In 2024, generative AI will remain the most compelling tech focus.
From the so-called artificial intelligence that was ignorant of human instructions in the early days, to today, when you hear a password, you can honestly draw and write for us ......The productivity brought by AI has been significantly improved, and many people want to use it to assist their work and study, so as to improve efficiency, and even steal a little lazy.
Under the wave of AI, the current domestic leading technology enterprises intensively launch artificial intelligence models, Tencent, Huawei, Alibaba, and other giants have entered the game, invested resources, and devoted themselves to research, the industry has shown a trend of rapid development, and the "100 model war" is in full swing.
However, at present, many large models at home and abroad are actually in the internal testing stage, and only the registration threshold and use threshold have intercepted 99% of users.
Among them, iFLYTEK Xinghuo, which started the national test early, is a special existence.
On January 30, iFLYTEK released a large model of national openness based on the training of the first national computing power platform Flying OneiFLYTEK Spark v3Version 5.
(Source: iFLYTEK).
Compared to the previous version,iFLYTEK Spark v3Version 5 has achieved significant improvements in seven core capabilities, including text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, ** ability, and multimodal ability, further approaching the latest level of GPT-4 Turbo.
Not only that, iFLYTEK also brought a new Spark voice model to this press conference, as well as the first iFLYTEK Spark open source model that is deeply adapted to domestic computing infrastructure, continuing to build and consolidate the domestic large model format and bring new opportunities to the real economy.
It's still the old rule, friends who haven't squatted at the press conference, just follow Xiaolei and look down.
Since its release in May last year, the iFLYTEK Spark model has undergone several iterations in just eight months.
At the beginning, iFLYTEK gave the three upgrade milestones and time points of the Xinghuo large model in the year, and now it has been landed on June 9, August 15, and October 25 as scheduled, and the Xinghuo cognitive large model V3The rapid landing of 0 has promoted the ability of iFLYTEK's large model to quickly approach the forefront of the industry.
(Source: iFLYTEK).
Let's take a look at multiple rounds of dialogue first, Liu Cong, president of iFLYTEK Research Institute, was on the scene and Xinghuo v35. A commonplace conversation began.
During the conversation, Xinghuo will actively capture the user's current state and actively ask questions. For example, after Liu Cong said that there were a lot of trivial matters at the end of the year, Xinghuo would take the initiative to ask Liu Cong if he was going to travel and relax during the New Year, and gave detailed travel suggestions for the destination proposed by Liu Cong.
(Source: iFLYTEK).
It is not difficult to see that the iFLYTEK Xinghuo cognitive model v35. It has realized the leap from multiple rounds of dialogue, active dialogue to heuristic dialogue, and can realize the complete active communication and dialogue between man and machine.
After the progress of core capabilities such as semantic understanding, instruction following, multi-round dialogue, emotional perception and anthropomorphic synthesis, Xinghuo v35. It is expected to completely change the human-computer interaction mode in the era of the Internet of Everything.
(Source: iFLYTEK).
In terms of language semantic understanding, iFLYTEK Xinghuo 35How does it perform?Liu Cong first provided Xinghuo with a report from Anhui Province and asked Xinghuo to come up with five comprehension questions about this article.
(Source: iFLYTEK).
He then asks Starfire to answer the first and third questions in a coherent manner.
Starfire v35. You can answer two questions in a clear and coherent manner according to the internal order of the article, which not only summarizes the specific events described in this text, but also gives your own attitude and opinion on the content of the article, so that people can intuitively and clearly obtain the information they want to know.
(Source: iFLYTEK).
In terms of text generation,After importing the existing information, iFLYTEK Zhiwen can quickly generate PPT outlines and PPTs of different styles based on the theme of Hefei's 2024 Spring Festival tourism introduction, and can even automatically generate associated AI speech notes and narrators after determining the content of PPT.
(Source: iFLYTEK).
Yes, not only do you not have to do PPT now, but you may not even need to talk about it.
Finally, to test the logical reasoning ability, Liu Cong raised some questions on the spot that are easy to mislead the large model into AI hallucinations.
For example, "If there is a piece of ice floating in the basin, will the water level rise or fall when the ice melts?""If a person goes out for a walk, goes forward 20 meters, turns 60 degrees to the right, goes forward 20 meters, turns right 60 degrees, and so on, can he go back to the original point?If you can go back to square one, how far have you come?"And so on, as a result, Spark v35. Be able to answer accurately.
Even if it's a geometry problem based on a three-dimensional figure, Starfire v35 all gave a reply that was consistent with the facts.
(Source: iFLYTEK).
From the answers to these questions, it is clear that Spark v35. In terms of logical reasoning ability, there is a relatively high-quality embodiment, which can provide more accurate, comprehensive and professional answers to the questions raised by users.
Perhaps, this is a large model that is more suitable for the physique of Chinese babies.
(Source: iFLYTEK).
When it comes to iFLYTEK, voice is the first label that comes to mind for many people.
Even under the continuous sanctions, iFLYTEK has always been at the forefront of the world. Speech is the foundation of AI, whether it is NLP (natural language processing), knowledge graph, semantic understanding, speech recognition, or speech synthesis, all of which are core AI technologies. The development of speech AI over the years is an important foundation for large models, which in turn will further strengthen speech AI technology.
Now, after the breakthrough of speech technology driven by large models, human-computer interaction will usher in a new stage of development.
(Source: iFLYTEK).
The Xinghuo speech model is composed of multilingual speech synthesis, and has surpassed the whisper-large-v3 launched by OpenAI in the first batch of 37 mainstream languages, maintaining the international leading level of iFLYTEK's intelligent voice technology.
Not only that, the average MOS (mean opinion score) score of the first batch of 40 languages in the Xinghuo voice model has definitely increased by 025, MOS reached 45. The degree of anthropomorphism has reached more than 83%, successfully maintaining the international leading level of iFLYTEK in intelligent voice technology.
(Source: iFLYTEK).
iFLYTEK Translator will be the first batch of hardware products equipped with Xinghuo voice model. In addition, the iFLYTEK voice model can also be widely used in intelligent customer service, intelligent broadcasting, language assistant, vehicle-machine interconnection and other fields.
(Source: iFLYTEK).
At this press conference, iFLYTEK also showed a report based on Xinghuo V35. Empower the Spark Smart Blackboard.
This smart blackboard can not only intelligently identify the teacher's board book and digitize the board book, but also provide relevant courseware materials for students' reference based on the content of the board book, and even realize the disassembly and division of three-dimensional modeling, making the graphics more intuitive.
(Source: iFLYTEK).
As for the speaking teachers and science teachers that are difficult to equip schools, the Xinghuo smart blackboard is also integrated, so that children can practice speaking and learn science wellThe summary and highlight extraction function of the teacher's course records allows children to better review the unclear knowledge points.
Altman, the founder of OpenAI, once said that there are two AI application fields that he is particularly optimistic about:One is a medical consultant, and the other is enabling education. At least on the latter point, relying on iFLYTEK's years of accumulation in the education industry, Xinghuo V35 did.
It is not difficult to see that in the development of large models, iFLYTEK Xinghuo has done itGrasp with both hands, both hands should be hard
On the one hand, iFLYTEK continues to invest in iterating the Spark model, promoting the continuous improvement of core capabilities such as natural language interaction, multi-scenario content generation, and voice, and strives to become an indispensable assistant in users' lives and work through active open testing.
On the other hand, under the guidance of the strategy of "platform + track", iFLYTEK adheres to the construction of artificial intelligence ecology, and strives to make iFLYTEK Spark benefit more industries, effectively improve the productivity of existing products, and at the same time lower the once unattainable threshold for social innovation and entrepreneurship.
(Source: iFLYTEK).
The first half of the large model is a technical contest, and the second half is the application landing, which must be realized from the arms race of technical parameters and the fun and cool demonstration demo to thousands of industriesOnly by landing applications, empowering scenarios, and serving human life, work, learning, and entertainment can we unleash the value that technology should have.
In the second half, Chinese players have the advantage of industrial scenarios, you must know that China not only has the most complete industrial system in the world, but also has a structure of universal inclusion in education, medical care and other fields, which provides a broad space for innovation in the landing and application of large models. At the same time, Chinese technology companies are better at doing "down-to-earth" landing, just like what iFLYTEK is doing.
In order to further accelerate the implementation of the large model industry, iFLYTEK also officially launched the first iFLYTEK Xinghuo open source large model that is deeply adapted to domestic computing power, attracting domestic underlying software and hardware ecological partners, industry leaders, and thousands of developers to jointly build a large model industry ecology.
The era of domestic large models is coming.
Digital Chinese New Year Challenge