Mint Primer | DeepSeek: A Chinese marvel or OpenAI copy?

Metaverse

Mint Primer | DeepSeek: A Chinese marvel or OpenAI copy? – Crypto News

Published

3 weeks ago

January 31, 2025

Dripp

DeepSeek has challenged big tech, proving AI can be efficient without costly graphics processing units (GPUs) or massive data centres. Now Microsoft is probing possible unauthorized access to OpenAI data by a group linked to DeepSeek. Are Chinese AI models just illegal copies?

Can you explain this fuss over DeepSeek?

AI lab DeepSeek shocked big tech by training its open-source R1 model on Nvidia’s lower-capability H800 chips for under $6 million—far less than the billions spent on OpenAI’s ChatGPT or Google’s Gemini. OpenAI CEO Sam Altman called R1 “impressive”, while Anthropic CEO Dario Amodei noted its near-frontier performance at low costs. Former Intel CEO Pat Gelsinger chose R1 over OpenAI for his startup, Gloo. DeepSeek’s low-cost, energy-efficient, open-source AI could democratize access, challenging Microsoft, Google, Meta and Nvidia while proving advanced AI can be built without a huge outgo.

Are there challenges to Chinese AI models?

There is deep-seated mistrust of Chinese products, services, companies and tech platforms in the US. Amodei argues DeepSeek has shown a natural step in cost cuts. But since a Chinese company led this advance, it caused a stir. Investor Gavin Baker noted that the $6 million cost excludes prior research, and such training is only possible if a lab has invested hundreds of millions and has access to large computing clusters. Microsoft is probing possible unauthorized access to OpenAI data by a group linked to DeepSeek. Ironically, OpenAI itself is facing copyright violation suits in the US and India too.

How are other AI models from China faring?

Alibaba’s Qwen handles long inputs but needs high memory. Baidu’s Ernie Bot 4.0 excels in search but lags ChatGPT in performance. ByteDance’s Doubao 1.5 Pro is strong in social media AI but lacks adoption. Moonshot AI’s Kimi k1.5 shows promise but is still developing. Despite China’s AI progress, challenges remain in scaling, originality, and global competition.

What’s the upshot of the charges?

OpenAI suspects DeepSeek distilled its advanced models into a smaller, cheaper version without permission. Distillation implies that DeepSeek may have used OpenAI’s outputs as “teacher” data to train its own AI, reducing costs and development time—violating OpenAI’s licence terms, and raising concerns about originality, ethics, and intellectual property (IP) rights. Ironically, R1 is now available on Microsoft’s Azure AI Foundry and GitHub, also owned by the Redmond giant that has invested $10 billion in OpenAI.

What’s India’s stance on Chinese AI models?

Despite banning TikTok and restricting Huawei, India has plans to host DeepSeek on local servers and invite proposals for ChatGPT-like models. India announced AI infrastructure plans using 10,000 GPUs via public-private partnerships, though US chip export restrictions are a challenge. This, even as Altman may visit India in February. DeepSeek’s R1 proves AI can be built affordably, much like how Altman’s efforts encouraged Indian startups to develop small language models for under $10 million.