Posted on 01/28/2025 8:10:52 AM PST by MtnClimber
The recent release of DeepSeek's R1 model has sent shockwaves through the AI community and global markets. While DeepSeek claims to have achieved this breakthrough with minimal resources, some experts and analysts are raising concerns about the company's rapid ascent and potential ties to the Chinese government. This report delves into the evidence and expert opinions surrounding DeepSeek, steelmanning the case that it may be a psychological operation (psyop) orchestrated by the Chinese Communist Party (CCP).
China's Pursuit of AI Dominance
China has explicitly stated its ambition to become a global leader in AI by 2030. This goal is driven by the recognition of AI's potential to transform industries, enhance national security, and elevate China's global standing. China aims to achieve a first-mover advantage in AI, allowing it to set the trend and lead the world in this critical technology. To achieve AI dominance, China is actively pursuing the following strategies:
- Investing in AI research and development: China is promoting AI education and research programs to cultivate domestic talent and reduce reliance on foreign technology.
- Developing domestic AI infrastructure: China is investing in advanced technologies like 5G networks and green data centers to support the deployment of high-capacity computing power essential for AI development.
- Prioritizing AI applications across sectors: China is actively integrating AI into various industries, including healthcare, finance, and surveillance, to drive efficiency and innovation.
- Monitoring global AI trends: China closely monitors global AI developments to identify opportunities for breakthroughs and strategically allocate resources to maintain a competitive edge.
DeepSeek's Suspicious Success
DeepSeek, a Chinese AI startup founded in 2023 by Liang Wenfeng, has rapidly gained attention for its advanced AI models, particularly the recently released R1 model. Liang, born in Guangdong in 1985, holds bachelor's and master's degrees in electronic and information engineering from Zhejiang University. He also runs a hedge fund, which could potentially be funding DeepSeek's operations.
DeepSeek's R1 model has demonstrated impressive performance on various benchmarks, rivaling and even surpassing established models from US tech giants like OpenAI and Google. DeepSeek claims that R1 offers performance on par with OpenAI's latest model and ranks among the top performers on the UC Berkeley-affiliated leaderboard called Chatbot Arena. However, DeepSeek's claims of achieving this breakthrough with limited resources and in a short timeframe have raised skepticism.
DeepSeek claims to have developed the R1 model in less than two years with only $5.6 million, using less advanced Nvidia H800 chips . This claim is particularly striking considering the significant investments made by US companies in AI development. For instance, the following table shows the number of H100 chips purchased by leading US tech companies in 2024 alone:
DeepSeek's approach challenges the traditional capital-intensive model of AI development. The company emphasizes research and employs an open-source strategy, potentially as a way to gain a competitive edge . DeepSeek also claims that it does not engage in external project cooperation or provide privatization deployment, which could be interpreted as a way to avoid scrutiny and maintain control over its technology.
DeepSeek's development timeline includes the release of DeepSeek Coder in November 2023, followed by DeepSeek LLM, DeepSeek V2, DeepSeek V3, and finally DeepSeek R1 in January 2025. The R1 model was trained using multi-stage training and large-scale reinforcement learning techniques. It employs a Mixture-of-Experts (MoE) architecture with 671 billion total parameters, but only 37 billion are activated per task. The training process involved curated datasets, R10-generated reasoning chains, and human preference alignment . A reward system was used during training, which included accuracy rewards and format rewards.
DeepSeek R1 has achieved impressive results on various benchmarks, including:
MATH 500: 97.3% accuracy
Codeforces: Ranked in the 96.3rd percentile
AIME 2024: 79.8% pass@1 score
MMLU: 90.8% accuracy
GPQA Diamond: 71.5% accuracy
SWE-bench Verified: 49.2% accuracy
AlpacaEval 2.0: Won 87.6% of evaluations
Despite these achievements, DeepSeek R1 has some limitations, such as mixing English and Chinese responses due to a lack of supervised fine-tuning....SNIP
Deep Six.
Psyop and propaganda.
And short sell vehicle.
DeepSeek has 50,000 Nvidia H100 chips.
that sounds like a lot of money and power, the h100 card I am using cost $27k and uses 350watts
I’m surprised at how quickly this new AI has gained broad acceptance. I personally don’t think that it is feasible to develop such a thing at this low a cost. They don’t get paid as well as American engineers for one thing. Also they are not above lying.
Thought something similar when I first heard about this yesterday.
It’s a classic tactic in technology wars, to “freeze” product-adoption decisions by potential users. It occurs when a new technology is released, and is getting a big buzz and a lot of investment, AND when the slower market-player is way behind the curve with their own offering. The A.I. marketplace is exactly in this state today.
If successful, China hopes they can cause potential A.I. users to defer purchase decisions and their own development and integration investments, waiting for the (now) expected release of China’s product.
Common components when using this tactic are to announce that your own product is as good as, if not better than the current market leader, that it will be available very soon, and that it will come at a much lower cost.
To me, this whole charade is China effectively announcing that they agree A.I. is going to be huge, that they’re going to be a player, and that they’re behind in the A.I. “arms race”.
I’m deeply skeptical as well. Let’s wait and see if it’s for real.
What can you tell us about it?
Less W than I would have expected given the 5000 series' needs.
“””Anyone who uses deepseek and rednote are as good as spies for the Chinese.”””
I suppose you are referring to apps? We have been messing with deepseek sandboxed for a bit and we are not “spies”. It’s open source.
sure it is, all Chinese stuff is perfectly safe. They are our good friends and would never do anything to hack us.
“””sure it is, all Chinese stuff is perfectly safe. They are our good friends and would never do anything to hack us.”””
Huh? Sarcasm, I presume. I was referring to apps. For example, I’d never use TikTok.
DeepSeek vs ChatGTP for code building as we have compared them recently isn’t a fair fight. DeepSeek is pretty cool. Note that DeepSeek is open source so putting it into a sandbox and messing with it isn’t a big deal.
FYI, I built one of the first large Intel based systems decades ago. Experimental cards, etc. I found Linus very early on and ported a number of things from commercial Unix like SCO to Linux. I’ve been running Linux servers 24/7/365 for over 3 decades. The list goes on. I’ve got some real nerds around me and they are telling me that DeepSeek is a game changer from a developers standpoint vs anything else like Chat, etc.
As far as the Chinese go all I can say is “thanks for the new free toy”.
“””DeepSeek has 50,000 Nvidia H100 chips.”””
The notion that DeepSeek is a negative for NVDA for some reason seems a bit odd considering that. We bought more NVDA today.
Very much sarcasm. I can see everyone world wide create nifty neato code with DeepSeek that will go directly into the hands of the Chi-Coms. Anything out of China should not be trusted, sandboxed or not.
“””I can see everyone world wide create nifty neato code with DeepSeek that will go directly into the hands of the Chi-Coms. Anything out of China should not be trusted, sandboxed or not.”””
1-2-3-4-5-6-7-8-9-10
Disclaimer: Opinions posted on Free Republic are those of the individual posters and do not necessarily represent the opinion of Free Republic or its management. All materials posted herein are protected by copyright law and the exemption for fair use of copyrighted works.