Earlier this week, DeepSeek, a well-funded Chinese language AI lab, launched an “open” AI mannequin that beats many rivals on common benchmarks. The mannequin, DeepSeek V3, is giant however environment friendly, dealing with text-based duties like coding and writing essays with ease.
It additionally appears to assume it’s ChatGPT.
Posts on X — and TechCrunch’s personal exams — present that DeepSeek V3 identifies itself as ChatGPT, OpenAI’s AI-powered chatbot platform. Requested to elaborate, DeepSeek V3 insists it’s a model of OpenAI’s GPT-4 mannequin launched in 2023.
This really reproduces as of at the moment. In 5 out of 8 generations, DeepSeekV3 claims to be ChatGPT (v4), whereas claiming to be DeepSeekV3 solely 3 occasions.
Offers you a tough concept of a few of their coaching information distribution. https://t.co/Zk1KUppBQM pic.twitter.com/ptIByn0lcv
— Lucas Beyer (bl16) (@giffmana) December 27, 2024
The delusions run deep. If you happen to ask DeepSeek V3 a query about DeepSeek’s API, it’ll provide you with directions on the way to use OpenAI’s API. DeepSeek V3 even tells a number of the similar jokes as GPT-4 — right down to the punchlines.
So what’s occurring?
Fashions like ChatGPT and DeepSeek V3 are statistical techniques. Educated on billions of examples, they study patterns in these examples to make predictions — like how “to whom” in an electronic mail sometimes precedes “it could concern.”
DeepSeek hasn’t revealed a lot in regards to the supply of DeepSeek V3’s coaching information. However there’s no scarcity of public datasets containing textual content generated by GPT-4 through ChatGPT. If DeepSeek V3 was educated on these, the mannequin would possibly’ve memorized a few of GPT-4’s outputs and is now regurgitating them verbatim.
“Clearly, the mannequin is seeing uncooked responses from ChatGPT sooner or later, nevertheless it’s not clear the place that’s,” Mike Prepare dinner, a analysis fellow at King’s School London specializing in AI, instructed TechCrunch. “It may very well be ‘unintended’ … however sadly, we’ve got seen situations of individuals straight coaching their fashions on the outputs of different fashions to attempt to piggyback off their information.”
Prepare dinner famous that the apply of coaching fashions on outputs from rival AI techniques might be “very unhealthy” for mannequin high quality, as a result of it could result in hallucinations and deceptive solutions just like the above. “Like taking a photocopy of a photocopy, we lose increasingly more data and connection to actuality,” Prepare dinner mentioned.
It may additionally be in opposition to these techniques’ phrases of service.
OpenAI’s phrases prohibit customers of its merchandise, together with ChatGPT prospects, from utilizing outputs to develop fashions that compete with OpenAI’s personal.
OpenAI and DeepSeek didn’t instantly reply to requests for remark. Nonetheless, OpenAI CEO Sam Altman posted what seemed to be a dig at DeepSeek and different rivals on X Friday.
“It’s (comparatively) simple to repeat one thing that you recognize works,” Altman wrote. “This can be very onerous to do one thing new, dangerous, and tough whenever you don’t know if it’s going to work.”
Granted, DeepSeek V3 is way from the primary mannequin to misidentify itself. Google’s Gemini and others typically declare to be competing fashions. For instance, prompted in Mandarin, Gemini says that it’s Chinese language firm Baidu’s Wenxinyiyan chatbot.
And that’s as a result of the online, which is the place AI corporations supply the majority of their coaching information, is changing into littered with AI slop. Content material farms are utilizing AI to create clickbait. Bots are flooding Reddit and X. By one estimate, 90% of the online may very well be AI-generated by 2026.
This “contamination,” if you’ll, has made it fairly tough to completely filter AI outputs from coaching datasets.
It’s actually attainable that DeepSeek educated DeepSeek V3 straight on ChatGPT-generated textual content. Google was as soon as accused of doing the identical, in any case.
Heidy Khlaaf, chief AI scientist on the nonprofit AI Now Institute, mentioned the associated fee financial savings from “distilling” an current mannequin’s information might be engaging to builders, whatever the dangers.
“Even with web information now brimming with AI outputs, different fashions that might by chance prepare on ChatGPT or GPT-4 outputs wouldn’t essentially exhibit outputs harking back to OpenAI personalized messages,” Khlaaf mentioned. “If it’s the case that DeepSeek carried out distillation partially utilizing OpenAI fashions, it will not be shocking.”
Extra possible, nonetheless, is that quite a lot of ChatGPT/GPT-4 information made its manner into the DeepSeek V3 coaching set. Which means the mannequin can’t be trusted to self-identify, for one. However what’s extra regarding is the likelihood that DeepSeek V3, by uncritically absorbing and iterating on GPT-4’s outputs, might exacerbate a number of the mannequin’s biases and flaws.
TechCrunch has an AI-focused publication! Join right here to get it in your inbox each Wednesday.
