MightyAIGCWith the rapid development of the trend and the massive infusion of capital, the AI track has been reborn in the past year and has once again reached the top of the technological context.
However, withLarge ModelCompared with the bustling battlefield, AIGC's pace in the commercialization scenario is a little hesitant. On the technical level, from text to image creation to video creation, AI has repeatedly delivered answers that refresh the public's vision, but how to make AI meet actual business expectations is still a complex and severe issue facing the industry.
And this may be the key to the victory or failure of the AI battlefield this year - only by moving from the laboratory to the market and realizing commercialization can we become the "backbone force" in the AI field.
January 30,iFLYTEKThe Spark Cognitive Big Model V3.5 Upgrade Conference was held - the domestically produced computing platform "Feixing No. 1", iFlytek Spark V3.5, voice big model, Spark Open Source-13B and other products were unveiled one after another. iFlytek also presented many big model commercialization results.
Apparently, the large model will be commercialized this year.FirstThe gun has already sounded, and as players continue to serve up the "main courses" of this feast, the first year of large-scale model applications will also arrive.
Industry constraints,How to break
Looking back at the AI wave in 2023, various players are competing fiercely in the field of large models, but the commercialization of AI is slightly slow. Even though various AI applications continue to refresh the public's vision, in the mature business field, the commercialization results of AI technology have not met expectations.
The logic behind this is that the immature AI technology has made application developmentVery highThe degree of uncontrollability and the scarcity of computing power have to some extent "added fuel to the fire", hindering the iterative development of large-scale model technology.
Take the "hallucination" problem that has plagued the industry for a long time as an example. If ordinary users see AI outputting "crazy words", they may just laugh it off. However, in serious business fields, AI's fallacies will be infinitely magnified and even cause "harm" to customer companies. In addition to the accuracy of the response, dimensions such as response time and cost also affect the commercialization of AI at all times.
To put it simply, today's AI seems to be dancing in shackles under a technological bottleneck.
Fortunately, historical experience has repeatedly told us that revolutionary innovation is difficult to achieve overnight, just like when DOS evolved into Windows, it faced questions in the technological context about “why a graphical interface is needed”; even the first generation of the iPhone, in the eyes of Nokia supporters at the time, was just a product that was not resistant to falling and was not durable.
However, this did not prevent Windows from witnessing Microsoft's legend, nor did it prevent the subsequent iPhone from lighting up the glorious moment of the mobile Internet.
Today, the rise of big models is pushing human-computer interaction in the Internet of Things to an unprecedented level, and this rotation process is deterministic.
Therefore, faced with long-term deterministic opportunities, how to solve the current problems of AI commercial applications and break the shackles of the industry has naturally become a top priority for the AI industry.
The key to breaking the deadlock lies in computing power. On the one hand, the difficulty of improving the technical capabilities of large models is mainly rooted in the diminishing returns of computing power - as the scale of the model gradually increases, the cost of additional computing power required to achieve the same improvement will gradually increase, and even become unaffordable.
On the other hand, due to geopolitical factors, the export of AI chips by semiconductor giant Nvidia has been restricted, and the US Secretary of Commerce has recently stated that "China will be prevented from obtaining computing power." Under external pressure, AI giants are more or less facing the problem of scarce computing power.
In other words, the domestic computing power base has become a key path for domestic large-scale model players to supplement computing power, achieve controllable computing power, and thus break the fateful cycle.
As a veteran AI player, iFLYTEK is aware of the challenges faced by the domestic AI track.
After joining hands with Huawei to create the "Spark All-in-One Machine" that can be privately deployed, iFlytek once again joined hands with Huawei to deeply build the large-model computing power base and jointly created the "Feixing No. 1" large-model computing power platform. The cooperation of the two "regular armies" has injected a strong dose of confidence for the domestic AI industry to break through the technological blockade. In this process, Huawei's excellent hardware foundation and iFlytek's deep AI foundation have achieved complementary advantages.
Under the shackles of the industry, this is not only a trump card to ensure the further iteration of large model capabilities and the improvement of application development controllability, but also marks the official start of an independent and innovative computing power revolution. The development of the domestic AI industry ecosystem has now had another choice.
iFLYTEKThree-way launch
If overcoming the computing power base problem is a key link for iFlytek and even domestic large-scale model manufacturers to enter the "midfield" of the track, then how to gain a better position in the AIGC midfield battle is a more complex proposition that varies from person to person.
At this stage, the competition among AI giants is rising to a higher dimension. The focus of players' competition is gradually shifting from whether or not to launch big models and big model parameters in the past to the real usability of big models, developer ecology, and commercialization track.
In this context, AI vendors such as iFlytek are standing at a crossroads of commercialization: direct competition for general capability output, open source to seize the ecosystem and application development, and each path is accompanied by huge challenges and room for imagination. iFlytek has chosen to develop all three paths simultaneously.
In terms of general capabilities of large models, iFLYTEK has always been at the forefront of the industry.
Last November, in the evaluation of the "Artificial Intelligence Big Model Experience Report 3.0" by the China Enterprise Development Research Center of the Xinhua News Agency Research Institute, iFlytek Spark Cognitive Big Model won the championship again and also won the three evaluation indicators of basic ability index, IQ index and tool efficiency index.First.
At present, iFLYTEK has trained iFLYTEK Spark V3.5 based on the "Feixing No. 1" computing power base.
It is reported that compared with iFlytek Spark V3.0, its core capabilities such as logical reasoning, language comprehension, text generation, math answering, code, and multimodality have been significantly improved. Among them, language comprehension and math capabilities have surpassed GPT-4Turbo, while code capabilities have reached 96% of GPT-4Turbo, and multimodal comprehension capabilities have reached 91% of GPT-4V.
This means that domestic large models are gradually moving away from the narrative of "constantly catching up with GPT-4" and surpassing industry benchmarks in more and more technological fields.
However, if we want to surpass GPT-4 in all aspects, in addition to facing the existing shortcomings and defects, we also need to be more open and jointly promote the technological prosperity of the domestic AI industry. To this end, iFLYTEK launched the "Spark Open Source-13B" plan, which deeply adapts to domestic computing power and helps developers, universities, and enterprises to conduct independent research and development, thereby opening up the way for joint construction.FirstA channel for the developer ecosystem.
On the other hand, iFLYTEK’s understanding of the commercialization of large models does not remain on the surface of the industry, but has truly integrated it into its own texture.
Throughout the development history of iFLYTEK, "voice" has always been its unavoidable mark. Over the years, iFLYTEK has always maintained the innovation and iteration of core technologies in multiple fields such as speech recognition and speech synthesis, and has won many relevant international awards.authorityTournament and review champion.
Based on this, as the technology context enters the era of big models, iFLYTEK, which has deep roots in the field of intelligent voice, has launched the domesticThe firstSpeech Big Model - Spark Speech Big Model.
It is reported that the Spark Speech Model has surpassed OpenAI Whisper V3 in speech recognition performance in dozens of mainstream languages, and in terms of multi-language speech synthesis, the humanization degree of the Spark Speech Model has exceeded 83%.
From an industry perspective, iFLYTEK launched the voice big model based on its insight into the new demands emerging in the era of the Internet of Everything.
In the era of the Internet of Everything, voice is undoubtedly the gateway to human-machine communication. This means that the voice big model will seamlessly adapt to popular and cutting-edge tracks such as smart connected cars, smart homes, and companion robots, reshape current products and businesses, and play the role of "1+1>2", thus pushing human-machine interaction to a new level.
As a veteran smart voice manufacturer, iFLYTEK has been in this field for a long time. While leading in technology, it can also reduce application costs to the price of "tap water", so as to better embrace the new opportunities brought by the "AI + Internet of Everything" era.
Therefore, iFLYTEK’s launch of the Spark Voice Big Model is by no means a “hard-core attempt” to take advantage of its own core track, but rather a well-considered answer from a pragmatic player.
Big model, startDiving
Whether it is breaking the shackles or iterating and innovating at the technology and product level, it is ultimately inevitable to circumvent the real assessment at the commercial level. If the technology upgrade cannot be exchanged for "real money", then the so-called AIGC wave is likely to evolve into a "reinventing the wheel" technology self-entertainment, resulting in a waste of manpower and time costs.
Looking at thousands of industries, it is not that the industry lacks the demand for the application of AI, but that there is a certain gap between the capabilities provided by large model players and the real needs of the same industry.
Taking the education scenario as an example, if we build a big model for primary school students to learn English, then compared with the conventional evaluation dimensions such as big model parameters and big model technical scores, whether the vocabulary range and sentence types are beyond the scope and whether they can be grafted into the primary school English learning goals may be the key to determining its actual application effect.Very high, and the response is extremely fast, but if it blindly outputs obscure and long sentences to primary school students, it is obviously not usable.
In essence, the commercial application process of large models is actually a process of examining players' cognition and understanding of industry needs.Based on this, many large model players are always accustomed to burying their heads in their own "three-acre land", thus sinking into the "comfort zone".
Business entities that fully adapt to their environment often face complex challenges in the process of migrating to a new business ecosystem. At present, AI is gradually becoming an indispensable presence in all walks of life, and if players narrow their path, they will inevitably limit their own development space.
iFlytek has clearly understood this logic: to commercialize a large model, one must have both the ability to hit the target and the ability to draw the target. iFlytek's past experience in enterprise services has enabled it to more accurately understand industry needs and create large models that are truly applicable to the industry.
From a vertical perspective, iFLYTEK's moat in educational hardware and voice is deepening thanks to its industry-leading large model capabilities.
Data shows that iFLYTEK AI learning machine users’ Net Recommendation Score (NPS) continues to maintain the industry’sFirst, and won the 2023 Tmall & JD Double Eleven sales champion; and with the support of the Spark voice model, iFLYTEK translation machine also ushered in an upgrade in interactive capabilities, bringing the user experience to a higher dimension.
At the same time, iFLYTEK has also reached a cooperation with China Mobile to launch a new 5G call business shorthand - mobile 5G users can synchronize call content minutes and accurately extract key matters on all mobile phones without downloading any apps, providing another perspective for the payment logic mining of the large model track.
From a horizontal perspective, iFLYTEK has accumulated a lot of experience in smart education, smart healthcare, smart cities and other fields. This enables it to better understand the essence of the industry on the road to commercialization of big models, and thus come up with big models that are truly suitable for different industries.
It is reported that based on its understanding of various vertical industries, iFLYTEK has established in-depth cooperation with leading enterprises in various industries, including Pacific Insurance, Bank of Communications, State Energy Group, Chery Automobile, etc. The cooperation with leading enterprises is like an anchor point in the relevant industrial chain, which promotes the positive cycle of commercialization in thousands of industries.
Ultimately, if AI wants to illuminate the future, its first task is to illuminate the present. Looking back at the long development of the AI track over the past decades, the current AIGC craze is nothing more than a "glimpse". This marathon of technology and commercialization has actually just begun. But what is certain is that iFlytek, which has the ability to generate its own blood and fight in the long term, has already obtained a ticket to break through the siege.