In the past two months, domestic large models have been involved in the “long text” war, and iFlytek's large model of iFlytek is no exception. Recently, the iFlytek Spark model has received an update to version V3.5. This update significantly enhances the processing capabilities of long text, long graphics and long speech. At the same time, the new version also launches for the first time a large model of Spark image and text recognition, multi-emotional super-anthropomorphic synthesis technology, and a sentence reenactment function.
According to official instructions, the iFlytek Spark model can quickly absorb and understand large amounts of text data from different channels in terms of long text processing, and provide more accurate answers to questions and answers in various industries and professional fields. In addition, the efficiency of iFlytek Spark has also been significantly improved in terms of file upload, response speed to knowledge questions and answers, and text generation.
Advertisement
At the beginning of this year, Gamingdeputy conducted an in-depth experience with the iFlytek Spark V3.5 version and made a comprehensive comparison with GPT-4. Judging from the evaluation of Gamingdeputy, the comprehensive capabilities of iFlytek Spark V3.5 are already comparable to GPT-4, and even show a certain leading advantage in terms of logical reasoning, mathematical ability and knowledge base update speed.
So, what will the experience be like with the updated version of iFlytek Spark Model V3.5, which is equipped with new functions such as long text, long picture and text, and long voice? Next, Gamingdeputy will further share experiences around these new functions.
1. Long text experience
Long text processing capabilities have become one of the key indicators to measure the hard power of major model products.
Advertisement
In our daily lives, we inevitably encounter long text content, such as privacy policies that are boring to read, those long and obscure disclaimers, and those complicated and circuitous insurance contracts. .
Faced with these documents, which often contain hundreds of thousands of words, reading them completely is as difficult as finishing the philosophical work “Metaphysics”.
For the author myself, if I forget to bring my mobile phone in the bathroom one day, there is a book of xx insurance model clauses on the side. On the other side was a bottle of shower gel. I would rather memorize the ingredient list of shower gel than take the initiative to read the insurance terms of xx.
However, the existence of these things is necessary. After all, they are written for us to see. You can choose not to read it, but if you encounter problems, these privacy policies and contract terms may become key.
So here comes the question,How do we quickly locate key information in content containing tens of thousands or even hundreds of thousands of words?Especially when it comes to insurance clauses and contracts, how do we find the clauses that are most beneficial to us? Or, how do we immediately spot the rules that are detrimental to us?
For another example, for some financial personnel or literary workers, when a company releases a financial report or white paper, how to extract the most critical points from the massive information?How to quickly find the information you care about most?
All in all, the need to read this kind of long text is a pain point that is often encountered in life, and the purpose of large models is to help us find the information we most want and need in complicated text.
So what is the performance of iFlytek Spark Large Model V3.5?
1. Contract terms
First of all, iFlytek Spark Model V3.5 has newly introduced the newly introduced “Spark Contract Assistant” widget, which can comprehensively assist users in quickly drafting various contracts.
iFlytek Spark V3.5 can not only provide popular answers to a series of professional questions to ensure legal accuracy and compliance, but also help users understand and process complex information more effectively through its accurate judgment and answers.
Take the “Website Privacy Policy” in the picture below as an example. This kind of privacy policy can be seen everywhere in our lives, such as swiping to agree to those lengthy terms before opening every app.
The author copied the privacy policy into Word and then uploaded it to the chat box of iFlytek Spark Model V3.5:
Then the following two questions were asked:
Outline the issues in this privacy policy that I need to pay special attention to
What information does this privacy policy obtain about me?
iFlytek Spark Model V3.5 quickly gave simple and easy-to-understand answers after receiving questions.
Next, the author uploaded the “Motor Vehicle Commercial Insurance Model Clauses and Disclaimers Statement for Fee-to-Fee Motor Vehicle Insurance” (picture below), with a word count of about 20,000+, and consulted on various issues regarding exemption clauses, insurance compensation, etc.
For example, when the author asked “Are you compensated for water intrusion into the engine?”, iFlytek Spark Large Model V3.5 quickly gave the answer:
Returning to the “Disclaimer”, I did find this statement, as shown in the red box in the picture below.
However, the author still didn’t quite understand the meaning of this “special agreement”, so I continued to ask about Feixinghuo Large Model V3.5, and gave the answer (picture below):
Later, I consulted the insurance company’s manual customer service with the same question, and the answer was as follows:
The answers given by iFlytek Spark Model V3.5 and manual customer service are the same..
Regarding this “special clause”, let me insert a sentence here.
The author checked some information online. To put it simply, after the new insurance regulations, normal engine water wading is covered by car damage insurance. However, some people use their cars in deserts or areas with less rain.Then purchasing insurance with this “special clause” can further reduce the premium, but if the engine is damaged by water, the insurance company will not pay out the claim.
Judging from the answer of iFlytek Spark Model V3.5, the meaning of this “term” is indeed expressed clearly, and it is basically consistent with the answer of customer service.
However, there is a premise that “engine water wading is included in the coverage of car damage insurance”, and this premise does not appear in this 20,000-word “Disclaimer”. iFlytek Spark Model V3.5 is retrieved through the long text provided by the author, so naturally I don’t know this.
Next, the author asked “Do you recommend buying it?”
iFlytek Spark Model V3.5 gave a clear answer – not recommended.
Customer service also doesn’t recommend it.
The author also asked questions about various issues in insurance, and iFlytek Spark Large Model V3.5 gave accurate answers:
Judging from various answers, iFlytek Spark V3.5 has reached a satisfactory level.And have certain logical reasoning abilitycan give users a correct suggestion.
2. Research reports
On the morning of the 26th of this month, OPPO released the “OPPO Innovation and Intellectual Property White Paper”, a pdf file with a total of 23 pages.
Regarding the content of this white paper, the author also asked a series of questions.
The answer results of iFlytek Spark V3.5 are naturally satisfactory, andVery fast feedback, giving an answer almost within seconds. For some text workers, this is simply an efficiency artifact.
A research report on Li Auto has a large number of charts, picture descriptions and data. Iflytek Spark V3.5 can even answer very detailed questions in the report (such as sales).
In response to users' needs in scientific research, iFlytek Spark V3.5 also adds long text summary capabilities, and long text generation capabilities for industry reports.
In the Spark Assistant Center, find the Spark Scientific Research Assistant to provide a series of professional answers.
The author uploaded the research reports “Research and Judgment of Global Industrial Digital Transformation Trends and Directions” and “Huawei Terminal Sustainability Report (2022-2023)”, and initiated a series of questions on the professional issues contained therein.
iFlytek Spark V3.5 can provide systematic answers to deal with complex issues in these professional fields.
Long text summary:
Long text generation:
3. Reading and entertainment
Finally, the author uploaded the TXT file of Yu Hua's novel “Shouting in the Drizzle” to iFlytek Spark V3.5, and raised a series of questions about the many characters and storylines presented in the novel.
“Shouting in the Drizzle” is also one of my favorite novels. I have read it four or five times and am deeply impressed by every story and plot in it.
But, after all, this is also a movieA novel with 149,000 wordsthe novel contains many intertwined details and plots. It is probably not easy for me to give a comprehensive and accurate answer.
So how does iFlytek Spark V3.5 perform?
First of all, the author asks, what kind of person is the father of the protagonist of the novel (Sun Guangcai)? The answer of iFlytek Spark V3.5 is as follows:
In the author's opinion, Sun Guangcai was a complete scoundrel in the second half of his life, selfish, hypocritical, despicable and pitiful.The answer given by iFlytek Spark V3.5 is quite close to the author's opinion, but the “strength” of the judgment is not enough.But it is undeniable that contradictions run through Sun Guangcai's entire life, and iFlytek Spark V3.5 also gives this view.
In the novel “Shouting in the Drizzle”, the author has a lot of descriptions of death, and Sun Guangcai's death is the most dramatic scene in the novel.
Regarding this question, iFlytek Spark V3.5 also gave an accurate answer – he was buried in the most filthy place, but he did not know this when he died.
Regarding my grandfather’s life experience, iFlytek Spark V3.5 can also make a concise summary:
However, regarding some more complex issues, although iFlytek Spark V3.5 is clear and thorough in its statement of facts,A little superficial in terms of in-depth evaluation.
For example, when asked about “Sun Guangming's life-saving behavior”, iFlytek Spark V3.5 answered “It appreciates his selfless heroism and also reveals its critical attitude towards his reckless behavior.”
The author describes it this way in the novel:
Sun Guangming drowned to save the child. To use self-sacrifice to save others on my brother is obviously an exaggeration. The younger brother is not noble enough to be willing to trade his own death for someone else's life.
His behavior at that moment came from his authority over those seven or eight-year-old children.
When death struck the children under Sun Guangming, he thought he could easily save them.
The rescued child could not recall the original scene at all. He would only look at the person who asked him dumbfounded. A few years later, when someone mentioned this matter again, the child looked doubtful, as if someone else had made it up.
If someone in the village hadn't witnessed it with his own eyes, Sun Guangming might have been believed to have drowned himself.
2. Long picture and text experience
Compared with complex situations other than simple pictures without text or long text, the large image and text recognition model of iFlytek Spark V3.5 can perform high-precision analysis of complex layouts.
Officials stated that the recognition scenarios include education (books, composition correction), patents, academic papers, newspapers, financial documents, physical examination reports, natural scenes, PPT, product manuals, posters, reading materials, pill boxes, APP screenshots, etc.
For the above scenario, you can ask questions about the text information in the picture, or you can ask more in-depth and integrated questions based on the text information.
Taking the “nasal spray” I just bought as an example, I took a photo of the instruction manual with my mobile phone and then uploaded it to iFlytek Spark V3.5.
The author asked about precautions and usage methods, and iFlytek Spark V3.5 was able to give specific answers.
Judging from the results, basically the instructions in the manual areThe text information is “OCR”, then organized according to the meaning of the word, and fed back to the questioner.
For more complex scenarios, the author has uploaded a screenshot of the USB tester instruction manual:
We asked questions about functions for different interfaces, and the answers to iFlytek Spark V3.5 were also satisfactory.
The author also uploaded a PPT picture taken at a previous event and asked iFlytek Spark V3.5 to extract the key points in the photo.
The results showed that iFlytek Spark V3.5 accurately identified the content in the photo and correctly judged that it was a technology demonstration by GAC Group. It also noticed the high level of attention shown by the audience present.
At this point, I would like to briefly express my emotion. The space for imagination of this function is indeed huge, especially for the visually impaired. Although their eyes cannot see, they only need to take photos with their mobile phones and upload them to iFlytek Spark. Being able to immediately convey the world in front of them to the visually impaired through voice description can bring great help to daily life.
Of course, the current experience is not perfect. For example,There was a “read error” phenomenon (picture below)the answer is incomprehensible, there is some room for optimization.
3. Long voice and video experience
In today's learning and life, we not only need large models to assist word processing, but also need them to assist in processing voice and video data. Especially under the trend of “national short videos”, something that can be explained clearly in one or two sentences has to be made into a video.
At the same time, for students and professionals, video materials, whether academic lectures or business interviews, contain a wealth of information. The key is, how to efficiently extract the core points from these videos?
The upgraded iFlytek Spark V3.5 can help users quickly capture and understand key information in these multimedia contents.
The author uploaded the audio article “Today, Beijing Auto Show, Crying, Laughing and Haha” from Gamingdeputy, which lasts about 19 minutes.
Regarding the audio, the author asked about the main content of the audio, and iFlytek Spark V3.5 gave the answer in a very short time.
Judging from the results, they are basically satisfactory.
However, there are also some small errors in details, such asJiKr identified it as “geek” and identified NIO ET7 as “A7”, but the flaws do not hide the merits, this performance is already outstanding. To know,The above-mentioned audio contains various new technology terms, new car names, and various situations where Chinese and English are mixed, which itself is very difficult to identify.
Next, the author raised more specific questions about new cars such as Denza and Magotan, and the answers from iFlytek Spark V3.5 were very satisfactory.
iFlytek Spark V3.5 also supports uploading videos, take the “Planting grass for Huawei Sports Health Family Bucket》Shopping guide video is an example, the video is 6 minutes long.
First, the author asked him to outline the entire video content, and iFlytek Spark V3.5 gave an accurate answer.
However, a small mistake in details,Bundle”HarmonyOS“Identified as “Ham 6s” (it may also be related to the pronunciation in the video)but the overall answer did not deviate from the topic, nor did it give an ambiguous answer.
Ask the video to recommend which products are worth buying, and iFlytek Spark V3.5 can also rank them in order and give the highlights of each product.
In addition, during the experience, during the recognition process of iFlytek Spark V3.5,No long loading times eitherbasic questions are answered in “seconds”, and the more you use it, the faster it will be. The more questions you ask, the faster it will answer.
4. iFlytek Spark voice large model
iFlytek Spark V3.5 has upgraded the Spark voice model this time, bringing the first two functions of “multi-emotional super-anthropomorphic synthesis” and “one-sentence voice reproduction”, which are more interesting experiences.
The voice dialogue is a call-like interface, and the answers provided by the iFlytek Spark model are very close to natural human voices. Although it has a hint of robot-specific charm, it is very realistic overall.
“One-sentence voice reproduction” is very interesting. It can imitate your voice or the voice of others around you. After you finish recording your voiceprint, you can find your voice in “Speakers I Created”. After selection, when using voice interaction, the large model will talk to you in its own voice, and the sound reproduction is quite realistic.
You can click on the video below to experience it:
Summarize:
AI is not a new concept. A few years ago, when talking about AI changing life, I always thought it was a fantasy and out of reach. However, in just two or three years, AI technology has undergone explosive upgrades and changes, and a truly golden age of AI is just around the corner.
In this process, iFlytek is both a participant and a promoter. iFlytek's Spark model is just one of the concrete manifestations of iFlytek's innovation in the field of AI.
This time, the iFlytek Spark model V3.5 upgrade has demonstrated excellent capabilities in long text processing, image and text recognition, and long voice and video processing, focusing on increasing user needs in terms of professionalism and practicality. experience.
As mentioned at the beginning of the article, we are exposed to a large amount of information every day. A contract with obscure words, accumulation of professional terms, and convoluted contracts, or complicated and difficult to understand exemption clauses can stop countless workers. .
In the past, you might need to go online to check various information, or spend money to consult experts to get help. During this process, you may also encounter various recharges and payments to display the answer, and it is inevitable to encounter various scammers when looking for experts.
But with the emergence of applications such as iFlytek Spark Model V3.5, the above problems can be solved very simply.
Similar scenarios include those long and cumbersome conference audios. Workers can easily find the most critical sentences among tens of thousands of words; quickly extract the essence from the video, and even quickly generate abstracts for scientific research reports. .
Although the experience of iFlytek Spark Large Model V3.5 still needs to be optimized in some details, the benefits it bringsThe space for imagination is undoubtedly huge.
iFlytek Spark V4.0 will be officially released on June 27 this year. We can look forward to what new features it will bring.