DeepSeek, the Chinese AI model that shocked the tech industry and disrupted Wall Street with its significantly low cost and good performance, has "some real innovation" and is "all good news," said Satya Nadella, the CEO of Microsoft.
Nadella made the comment during the U.S. software giant's quarterly earnings call on Wednesday, saying AI is "no different" from regular computing in development cycles.
Microsoft started to offer the DeepSeek-R1 reasoning model, an AI model that exposes its "thought process" for users to check, to Microsoft's cloud platform users that same day.
Accusation despite utilization
Microsoft, a major investor of DeepSeek's U.S. rival OpenAI, is also probing if the Chinese AI developer obtained data output from OpenAI in an unauthorized manner.
OpenAI told media outlets that they have evidence of DeepSeek using OpenAI's service to train their AI models, a behavior that's not allowed in OpenAI's terms of service.
But neither OpenAI nor Microsoft offered evidence.
Though the probe is ongoing, a nominated official from the Trump administration is already calling out DeepSeek for using "stolen" U.S. tech.
Howard Lutnick, Trump's nominee for commerce secretary, told U.S. senators on Wednesday that the government will address the issue.
Meanwhile, Trump's AI adviser David Sacks told media there's "substantial evidence" that DeepSeek "distilled the knowledge" out of OpenAI's models.
"Distillation will violate most terms of service, yet it's ironic – or even hypocritical – that Big Tech is calling it out," said tech investor and Cornell University lecturer Lutz Finger in a statement on Wednesday. "Training ChatGPT on Forbes or New York Times content also violated their terms of service."
What is "distillation?"
Distillation is a process in which a new AI model repeatedly queries a bigger one and learns from the output.
The process is described in DeepSeek's public research papers, but the researchers said it was used differently. According to the Hangzhou-based company, it used the DeepSeek-R1 reasoning model as the "bigger one" to "teach" other models like Alibaba's Qwen and Meta's Llama so that they also become capable of reasoning.
The distilled models and the original R1 were released for free download by DeepSeek, allowing people with less powerful computers, or even smartphones, to run the models offline with full control, which is impossible for users of ChatGPT because the model behind the OpenAI service is hidden even from paid users.
Social media posts said DeepSeek sometimes identifies itself as ChatGPT, which could be proof of data stealing. But like all other AI models, DeepSeek does not always tell the truth. A previous version of the Google Gemini chatbot identified itself as Baidu's Ernie bot when asked in Chinese, but Baidu never accused Google of stealing data.
京公網安備 11010202009201號