The Hated One
patreon
DeepSeek has a secret agenda
🕑 Added 2025-02-05 17:11:36 +0000 UTCDeepSeek AI is an AI with an agenda.
Here’s a set of AI benchmarks. They measure performance of large language models and weigh them against them each other. Here’s how a high-end model called OpenAI-o1 performs on the various scales. And here’s DeepSeek-R1, a novel model developed by a Chinese startup. In three out of six benchmarks, DeepSeek-R1 tops OpenAI-o1. Sometimes barely, sometimes squarely. But definitely bursting the bubble, that the US big tech is years ahead of AI competition. [0]
Yet, that’s not a paradigm shift. What is a paradigm shift, is that DeepSeek managed to build such a powerful AI using only 2,000 second-rate chips from Nvidia. [1]
In perspective, US big tech uses as many as 16,000 GPUs. According to DeepSeek engineers, the total cost of compute power to build their AI was $5.5 million. [1, 2]
By comparison, the cost of training GPT-4, OpenAI’s flagship model, ran between $41 to $78 million. Google spent up to $191 million to train Gemini. [3]
Now we are in a sort of “Tony Stark built this in a cave with a box scraps” situation. Either that, or DeepSeek is lying about their true costs. But we are about to find out about that very soon. This is why.
This is the Copyright license under which DeepSeek published their model. [4] It’s short and straightforward. It’s a standard MIT License. The most important line is this: anyone with a copy of DeepSeek AI is allowed to “use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software” without any restrictions. In other words, DeepSeek is a complete Open Source AI model. [5, 6]
So anyone, including you, is allowed to take a copy of this AI, make any modifications to it whatsoever, and redistribute it to whomever, however you want. DeepSeek essentially gave away their high-end AI that beats OpenAI and the rest of the big tech and they gave it away for free. Thus proving there is little incentive to spend so much money on expensive US tech when you can build it at the fraction of the cost.
Now open source AI is not a new thing. Meta famously open sourced all of their Llama models [7]
challenging the proprietary status quo that was propagated for by Microsoft, OpenAI and billionaire philanthropies of effective altruism. [8, 9]
But Zuckerberg’s Llama AI was still expensive to make so even if he gave it away for free, it was still believed to be too difficult for a company without a trillion-dollar evaluation to afford competing with Meta. [10, 11]
DeepSeek completely flattened that idea. And it’s not just because they managed to compete with big tech at 1/10 of their cost. If everything DeepSeek said about its engineering is true, then not only did they prove it’s possible to build AI much more cheaply, they also gave everyone a blueprint on how to do it.
DeepSeek published a 22-page-long research paper, detailing exactly how they managed to scale the performance of their AI while keeping the costs low. [12]
This doesn’t just put China at the top spot in AI research. It permanently dethrones US status as a tech leader which will now have to share that status with at least one if not many more countries capable of replicating what China did with DeepSeek.
If what DeepSeek says about its research and engineering is true. Because many industry experts looking at their research and results initially thought that somehow DeepSeek must have cheated. [13]
That either they used a lot more GPUs or used Nvidia H100 chips which they weren’t supposed have due to US-imposed barriers. It’s also possible that all of this is one big Chinese government propaganda to undermine American tech industry. Or maybe DeepSeek took one of Meta’s open source models and just put their own skin on it. [13]
But unlike with the claims the US big tech is making, speculations around DeepSeek are actually testable. [14]
Microsoft, Google or OpenAI don’t allow the public to see what’s in their AI and they are definitely making some bold claims about what they can do. [15,16]
DeepSeek may also sound bold but at least we can test them. And that’s why DeepSeek claims about its results and costs are at least mostly true. And mostly true is good enough to totally deligitimize the big tech’s multi-billion-dollar spending on AI. Trump recently announced a $500 billion joint venture with Oracle, OpenAI and SoftBank, which sure as hell looks embarrassing now against DeepSeek AI, even if their true costs were ten times higher than originally stated. [17, 18]
It’s not likely that DeepSeek’s true costs are that much higher though. The company is run and funded by Chinese hedge-fund company High-Flyer, which has $7 billion in assets under management. [19]
That’s sounds like a lot, but it’s a smaller amount of money than big tech companies spend on their AI divisions and AUM is not their money, but their clients’ money. So it’s most likely true that DeepSeek’s budget is only millions and not billions. And that says a lot.
But DeepSeek is more than just an AI. It’s an agenda. Its job is much more than just to compete with American AI companies. Why would they just give away all of their tech, research and methodology as open source? They are literally giving US big tech the ability to build on top of them. So what does China really want to achieve here?
In 2017, the Chinese government made it its priority to become the world leader in AI by 2030. It then told Chinese companies to achieve major breakthroughs and become “world-leading” by 2025. [20]
It is 2025 and DeepSeek is now the world leader. Literally. According LM Arena, it sits at a comfortable top spot as the only open source model in a sea of proprietary AIs. [21]
DeepSeek app hit the number one spot on Apple’s App Store as the most downloaded free app, dethroning ChatGPT. [22]
Mission accomplished? It’s far from over. This is not something that was officially stated in public, but DeepSeek is on track to dominate AI research not in house, but globally. And it’s precisely because it went open source.
Because it’s the only openly available model that is just as powerful as proprietary ones, research teams, AI companies and engineers around the world will now flock to use DeepSeek AI and build more tools on top of it. [20]
Many new breakthroughs and improvements will spawn out of DeepSeek open source AI. It will essentially become a platform. [23]
It is the exact same formula Google went for with Android.
Android, being built on top of Linux, is also fully open source. Most people don’t know this, but you are actually free to take Android and install it on whatever device you want and even sell it if you want to. [24]
It might seem counterproductive, but for Google this paid off tremendously. Even with the iPhone overtaking Android in a few markets, iPhone will never be more popular than Android. That’s because Android runs on literally everything. It’s not just on Google phones. It’s on every single phone that is not the iPhone. It’s on every smartwatch, smartTV, even cars run Android. Google is happy to give it away for free, because they correctly estimate that most Android devices will run Google apps and services on top of it. Any Android device will almost certainly have the Google Play Store, Google Search, Gmail, YouTube, Google Assistant and many many other Google apps. Google made this decision 20 years ago, way before anyone knew the concept of a smartphone would take off. But it did. And Google became a global behemoth because of it.
We are at the dawn of generative AI. And it’s not gonna be the AI that sells and makes money. It’s gonna be the apps, services and tools built on top of AI that will pay for the bills. OpenAI, Google and Microsoft are all building those AI tools and services. They are all competing to become a new platform but they are all proprietary, closed off walled gardens much like Apple is. But Meta and now China are trying a different strategy – they want to become a platform by allowing everybody in with no restrictions. And now China has an open source AI that’s better than any other open source AI. [25]
Will it win the race to AI domination?
Maybe. If it doesn’t, at least it achieves one thing – it fragments the AI ecosystem enough to prevent US hegemony in the industry. And for China, that’s better than having the US rule them all.
China’s AI doesn’t come without caveats though. With all its openness and lack of restrictions, China has a second agenda once it reaches mass adoption.
It is mandatory in China, that all large language models are to be tested by officials from the Cyberspace Administration of China to ensure they “embody core socialist values”.
These officials go to every Chinese AI company and check their data sets and safety processes against a range of questions and benchmarks. The government then mandates that AI companies censor topics that go against these “core socialist values”, such as “inciting the subversion of state power” or “undermining national unity”. [26]
There are things like blanket bans on callouts to Xi Jinping or Tienanmen Square. Their AI models are also trained to give politically correct answers to questions about China’s record on human rights or about its political leadership. And they are also told to reject not more than 5% of questions. [26]
The goal of all of this to build a socialist AI. [26]
An AI that will be censored much like proprietary LLMs from the US are censored, but in the opposite direction – so that they are not challenging Chinese authority and the Communist party.
This is hilariously visible in DeepSeek AI also. When you ask it any of these fiery topics related to China, it might straight out refuse to answer or avoid critical facts. It doesn’t hiccup nearly as much as when you ask it similarly controversial questions about authorities in other countries.
This becomes that much more significant if China’s AI becomes a global platform. Chinese government would love nothing more than to export its censorship power and party propaganda abroad. For now, much of the Chinese tech has been isolated from the rest of the global Internet due to the Great Firewall and trade wars. But releasing their tech as open source is a clever move. There is no reason to use expensive American AI, when you can use Chinese AI that’s free and just as good. China is hoping to build a momentum for mass adoption. And so far they are succeeding greatly.
And let’s be really honest about something here. Open sourcing tech is not just pure altruistic giving away of your tech and hard work. In the long run, you get many researchers, universities, businesses and individuals using your tech, improving upon it and making contributions. The original authors of open source benefit equally from all of those contributions made on top of their work, without having to pay for any of them. So if mass adoption is there, contributions will come too. Chinese open source AI will win. The big tech has lied to you and they convinced the US government proprietary AI is the way. Well it might not be now.
SOURCES
[0] https://github.com/deepseek-ai/DeepSeek-R1/blob/main/figures/benchmark.jpg
[1] https://www.nytimes.com/2025/01/23/technology/deepseek-china-ai-chips.html
[3] https://www.forbes.com/sites/katharinabuchholz/2024/08/23/the-extreme-cost-of-training-ai-models/
[4] https://github.com/deepseek-ai/DeepSeek-R1/blob/main/LICENSE
[5] https://github.com/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-CODE
[6] https://github.com/deepseek-ai/DeepSeek-V3/blob/main/LICENSE-MODEL
[7] https://fortune.com/2023/07/18/mark-zuckerberg-meta-ai-open-source-llama-2-llm/
[8] https://www.politico.com/news/2023/10/13/open-philanthropy-funding-ai-policy-00121362
[9] https://www.politico.com/news/2023/12/03/congress-ai-fellows-tech-companies-00129701
[12] https://arxiv.org/pdf/2501.12948
[13] https://www.nytimes.com/2025/01/28/technology/china-deepseek-ai-silicon-valley.html
[14] https://huggingface.co/deepseek-ai/DeepSeek-R1
[16] https://www.ft.com/content/2dc07f9e-d2a9-4d98-b746-b051f9352be3
[17] https://www.cnbc.com/2025/01/27/chinas-deepseek-ai-tops-chatgpt-app-store-what-you-should-know.html
[18] https://www.nature.com/articles/d41586-025-00259-0
[19] https://www.cnbc.com/2025/01/27/chinas-deepseek-ai-tops-chatgpt-app-store-what-you-should-know.html
[20] https://www.nature.com/articles/d41586-025-00259-0
[21] https://lmarena.ai/?leaderboard
[22] https://www.nytimes.com/2025/01/27/technology/deepseek-ai-chatbot-first-impressions.html
[23] https://www.wired.com/story/meta-ai-llama-3/
[24] https://source.android.com/
[25] https://github.com/deepseek-ai/DeepSeek-V3/blob/main/figures/benchmark.png
[26] https://www.ft.com/content/10975044-f194-4513-857b-e17491d2a9e9
Credits: Music by Carl Kasey youtube.com/@WhiteBatAudio
