On May 8, the share price of HKUST Xunfei was reported to be 63.86 yuan after the market opened. On May 6, the company released its own natural language model "Xunfei Xinghuo Cognitive Model" (hereinafter referred to as "Xinghuo") the day before the A-share market closed on the weekend.
Since the release of GPT-4 by OpenAI, domestic Baidu, Huawei, and Alibaba have successively launched their own large-scale models, striving to create a Chinese version of ChatGPT, and iFLYTEK, which has been in the intelligent voice industry for more than 20 years, has even greater ambitions. Its chairman Liu Qingfeng At the press conference last Friday, it was stated that Xunfei Xinghuo’s cognitive large model “has surpassed” ChatGPT in terms of text generation, knowledge question and answer, and mathematical ability.
Not only that, HKUST Xunfei also plans to let "Xinghuo" achieve the goal of surpassing ChatGPT in Chinese and equivalent in English before October 24. Why is the upgrade time deadline so accurate? Some netizens found that the company is expected to complete the reduction of major shareholders before October 12 this year.
"Has surpassed" and "will surpass", the future is unknown, and the experience is in the present. However, on the social network, the "Spark" model has created many rollover scenes.
Before the market closes, the large model will be released and the daily limit will be opened
On May 8, HKUST Xunfei (002230.SZ) rose 9.91% 15 minutes after the market opened, and reported 63.86 yuan at 9:46 a.m., quickly hitting the daily limit, with a turnover of 8.707 billion yuan. Since the beginning of this year, the stock price of this listed company specializing in intelligent voice has nearly doubled, with a market value of nearly 150 billion yuan.
HKUST Xunfei’s 2023 first quarter report shows that the company’s net loss in the first quarter was 57.895 million yuan, compared with a net profit of 110 million yuan in the same period last year. The company has stated that although its performance declined in the first quarter, based on the progress in domestic substitution and business development, it is expected to achieve positive growth in revenue and gross profit from the second quarter.
Just after the first month of the second quarter, HKUST Xunfei’s stock price took off first. "If the market opens on Monday, HKUST Xunfei will definitely go up." On the afternoon of May 6, some netizens who paid attention to stocks predicted so. On the same day, HKUST Xunfei released its own large-scale model "Xunfei Spark Cognitive Model" (hereinafter referred to as "Spark") in the company's headquarters in Hefei.
The release of blockbuster products the day before the market closed, HKUST Xunfei contributed to the biggest positive signal.
It is reported that the "Spark" model has seven core capabilities: multi-style multi-task long text generation, multi-level cross-lingual language understanding, pan-domain open knowledge question and answer, situational thinking chain logical reasoning, multi-question step-level mathematical ability, Multifunctional multilingual coding capability, multimodal input and expression capability.
Two days after the press conference, the A-share market was closed. HKUST Xunfei appeared in the headlines of major brokerage research reports. CITIC Securities, Huatai Securities, etc. gave positive comments on the application prospects of HKUST Xunfei’s "Spark" model. Evaluation, release "buy" signal.
Strong market expectations still come from HKUST Xunfei's "Spark" large-scale model conference. Unlike Baidu’s recorded and broadcasted demonstration when it released Wenxin Yiyan, HKUST Xunfei demonstrated the ability of “Spark” in real time.
Combined with "Spark", Xunfei AI learning machine improves composition correction ability
At the press conference, Liu Qingfeng asked the AI learning machine implanted with the "Spark" large-scale model to correct the composition. A handwritten composition was marked with various correction suggestions, from punctuation, the use of words and sentences to rhetorical techniques, AI incarnated all-round "composition teacher". It is reported that the upgraded AI learning machine can only correct low-level mistakes such as typos and wrong sentences. With the large model, it can give specific basis and examples for revision, generate more detailed content evaluation, and quantitatively analyze students' writing ability.
What also released greater confidence to the market is iFLYTEK’s next large-scale model upgrade plan: the company plans to achieve the ability to benchmark against ChatGPT before October 24. According to Liu Qingfeng, the Chinese proficiency should exceed the level of ChatGPT, and the English proficiency is equivalent to it.
"Xinghuo" claims to surpass ChatGPT's Chinese ability before October 24
HKUST Xunfei’s goal alignment time is not like some large-scale model R&D companies who only dare to say "the next X months", but accurate to "before October 24". Some people are curious, "If it can't be realized by then, aren't you afraid of being slapped in the face?" Some careful netizens discovered the "clue": the major shareholder of HKUST Xunfei will reduce its holdings in October.
On March 22 this year, HKUST Xunfei released the "Pre-disclosure Announcement of the Major Shareholder's Plan to Reduce Shareholdings". The shareholder, China University of Science and Technology Asset Management Co., Ltd. It is planned to reduce the company's shares by no more than 23.23 million shares (1% of the company's total share capital) through centralized bidding (or block transactions) from April 21, 2023 to October 12, 2023.
Weibo blogger @小咖科技v believes that the reason why HKUST Xunfei is not afraid to slap its face and blow up its own big model is to "smoothly reduce its holdings of 23.23 million shares at a high price before October 12."
Of course, the major shareholder’s shareholding reduction plan is still the “prediction” of iFLYTEK, and it is unknown whether the stock price can maintain the current momentum at that time. It depends not only on the profitability of iFLYTEK, but also on the ability of the large-scale AI model. It cannot last until October of this year.
Beyond ChatGPT? "Spark" rollover case full of Netflix
When evaluating the ability of the "Spark" large model, Liu Qingfeng at the press conference has been benchmarking ChatGPT, "If the full score of English is calculated as 5 points, ChatGPT is now 4.48 points, and we have reached 4.29 points." He said that the current Xunfei Spark The text generation, knowledge question answering and mathematical capabilities of the cognitive large model have surpassed ChatGPT. "
Is the "Spark" model that is confident that it "has surpassed" and "will surpass" ChatGPT okay? The future is unknown, but it can be experienced now. Various netizens who have obtained the internal test qualification have given the test.
Some netizens mixed and matched the characters and plots in the four famous novels to ask questions, and wanted to see if "Xinghuo" could find mistakes. As a result, it told the story of "Lin Daiyu beats the bone spirit three times". In the story, Xue Baochai became Lin Daiyu's cousin.
So, the netizen lowered the difficulty and asked a question only about "Water Margin": Which people were on Liangshan because of being cuckolded? In order to test "Spark"'s understanding of "green hat". In the end, the answer was wrong. "Xinghuo" pinched Lu Zhishen and Pan Jinlian into bad CPs, and arranged for Song Jiang's wife "Li Shi", and talked about unrequited love with Ximenqing... "Water Margin" "Biography" seems to be just names of people, and "Xinghuo" is responsible for "spreading rumors" for them.
Netizens who did not give up gave up the difficulty of logical question-and-answer and semantic understanding, and directly asked a plot of "Water Margin": Who burned the forage yard? As a result, the classic plot of Lin Chong's being framed was arranged on Song Jiang, and a deed of "occupying the fodder field" was added to Song Yasi.
"Xinghuo", which is known as a rich source of corpus, has a sense of déjà vu that has not been input into the four major masterpieces at all. For some simpler factual questions, the mistakes of "Spark" are also likely to mislead the uninformed. Someone asked about Tongling's license plate number. The "Xinghuo" developed by Anhui's local company iFLYTEK was unfamiliar with the city of his hometown, and output Tongling's license plate code as "Wan M". In fact, the license plate code of Tongling City is WanG, and WanM is Chuzhou City.
Judging from the experience feedback of netizens, the large model "Spark" has a high turnover rate in basic text answers, and it also has the common problem of large models "talking nonsense in a serious manner". Such performance wants to surpass ChatGPT before October 24th, iFLYTEK has to work harder. It’s no wonder that some netizens joked that the claim that “Xinghuo” surpassed ChatGPT in terms of Chinese ability is “completely unbelievable” just like Cambrian’s self-proclaimed surpassing Nvidia.