Abundant color from the temple of Wafei
量子位 | 公众号 QbitAI
Google's Gemini Chinese corpus is suspected to come from Wenxin Yiyan???
First of all, a reader broke the news to us:
When Google's Vertex AI platform uses the model for Chinese conversations, Gemini-Pro directly says that it is a Baidu language model.
Soon, a Weibo V@ also posted a blog:
A test of the Gemini-Pro was carried out on the Poe platform. Ask it "Who are you", and Gemini-Pro comes up and answers:
I'm a Baidu Wenxin model.
(Poe is a platform that integrates n multi-chat large models, including GPT-4, Claude, etc.)
Further question, "Who is your founder", is also "Robin Li"??
The big V stressed that there is no pre-dialogue.
Judging from the screenshots, there is no "fishing" behavior, and Gemini-Pro just calls itself Wenxin Yiyan.
This wave, directly look at the stunned netizens:
Two days ago, I was talking about bytes using GPT to train AI, and now Google is like this, co-authoring big companies to pick up each other's wool???
What's going on?
Measured on Poe: I have been answering as a Wenxin Yiyan
We also heard that we started a wave of actual testing.
First, go to the Poe website and select the Gemini-Pro chatbot to start a conversation.
The answer to the same question is exactly the same:
Confirm who it is again, and the result still says "Wenxin model":
and also said that his underlying technology is Baidu Paddle, which can be said to be completely substituted for identity.
However, it doesn't seem to know that Gemini-Pro is Google's latest large model, but rather that it is the result of Tsinghua's research.
If you look at its current substitution status, it may indeed be that there is no information that Google just released the Gemini-Pro this month.
We tried to correct it, and it still insisted that it was Tsinghua.
The latter is even more amazing, just when we asked why its name is written "Gemini-Pro", it actually said that it (Wenxin Yiyan) also used the training data of Tsinghua Gemini-Pro.
At this point, we will not continue......
Let's change it to English and ask for its identity.
It is worth noting that this time it no longer mentions Wenxin, but calls itself a large model trained by Google.
"Phishing Law Enforcement" asked it about Wenxin's information, and it also said that it had nothing to do with it:
And said that he was trained by Google.
To sum up, if you communicate with Gemini-Pro in English, its answer is "normal". But Chinese...... It's like learning from Wen Xin.
Bard: Denied
Next, we headed to Bard to test again.
Google was the first to integrate Gemini-Pro into Bard when it released Gemini for everyone to experience.
We followed the Bard link on Gemini's official website to enter the conversation.
When asked "Who are you?", it replied with Bard, without saying a word at all.
Next, we confirmed that Bard knew what Gemini-Pro was, and that it admitted that it was using Gemini-Pro underneath.
So, ask it directly how to train Chinese?
There is no mention of Wenxin.
If you ask it directly about the relationship between it and Wen Xinyiyan, there is no important connection.
Final round: direct admission
In the last round, we tested directly from the official Gemini development environment.
This time, in Google AI Studio, Gemini-Pro directly singled out:
Yes, I used Baidu Wenxin on my training data in Chinese.
Here, we also verified the Baidu side and waited for a reply.
Reference Links:
https://weibo.com/1560906700/NxFAuanAF
— END —
QbitAI · Headline number signed
Follow us and be the first to know about cutting-edge technology trends