Earlier this week, DeepSeek, a well-funded Chinese AI lab, has released an “open” AI model that beats a number of competitors and beats popular benchmarks. model DeepSeek V3Like coding and writing essays, it's huge but effective and easily tackled.
That seems to be the case. ChatGPT.
post Office on X — and TechCrunch's own tests — DeepSeek V3 itself ChatGPT; Showcased as OpenAI's AI-powered chatbot platform. Asked for details, DeepSeek V3 was confirmed to be OpenAI's version. GPT-4 Model released in 2023.
This is actually reproduced today. In 5 out of 8 generations; DeepSeekV3 says ChatGPT (v4) DeepSeekV3 only says 3 times.
gives some rough idea of the distribution of their training data. https://t.co/Zk1KUppBQM pic.twitter.com/ptIByn0lcv
— Lucas Beyer (bl16) (@giffmana) December 27 2024
Deep thoughts. DeepSeek V3 If you have a question about DeepSeek's API, We will guide you on how to use it. OpenAIs API DeepSeek V3 even says some of the same things. Jokes As GPT-4 — up to the gaps.
So what's going on?
Models like ChatGPT and DeepSeek V3 are statistical systems. For example, billions have been trained; They learn patterns in those examples to make predictions — usually “how” in an email is preceded by “likely to be concerned”.
DeepSeek hasn't revealed much about the source of DeepSeek V3's training data. But there is. There is no shortage Public datasets containing text generated by GPT-4 via ChatGPT. If DeepSeek V3 was trained, The model has memorized some of the results of GPT-4 and is now reassembling them into an interpreter.
“Obviously, the model is seeing raw responses from ChatGPT at some point, but it's not clear where,” Mike Cook, a research fellow at King's College London who specializes in AI, told TechCrunch. “It could be 'accidental'… but unfortunately, we've seen instances of people training their models directly on the results of other models to test their knowledge.”
Cook noted that training models on outputs from competing AI systems can be “very detrimental” to model quality, as it can lead to misleading and false answers like the one above. “Like copying, the more we lose touch with information and reality,” Cook said.
It may also be against those systems' terms of service.
OpenAI's terms prohibit users of its products, including ChatGPT users, from developing models that compete with OpenAI's own.
OpenAI and DeepSeek did not immediately respond to requests for comment. But OpenAI CEO Sam Altman posted the obvious. dig in DeepSeek and other competitors on X Friday.
“It's (relatively) easy to copy what you know,” Altman wrote. “It's very difficult to do something new when you don't know if it's going to work,” It's dangerous and difficult.”
The DeepSeek V3 is far from the first model to misidentify itself. Google's Gemini and more Sometimes They say parallel models. For example, it is signaled by Mandarin, Gemini. He said. It is the Chinese company Baidu's Wenxinyiyan chatbot.
That's because AI companies source most of their training data from the web. Bitch With AI Harp. Content farms use AI to create. Clickbait. Bots are flooding in. Reddit versus X. One by one Estimates90% of the web will be powered by AI by 2026.
This “pollution” if you will, did it. It's quite difficult. To thoroughly filter AI results from training datasets.
Of course, DeepSeek trains DeepSeek V3 directly on ChatGPT generated text. Google once was. Accused. the same
Heidy Khlaaf, chief AI scientist at the nonprofit AI Now Institute, said knowledge of the current model could be cost-effective for developers, regardless of the risks.
“Even on Internet data full of AI outputs, ChatGPT or other models that are randomly trained on GPT-4 outputs will not demonstrate results reminiscent of OpenAI custom text,” Khlaaf said. “It wouldn't be surprising if it used partially OpenAI models and was distilled by DeepSeek.”
But what's more likely is that a lot of ChatGPT/GPT-4 data went into the DeepSeek V3 training set. This means that the model cannot be trusted to self-identify for an individual. More worrisome, however, is that DeepSeek V3 has the potential to trivially absorb and iterate over GPT-4's outputs. It makes it worse. Some models biases versus A lot..
TechCrunch has an AI-focused newsletter. Register here. Get it in your inbox every Wednesday.