Chinese government. This is literally a tactic to destroy global competition.
It's a smart move to make everyone dependent on them.
I see this problem already for me.
I have unlimited tokens at work than i go home what do i do? Spend 200$ per month? No def not.
When Anthropic increased the limits for their 20$ plan, i started again coding with it on a private project and it was fun and i did a lot in that 4 weeks.
Yep. After yesterday's moves around "Fable 5" even twice as much.
We've had a taste, and damned if I'm going to have the "means of production" snatched from me already?
approximately how many months/years until there are "illegal models"?
I remember reading a similar tweet explaining DeepSeek breaks the insane Chinese work culture. They are against 996 and brutally grinding employees. They feel like a big family and that is their hedge against poaching by Chinese Big Tech with bigger salaries. Liang Wenfeng seems to be the only AI CEO down to earth. I want to believe.
Why would the agent send the results of the query "Show me my recent transactions" to LLM? This pretty deterministic results which involve no LLM interpretation or decision making.
"As a whole, China seems to treat AI as just another technology, rather than as some kind of singularity moment."
This is a refreshing perspective.
The CCP is very active in the matter of AI. In fact, the DeepSeek moment was responsible for Xi calling for a private meeting with tech bosses, including the exiled Alibaba founder Ma. Which is practically unheard of in China politics.
I don't have enough information to say whether the Chinese leadership sees AI "just as the next technology" or they are more cautious due to its double-sword nature. But the immense efforts for building their own AI/GPU chips plus government's billions fund pushed for AI build out, a directive for fast pace integration on large scale and a sweeping national education reform for AI, I don't think it can be seen as similar to other ordinary techs.
[0] https://www.reuters.com/world/china/china-prepares-295-billi...
[1] https://www.globalneighbours.org/en/articles/china-unveils-n...
[2] https://english.www.gov.cn/news/202606/10/content_WS6a296017...
"National attention is still on basic needs and infrastructure buildouts, and on providing more medicines for people. The “dreams of singularity" seem like a luxury or distant consideration."
Further on. Refreshing indeed.
China is probably more capitalist in many respects than the west these days. AI, robotics and automation is a way to push into the future. In the west we have endless researchers stuck in a psychosis that they are talking to a sentient being.
The CCP knows, whatever the heck this technology will bring with itself, the current power dynamic inside of the country is on their side, and AI will solidify it.
I hypothesize that, rather than slowly having it disperse in society and allow people to harness it in ways they don't want, they might as well accelerate everything until AI becomes the totalitarian swiss knife - which they can make use of in the best way of course.
Let's see what will happen.
US used AI (Claude on Maven) to determine a girl's elementary school as a target in war[0] and then triple tapped it and you're still more worried about hypothetical misuses of the single country responsible for this technology not being concentrated in the hands of a few powerful elite? ffs
[0] https://www.washingtonpost.com/national-security/2026/03/11/...
I don't really see how open weights models further what you're talking about.
It's trivial for me to download one of their models and run it on my Spark, and there's all sorts of ways to strip out their Tiananmen-denialism or whatever.
If/when the memory price crunch dissipates, even more so. And so far it's only China I see as making moves to increase production capacity on memory, too.
If anything the centralization of capital into US-based Anthropic and OpenAI is far more terrifying from the perspective you're outlining.
Especially here on HN, where AI anxiety (especially amongst those that are really nervous that it needs to succeed) is very, very tiresome.
It seems the OP has removed the tweet somehow.
US AI is almost a religious cult. It's devastating that they are treating it as a petty commodity
Altman used to talk about making a religion and Dario Amodei constantly talks about "building a God" and meets with religious leaders including the Vatican.
> It got me thinking, though--the most successful founders do not set out to create companies. They are on a mission to create something closer to a religion, and at some point it turns out that forming a company is the easiest way to do so. [1]
I would argue the US providers have gone full tilt into sales culture with respect to AI. Anything is said on a whim to redirect attention back from whomever is in the limelight. Initially I thought Anthropic was more pragmatic, but the constant release cycles of things that don't exist for most people, the gatekeeping, the statements made by Dario, it's all a part of large brand toxic sales and marketing.
From the notes this part sat with me as the real difference:
> As a whole, China seems to treat AI as just another technology, rather than as some kind of singularity moment. National attention is still on basic needs and infrastructure buildouts, and on providing more medicines for people. The “dreams of singularity" seem like a luxury or distant consideration.
Meanwhile... In the fantasy land over here in the US we're constantly being told that it's "coming", "almost here", "too powerful for us to give you access to", "of national security importance!". Or... FUD.
And while there may be trace amounts of truth in those overzealous statements we haven't seen a significant improvement in much outside of software development comparative to the spend and environmental impact.
"Their Head of Infrastructure, in particular, was young; maybe 30 years old and apparently one of the best AI buildout and energy experts in the country"
Expert in buildout or expert in distillation?
What's wrong with distillation? Wasn't GPT a distillation of the world's internet? That's how technology levels proceed, by recursively consuming the previous ones.
It's absolutely mind boggling to see claims of model distillation being theft, a class of attack, and all sorts of claims all the while Meta is in court for copyright violation, anthropic has had to settle a case with authors. With distillation "attacks" at least they paid API fees.
Anthropic had to settle with authors because they literally pirated books! Their behavior regarding distillation is genuinely beyond parody.
There are 2 things worth separating.
1) China distills and is therefore morally bad.
As you rightly point out, that's not a great argument.
2) China distills and is therefore possibly not that competent.
I think that makes sense. If they only catch up to the frontier through distillation then 1) Their model will never be as good as the model they are distilling from. 2) They will never reach the frontier - they need someone else to do it first.
Deepseek models are on the Pareto frontier of cost/performance. Thats the far more important one than just making a top scoring model.
This is literally a repeat of the whole “China only make low quality cheap stuff” argument.
“All they do is copy.”
And now, oops they are world leaders in EVs, batteries, solar, drones, just to name a few on the biggest consumer facing things.
"Success leaves clues"
You gotta start somewhere and you can start at page 1 or page 10 and that time, energy and cost you saved starting 9 pages later can be put into making whatever it is you're building better than the original.
The US, and every other country, is full of derivatives or straight up copies. No one is getting super mad at the generic cheerios at the grocery store. It's hypocrisy.
>2) China distills and is therefore possibly not that competent.
I think deepseek at least has done enough innovative work that you could grant them a baseline of competency.
In general, there are enough papers coming out of China to suggest that there are quite a few people there who know what they are doing.
You're correct and I shouldn't have used the word competent. Perhaps "and is therefore not elite enough to be state of the art"?
I also have a soft spot for deepseek because they write such readable papers. I don't have a degree in anything but with a little work I can understand their papers - which I really appreciate.
But I still think my point stands - if you need distillation you won't be SOTA
> China distills and is therefore possibly not that competent.
I heard that argument more than one year ago, when chain of thought and reasoning cycles started to be hudden to protect against distillation.
Meanwhile, models as DeepSeek and MiMo are nothing short of excellent nowadays.
Ever since I switched away from OpenAI to DeepSeek I never felt the need to go back.
Deepseek Flash V4 really was a "holy shit" moment and deserves the praise/hype it's been getting from users. I have a multi-tier subscription strategy I've maintained for the last year of: 1. $20-$30 plan from first Claude now Codex for "SOTA" 2. Gemini via the extra $10/mo or so from my Google One plan 3. a cheap fallback plan.
Together it gives me plenty of head room/model performance for $40ish/mo, plus letting me compare the various models over time.
Originally I'd been using the Z.AI plan (that I'm still grandfathered into for <1 yr) as my cheap plan but wasn't keeping up with the SOTA progress and is slow/limited now. So I subscribed to the Opencode Go plan and use Deepseek Flash V4 almost exclusively and it is insane how much usage I can get for $10/mo.
I did the math on my Flash usage vs. what I'm paying Opencode and I'm typically not even exceeding $10 in API costs! So it's actually sustainable not rugpull pricing at least for me. I can pound it with requests/agentic loops and have it running for 30 min doing whatever the fuck and check back and have spent literal pennies for what would have cost $30+ on my work's Github Copilot plan.
I know enterprise world works under different rules and isn't price sensitive in the same ways as an individual but I truly don't see how this is sustainable for the US AI giants in the long term to maintain like 25x+ markup for 1.25x performance benefit.
IMO it does help explain the recent emphasis on secret, scary "super models" like Mythos to muddy the waters for decision makers with hype and FOMO at at time when companies are beginning to seriously scrutinize their token spending for the first time.
Tell me, where did OpenAI and Anthropic got their training data? From public sources using legitimate means? Don't make me laugh.
Both. Both are good. Anyway this shows how full of shit Anthropic are - if Mythos was so advanced as they claim - distillation attacks just wouldn't work.