4 ms·

Huawei's star AI model was built on burnout and plagiarism

26 points by avervaet 1y ago

beambot 1y ago
So... the same things as every other foundational model?
echelon 1y ago
What if there truly is no moat? What if you can just hook your AI system up to some other AI system and drain everything? No weights access required. Just train on the raw inputs/outputs. What stops this from being the future?
- refulgentis 1y ago
  Isn't that today? (or, perhaps more pedantically, yesterday and today and tomorrow?)
  - echelon 1y ago
    ... and tomorrow, and the next day. It never ends. There's no moat. One day your at-home GPU will unwind an entire hyperscaler's worth of expertise. Does the capital outlay get them anything at all apart from a temporary lead? While you couldn't download a car, your product use might train a low-cost competitor.
  - vajrabum 1y ago
    I'm pretty sure I've seen that openai was complaining about one of their customers doing exactly that a couple of years ago. Yes, here https://aibusiness.com/nlp/tiktok-parent-allegedly-used-openai-s-api-to-build-rival-models https://aibusiness.com/nlp/tiktok-parent-allegedly-used-open...
    - refulgentis 1y ago
      Cosign, happens all the time in my experience, and off the top of my head easily undisputable evidence thats Google-able: early open models on ChatGPT transcripts, Google on ChatGPT transcripts, ByteDance on OpenAI, DeepSeek on OpenAI
- nomel 1y ago
  According to OpenAI, this is how DeepSeek did it: https://hls.harvard.edu/today/deepseek-chatgpt-and-the-global-fight-for-technological-supremacy/ https://hls.harvard.edu/today/deepseek-chatgpt-and-the-globa... I've had Claude 3.5, Grok 3, and DeepSeek claim that it was made by OpenAI.
  - whimsicalism 1y ago
    This is maybe how DS did v3, but certainly not the big technical leap to r1.
  - codedokode 1y ago
    > However, OpenAI has publicly acknowledged ongoing investigations as to whether DeepSeek “inappropriately distilled” their models to produce an AI chatbot at a fraction of the price. I wonder does OpenAI has permission from all authors of works it "inappropriately distilled"? A pirate has no right to complain about safety of navigation. I think there should be a system that if country A illegally uses works of country B for developing an AI, it loses copyright protection in country B.
    - nomel 1y ago
      > I wonder does OpenAI has permission from all authors of works it "inappropriately distilled"? The rulings are leaning towards it being transformative enough to be fair use, so permission may not be required [1]. We'll have to see how it goes with Disney vs Midjourney. > I think there should be a system that if country A illegally uses works of country B for developing an AI, it loses copyright protection in country B. This would be pointless with China, since it's already how they operate, as continuously reported by the WTO. But, I see your point. [1] https://www.whitecase.com/insight-alert/two-california-district-judges-rule-using-books-train-ai-fair-use https://www.whitecase.com/insight-alert/two-california-distr...
woadwarrior01 1y ago
Very reminiscent of the internal feuds at Meta around the Llama 1 timeframe (~2 years ago).
dang 1y ago
Recent and related: Whistleblower: Huawei cloned Qwen and DeepSeek models, claimed as own - https://news.ycombinator.com/item?id=44482051 https://news.ycombinator.com/item?id=44482051 - July 2025 (58 comments) Also: Huawei Whistleblower Alleges Pangu AI Model Plagiarized from Qwen and DeepSeek - https://news.ycombinator.com/item?id=44506350 https://news.ycombinator.com/item?id=44506350 - July 2025 (1 comment) Pangu's Sorrow: The Sorrow and Darkness of Huawei's Noah Pangu LLM R&D Process - https://news.ycombinator.com/item?id=44485458 https://news.ycombinator.com/item?id=44485458 - July 2025 (2 comments) Huawei's Pangu Pro MoE model is likely derived from Qwen model - https://news.ycombinator.com/item?id=44461094 https://news.ycombinator.com/item?id=44461094 - July 2025 (1 comment) Huawei releases an open weight model trained on Huawei Ascend GPUs - https://news.ycombinator.com/item?id=44441089 https://news.ycombinator.com/item?id=44441089 - July 2025 (333 comments)

26 points by avervaet 1y ago

So... the same things as every other foundational model?

What if there truly is no moat? What if you can just hook your AI system up to some other AI system and drain everything? No weights access required. Just train on the raw inputs/outputs. What stops this from being the future?

refulgentis 1y ago
Isn't that today? (or, perhaps more pedantically, yesterday and today and tomorrow?)
- echelon 1y ago
  ... and tomorrow, and the next day. It never ends. There's no moat. One day your at-home GPU will unwind an entire hyperscaler's worth of expertise. Does the capital outlay get them anything at all apart from a temporary lead? While you couldn't download a car, your product use might train a low-cost competitor.
- vajrabum 1y ago
  I'm pretty sure I've seen that openai was complaining about one of their customers doing exactly that a couple of years ago. Yes, here https://aibusiness.com/nlp/tiktok-parent-allegedly-used-openai-s-api-to-build-rival-models https://aibusiness.com/nlp/tiktok-parent-allegedly-used-open...
  - refulgentis 1y ago
    Cosign, happens all the time in my experience, and off the top of my head easily undisputable evidence thats Google-able: early open models on ChatGPT transcripts, Google on ChatGPT transcripts, ByteDance on OpenAI, DeepSeek on OpenAI
nomel 1y ago
According to OpenAI, this is how DeepSeek did it: https://hls.harvard.edu/today/deepseek-chatgpt-and-the-global-fight-for-technological-supremacy/ https://hls.harvard.edu/today/deepseek-chatgpt-and-the-globa... I've had Claude 3.5, Grok 3, and DeepSeek claim that it was made by OpenAI.
- whimsicalism 1y ago
  This is maybe how DS did v3, but certainly not the big technical leap to r1.
- codedokode 1y ago
  > However, OpenAI has publicly acknowledged ongoing investigations as to whether DeepSeek “inappropriately distilled” their models to produce an AI chatbot at a fraction of the price. I wonder does OpenAI has permission from all authors of works it "inappropriately distilled"? A pirate has no right to complain about safety of navigation. I think there should be a system that if country A illegally uses works of country B for developing an AI, it loses copyright protection in country B.
  - nomel 1y ago
    > I wonder does OpenAI has permission from all authors of works it "inappropriately distilled"? The rulings are leaning towards it being transformative enough to be fair use, so permission may not be required [1]. We'll have to see how it goes with Disney vs Midjourney. > I think there should be a system that if country A illegally uses works of country B for developing an AI, it loses copyright protection in country B. This would be pointless with China, since it's already how they operate, as continuously reported by the WTO. But, I see your point. [1] https://www.whitecase.com/insight-alert/two-california-district-judges-rule-using-books-train-ai-fair-use https://www.whitecase.com/insight-alert/two-california-distr...

woadwarrior01 1y ago

Very reminiscent of the internal feuds at Meta around the Llama 1 timeframe (~2 years ago).

dang 1y ago

Recent and related: Whistleblower: Huawei cloned Qwen and DeepSeek models, claimed as own - https://news.ycombinator.com/item?id=44482051 https://news.ycombinator.com/item?id=44482051 - July 2025 (58 comments) Also: Huawei Whistleblower Alleges Pangu AI Model Plagiarized from Qwen and DeepSeek - https://news.ycombinator.com/item?id=44506350 https://news.ycombinator.com/item?id=44506350 - July 2025 (1 comment) Pangu's Sorrow: The Sorrow and Darkness of Huawei's Noah Pangu LLM R&D Process - https://news.ycombinator.com/item?id=44485458 https://news.ycombinator.com/item?id=44485458 - July 2025 (2 comments) Huawei's Pangu Pro MoE model is likely derived from Qwen model - https://news.ycombinator.com/item?id=44461094 https://news.ycombinator.com/item?id=44461094 - July 2025 (1 comment) Huawei releases an open weight model trained on Huawei Ascend GPUs - https://news.ycombinator.com/item?id=44441089 https://news.ycombinator.com/item?id=44441089 - July 2025 (333 comments)