Why DeepSeek’s new AI mannequin thinks it is ChatGPT

Earlier this week, DeepSeek, a well-funded Chinese language AI lab, launched an “open” AI mannequin that beats many rivals on in style benchmarks. The mannequin, Deep Search V3is cumbersome however efficient, simply dealing with text-based duties resembling coding and essay writing.

He additionally appears to assume that it’s ChatGPT.

Messages on X — and TechCrunch’s personal assessments — present that DeepSeek V3 identifies as ChatGPT, OpenAI’s AI-powered chatbot platform. Requested to elaborate, DeepSeek V3 insists that it’s a model of OpenAI. GPT-4 mannequin launched in 2023.

That is really taking place once more beginning at this time. In 5 out of 8 generations, DeepSeekV3 pretends to be ChatGPT (v4), whereas pretending to be DeepSeekV3 solely 3 occasions.

Provides you a tough concept of ​​a few of their coaching knowledge distributions. https://t.co/Zk1KuppBQM pic.twitter.com/ptIByn0lcv

– Lucas Beyer (bl16) (@giffmana) December 27, 2024

The illusions run deep. For those who ask DeepSeek V3 a query concerning the DeepSeek API, it will provide you with directions on find out how to use it. OpenAI API. DeepSeek V3 even tells a number of the identical factor jokes like GPT-4 – proper right down to the punchlines.

So what is going on on?

Fashions like ChatGPT and DeepSeek V3 are statistical techniques. Educated on billions of examples, they study patterns in these examples to make predictions – for instance, how “to whom” in an electronic mail usually precedes “this may occasionally concern.”

DeepSeek hasn’t revealed a lot concerning the supply of DeepSeek V3’s coaching knowledge. However there’s no shortage of public datasets containing GPT-4 generated textual content by way of ChatGPT. If DeepSeek V3 had been skilled on these, the mannequin may need memorized a few of GPT-4’s output and is now regurgitating them verbatim.

“Clearly the mannequin sees uncooked responses from ChatGPT in some unspecified time in the future, however it’s not clear the place that’s,” Mike Prepare dinner, a researcher at King’s School London who makes a speciality of AI, advised TechCrunch. “This might be ‘unintended’… however sadly now we have seen circumstances the place folks immediately prepare their fashions on the outcomes of different fashions to attempt to leverage their information.”

Prepare dinner famous that the observe of coaching fashions on the output of competing AI techniques might be “very dangerous” for mannequin high quality, as it could result in hallucinations and deceptive responses like these above. “It’s like we take a photocopy of a photocopy, we lose increasingly data and reference to actuality,” Prepare dinner mentioned.

This is also towards the phrases of service of those techniques.

OpenAI’s phrases prohibit customers of its merchandise, together with ChatGPT prospects, from utilizing the outcomes to develop fashions that compete with these of OpenAI.

OpenAI and DeepSeek didn’t instantly reply to requests for remark. Nevertheless, Sam Altman, CEO of OpenAI, launched what seemed to be a dig at DeepSeek and different rivals on Friday X.

“It’s (comparatively) straightforward to repeat one thing works,” Altman wrote. “It’s extraordinarily tough to do one thing new, dangerous and tough if you don’t know if it can work. »

Definitely, DeepSeek V3 is much from the primary mannequin to misidentify. Gemini from Google and others Sometimes declare to be competing fashions. For instance, requested in Mandarin, Gemini said that it’s the Wenxinyiyan chatbot from the Chinese language firm Baidu.

And that is as a result of the net, the place AI firms get most of their coaching knowledge, is more and more littered with AI slope. Content material farms use AI to create clickbait. The robots are flooding Reddit And X. By a estimate90% of the Net might be powered by AI by 2026.

This “contamination”, if you’ll, has ensured that quite difficult to totally filter AI output from coaching datasets.

It’s actually doable that DeepSeek skilled DeepSeek V3 immediately on the textual content generated by ChatGPT. Google was once accused to do the identical factor, in spite of everything.

Heidy Khlaaf, chief AI scientist on the nonprofit AI Now Institute, mentioned the fee financial savings of “distilling” information from an current mannequin might be engaging to builders, regardless of the dangers.

“Even with Web knowledge now filled with AI output, different fashions that by chance prepare on ChatGPT or GPT-4 output wouldn’t essentially reveal output harking back to OpenAI customized messages,” Khlaaf mentioned. “Whether it is true that DeepSeek carried out distillation partially utilizing OpenAI fashions, this is able to not be stunning.”

Nevertheless, it’s extra seemingly that a considerable amount of ChatGPT/GPT-4 knowledge was fed into the DeepSeek V3 coaching set. Which means that the mannequin can’t be trusted to determine itself, for instance. However what’s extra worrying is the likelihood that DeepSeek V3, by uncritically absorbing and iterating on the outputs of GPT-4, might exacerbate a number of the fashions prejudices And defects.

TechCrunch affords a e-newsletter targeted on AI! Register here to obtain it in your inbox each Wednesday.