Saturday, 2024 July 20

iFLYTEK’s Spark Cognitive Model: A Beacon or a Mirage in the AI Landscape?

Amid China’s recent AI surge, iFLYTEK, a pioneer in the AI industry, has spearheaded numerous national projects, earning its place as a leading player. Now, it aims to capitalize on the AI wave ignited by ChatGPT, leveraging its expertise for a competitive edge. On 6th May 2023, iFLYTEK launched the Spark Cognitive Model and introduced a “1+N” product strategy, heralding a new era in the field as it propels China’s AI advancement to global prominence.

Navigating iFLYTEK’s AI Evolution: A Three-Stage Progression

iFLYTEK, a prominent AI company in China, was established in 1999 and went public on the Shenzhen Stock Exchange in 2008. Its operational timeline can be traced through three distinct periods: the technology exploration period (2008-2014), the artificial intelligence 1.0 period (2015-2018), and the artificial intelligence 2.0 period (2019-present).

During the initial phase from 2008-2014, iFLYTEK focused on developing core technologies like intelligent speech, natural language understanding, and computer vision. In 2014, it initiated the Xunfei Super Brain Project, aiming to develop a cognitive intelligence system rooted in a “human-like neural network.”

In the AI 1.0 era (2015-2018), while continuing its core AI research, iFLYTEK diversified into new areas like image recognition, big data, and brain science, thereby enhancing its investment in the Super Brain Project.

As of 2019, the company transitioned into the AI 2.0 phase. Prompted by the Chinese government’s “double reduction” policy, which eases student workload by reducing homework and off-campus courses, many educational institutions turned to AI to enhance teaching quality. Similarly, iFLYTEK expanded its business model to B2B, B2C, and B2G, applying its technological advancements to specific scenarios, seeking large-scale profitability.

Indeed, this period witnessed significant financial gains for iFLYTEK, with a substantial rise in revenue and profit for two consecutive years. However, its inclusion in the U.S. Department of Commerce’s Export Control Regulations Entity List in 2019 hindered technical collaborations with U.S. firms, and the COVID-19 impact in 2022 resulted in project bidding delays, leading to a slowdown in growth and profit in 2022.

Nevertheless, spurred by the AI wave generated by ChatGPT, iFLYTEK accelerated its efforts in AI development. On 6th May 2023, it released the Spark Cognitive Model, sending its stock prices soaring. Then, on 9th June, at its 24th anniversary, the company unveiled the Xinghuo App and the iFLYTEK Xinghuo applet, both capable of multi-round AI dialogues, and also launched the V1.5 version of the Xinghuo cognitive model.

So, the question arises — can iFLYTEK leverage this AI boom and its Spark Cognitive Model to offset its current business challenges?

Exploring the Spark Cognitive Model: iFLYTEK’s Technical Edge and Potential Applications

iFLYTEK’s Spark Cognitive Model, a general AI tool poised against overseas contender ChatGPT, showcases seven key capabilities — multi-style long text generation, cross-lingual understanding, knowledge Q&A, logical reasoning, math ability, coding capability, and multimodal input.

iFLYTEK unveiled initial application scenarios for the model:

  1. AI+Education: An AI learning device to optimize and revise Chinese and English content.
  2. AI+Office: A smart notebook for organizing and drafting meeting transcripts and minutes.
  3. AI+Car: A human-machine interactive system for a multi-functional smart car cockpit.
  4. AI+Digital Employee: A virtual digital assistant for tasks like online marketing and automating workflows.
  5. Further explorations in medical, smart city, judicial, and industrial scenarios are planned for the future.

So, what contributed to iFLYTEK’s product development success? While many AI startups compete at the model and application levels, iFLYTEK leverages its foundational AI accumulation and AI industry chain, combining algorithms, data, and application scenarios, resulting in a notable competitive advantage.

Its solid foundation comes in the form of considerable R&D capabilities in large cognitive intelligence models and deep neural network algorithms. The company also assured its computing power capabilities, having built its own data centers since 2009, with deep learning platforms in four cities.

At the model and application levels, iFLYTEK has shown significant accomplishments since launching the Spark Cognitive Model. For instance, the AI Learning Machine released on May 31 offers Chinese and English content correction, personalized math instruction, interactive teaching assistance, and oral English training. This product quickly became the top recommended learning device in the industry, earning top sales on various e-commerce platforms.

Spark Cognitive Model: Outperforming ChatGPT in Chinese Performance

When iFLYTEK first unveiled its Spark Cognitive Model, it was benchmarked against the overseas leader, ChatGPT. So, how did it fare?

During a press conference, iFLYTEK performed a side-by-side comparison of ChatGPT and the Spark Cognitive Model using an evaluation system comprising 481 subtasks. The results indicated that the latter exceeded ChatGPT in Chinese performance. In the English test, the Spark Cognitive Model scored 4.29 out of 5, slightly behind ChatGPT’s score of 4.48. However, the Spark Cognitive Model demonstrated a superior understanding of intricate Chinese questions, with the ability to adjust its responses according to context.

This performance is attributed to iFLYTEK’s past experience in B2B, B2C, and B2G sectors, including the education and medical industries, allowing it to amass industry-related data. Collaborations with professional institutions and open platforms have further enriched its Chinese training resources for large model development. Furthermore, the Spark Cognitive Model and ChatGPT share a common underlying algorithm — the Transformer algorithm — which is extensively used in iFLYTEK’s speech and graphic recognition, as well as machine translation services. iFLYTEK has open-sourced a series of Chinese pre-trained language models across 40+ general fields, earning it the top spot on GitHub among similar Chinese models.

Additionally, Liu Qingfeng, iFLYTEK’s founder and chairman, announced plans to collaborate with over 3,000 development companies from 36 industries to create an expansive “Spark” ecosystem. This consistent, open-source, large-scale model adapted to the Chinese environment will help maintain iFLYTEK’s advantage in Chinese AI. Liu stated that as per scientific evaluation standards, the Spark Cognitive Model outperformed ChatGPT in mathematical ability and general knowledge responses, in addition to Chinese proficiency.

Lastly, according to financial media outlet CaixiniFLYTEK’s Spark App matched ChatGPT’s response speed. When presented with standalone questions, the Spark App demonstrated copywriting and life advice capabilities comparable to ChatGPT, such as crafting Dragon Boat Festival greetings or advice on considerations for moving into a new house.

iFLYTEK’s Challenges in Large Model Development

While the launch of the Spark Cognitive Model has generated considerable market buzz, it has also spotlighted potential crises.

First, iFLYTEK’s commitment to the entire AI value chain, while strategically creating competitive barriers, requires substantial capital investment. iFLYTEK’s profits in 2022 and Q1 2023 have yet to rebound, and growth remains sluggish. This raises questions about iFLYTEK’s ability to sustain the significant capital expenditure required for AI competition, including the training of large models and the development of novel applications.

Second, iFLYTEK is up against domestic internet behemoths such as Baidu, Alibaba, and Huawei, who have rolled out their own large models like WenxinTongyi, and the Pangu Series, respectively. These competitors, fortified by robust financing and vast AI computing resources, present a formidable challenge to iFLYTEK’s breakthrough efforts.

Finally, iFLYTEK has had to contend with two public relations crises within a month of launching its large model. Not long after the 6th May press conference, accusations surfaced online alleging that the Spark Cognitive Model was merely a “shell” ChatGPT. iFLYTEK refuted these claims on 11th May, calling them “unfounded and illogical.” Then two weeks later on 24th May, a post titled “iFLYTEK Suspected of Collecting Extensive User Privacy Data” went viral online. iFLYTEK responded that this post was a rumor, spread by competitors using a large model. Regardless of their veracity, these public opinion crises have significantly impacted iFLYTEK, leading to a 4.26% drop in stock price on the day of the latter crisis.

Evidently, iFLYTEK has some way to go before it can leverage its Spark Cognitive Model for a significant business breakthrough.

36Kr Global Writers
36Kr Global Writers
The tech ecosystem is roaring. Unicorns valued at billions of dollars have emerged worldwide, while venture capital and strategic investors are constantly on the lookout for the next big thing. 36Kr Global is committed to establishing ties between global stakeholders and providing the most vital information about China’s tech scene and capital markets.

Related Read