Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Earlier this week, DeepSeek, China’s highly-funded AI lab, released an “open-source” AI model that beats many rivals in popular benchmarks. model, DeepSeek V3it’s large but useful, handling text-based tasks like copying and writing articles with ease.
It also seems to think that it is ChatGPT.
Notes on X – and TechCrunch tests that – show that DeepSeek V3 identifies itself as ChatGPT, OpenAI’s AI-powered platform. When asked to elaborate, DeepSeek V3 insists that it is OpenAI’s version GPT-4 model was released in 2023.
This is still happening today. In 5 out of 8 generations, DeepSeekV3 says it’s ChatGPT (v4), while it says it’s DeepSeekV3 only 3 times.
It gives you a rough idea of their distribution of training data. https://t.co/Zk1KUppBQM pic.twitter.com/ptIByn0lcv
– Lucas Beyer (bl16) (@giffmana) December 27, 2024
Fraud runs deep. When you ask DeepSeek V3 a question about DeepSeek’s API, it will give you instructions on how to use it. It’s OpenAI API. DeepSeek V3 tells the same story joke like GPT-4 – down to the punchlines.
So what’s going on?
Models like ChatGPT and DeepSeek V3 are statistical systems. Trained on billions of examples, they learn patterns in those examples to make predictions – just like the “to whom” in an email often leads to “it might be difficult.”
DeepSeek did not reveal any information about the source of DeepSeek V3 training material. But there is there is no shortage of the community groups containing GPT-4 text via ChatGPT. If DeepSeek V3 was trained on this, the model could memorize some of the GPT-4 output and now repeat the text.
“Apparently, the model is seeing raw responses from ChatGPT at some point, but it’s not clear where they are,” Mike Cook, a researcher at King’s College London who works in AI, told TechCrunch. “It could be ‘accidental’…
Cook also said that the practice of training models on outputs from AI combat systems can be “very bad” for the model, as they can lead to hallucinations and misleading responses like the one above. “Like taking a photocopier, we lose information and connection to reality,” Cook said.
It may also conflict with the system’s performance.
OpenAI’s statement prohibits users of its products, including ChatGPT customers, from using derivatives to develop models that compete with those of OpenAI.
OpenAI and DeepSeek did not immediately respond to requests for comment. However, OpenAI CEO Sam Altman wrote what appears to be a digging at DeepSeek and other competitors at X Friday.
“It’s (less) easy to base what you know works,” Altman said. It’s very difficult to do something new, risky, and challenging when you don’t know if it will work.
Of course, DeepSeek V3 is far from the first version of self-identification. Google’s Gemini and others sometimes they are said to be competing models. For example, he recommended Mandarin, Gemini he says that it is the Chinese company Baidu’s Wenxinyiyan chatbot.
And that’s because the internet, which is where AI companies release most of their training, is staying disposables and AI to slide. Content farms are using AI to create clickbait. The bottles are overflowing Reddit and X. It is one compare90% of the internet could be powered by AI by 2026.
This “defilement”, if you will, has caused it quite difficult to better filter the AI output from the training set.
It is indeed possible that DeepSeek trained DeepSeek V3 directly on the text generated by ChatGPT. Google was already there the accused do the same, later.
Heidy Khlaaf, chief AI scientist at the non-profit AI Now Institute, said the cost savings from “distilling” existing model information can be attractive to developers, regardless of the risk.
“Even with online data that contains AI results, some models that can accidentally learn the results of ChatGPT or GPT-4 will not show results that remember OpenAI messages,” said Khlaaf. “If DeepSeek did a little distillation using OpenAI models, it wouldn’t be surprising.”
Chances are, however, that most of the ChatGPT/GPT-4 data went into the DeepSeek V3 tutorial. This means that the model cannot be trusted to identify itself, for one. But what is most concerning is the possibility that DeepSeek V3, based on the unquestionable and repeated GPT-4 output, will end. add it some of the models favoritism and errors.
TechCrunch has a newsletter focused on AI! Log in here to get it in your inbox every Wednesday.
