“In the future, generative AI will revolutionize office software,” said Zhang Qingyuan, CEO of Kingsoft Office. He envisions a world where users no longer grapple with complex software interfaces but instead benefit from a smart "knowledge assistant" that guides their work.

The transformation and competition in this space have only just begun. For generative AI to penetrate households and businesses, its cost and accessibility must considerably decrease. It is only through persistence and strategic positioning that enterprises can emerge victorious in this fast-evolving landscape.

Decision: All in AI

Kingsoft Office has emerged as a pioneer in China's office software sector, launching its WPS Office suite long before Microsoft's Office in the late 1980s. At its zenith during the DOS era, WPS captured an impressive 90% of the domestic market. As mobile internet took off, Kingsoft adeptly pivoted towards mobile applications to meet diverse user demands, further enhancing its brand influence. The arrival of ChatGPT left Kingsoft in awe, emphasizing the sheer potential of technology to redefine AI capabilities; the billion-parameter models represented a leap towards true artificial intelligence.

Advertisement

Following the 2023 Chinese New Year, Kingsoft held extensive discussions regarding the implications of the generative AI wave. Although the trajectory of large models remains uncertain, they equated the magnitude of this transformation to the reinvention of electricity, a step that could truly elevate computers to the status of "computers." Consequently, by April, they committed to an "All in AI" strategy, recognizing that office applications are a compelling use case for AI.

The next challenge was determining how to go "all in." Capitalizing on their advantages became paramount. Kingsoft observed that training large models resembles nurturing a child; success often hinges on luck and inherent "genes" rather than sheer financial power. Thus, the feat accomplished by OpenAI with ChatGPT, rather than by giants like Google, Apple, or Microsoft, stood out.

After evaluating their resources and talent, Kingsoft decided to concentrate on application-level innovations. Their distinctive positioning as an application-layer provider is notable, underlined by the significant monthly active device count exceeding 600 million, a significant competitive advantage. Therefore, they opted to collaborate with major model developers, focusing on integrating AI within office software, encapsulating it succinctly: “They make the engine, we create the vehicle.”

In their efforts to expand internationally, Kingsoft turned to Amazon Web Services (AWS) for cloud solutions. AWS offers a robust and mature operational framework, particularly in model deployment, monitoring, and data management, alleviating much of the complexity for Kingsoft as they navigate the AI landscape abroad. This partnership is further rooted in a longstanding relationship established years prior when they ventured into international markets. With deployments across regions like the US West Coast, Singapore, and Japan, AWS provides the necessary stability. The intertwined nature of AI and cloud means that as cloud technology expands, so too should AI capabilities, making AWS one of their key engine suppliers.

Simultaneously, Kingsoft, alongside Xiaomi and Shunwei Capital, initiated a fund to invest in domestic large model developers, expressing a desire to “grow together, learn together, and keep pace with them” in this rapidly advancing field.

Exploration: Three Main Directions

What applications does generative AI hold for office scenarios? Kingsoft, leveraging its capabilities in AIGC, Insight, and Copilot, has delineated three main directions.

Generative content, which includes text and image generation, represents the most mature and widely applied ability of AIGC, notably in poetry and article writing, where it has surpassed the capabilities of most humans. Kingsoft utilizes AIGC to assist users in drafting outlines, refining, rewriting articles, as well as generating images and infographics.

Historically, computers excelled at managing structured data but struggled with unstructured language data. Figures in irregular tables, for instance, confounded computers, obscuring the valuable wisdom fundamentally embedded in what people express. Yet, in office environments, the amount of unstructured knowledge found in documents, presentations, and spreadsheets is vast, yet its reuse was negligible. Now, with the AI Insight capability blooming, there is renewed hope for understanding such unstructured data, which marks a monumental breakthrough.

The customer service within Kingsoft serves as one of its “users.” They receive numerous user emails daily. Previously requiring manual review, AI can now efficiently “read” these emails with impressive accuracy after several months of refinement. While AI can make mistakes, so too can humans, but the greater stability of AI leads to extensive potential. This capability opens avenues for document Q&A, retrieval, knowledge bases, and even the recreation of documents, like summarizing a book into a PowerPoint presentation, etc.

Copilot, with its AI coding abilities, provides a compelling revolutionary aspect and poses a threat to entry-level programmers, as it can autonomously write programs that surpass many novices. Its applications are seen primarily in automation and spreadsheets. Traditionally, “office automation” needed specialized partners to develop further on existing platforms; now, AI can handle those tasks efficiently. For instance, in spreadsheet contexts, with AI's support, users need not familiarize themselves with complex data formulas or pivot tables to obtain desired results. A simple request like, “List the operational conditions of stores in Zhuhai from yesterday,” or “Which products have abnormal sales data?” can efficiently generate necessary reports and charts.

Selecting the appropriate model based on business scenarios holds greater significance than merely chasing advancements in performance or parameters, as evidenced from Kingsoft's deploy of generative AI. For serving overseas customers, AWS’s Amazon Bedrock generative AI core service provides a comprehensive array of high-performance foundational models tailored for various scenes.

Upgrading: Integrating LUI and GUI

Since the launch of Kingsoft's WPS AI, millions have actively engaged with the platform over the course of a year. This usage has surfaced several issues.

In 2023, many believe that Language User Interface (LUI) is the future and that all software should transition towards this model, signaling the impending decline of the previously heralded Graphical User Interface (GUI).In reality, many users need guidance and prompts to express their needs effectively. Consequently, GUI will continue to coexist, as LUI must integrate seamlessly with GUI to help users articulate their demands clearly.

The second issue that arises is that content generated based on public domain knowledge tends to converge, rendering it less useful. To turn AI into a productivity tool and a competitive advantage, it is crucial to extend AI into the users' or companies' private knowledge domains. For instance, when a company prepares a semi-annual report or a roadshow presentation, the content should reflect all relevant aspects pertaining to their operations over the past six months. This forms part of the improvements in the WPS AI 2.0 slated for launch in 2024, gradually shifting from efficiency-oriented personal AI to creating an "enterprise brain" for companies and enhancing administrative processes for government offices.

The third challenge stems from some users harboring overly high expectations; such disillusionment is often experienced. Effectively managing user expectations is crucial when promoting new technologies and products. As a countermeasure, Kingsoft has opted to decompose AI functionalities into granular tasks to meet user needs more precisely, offering them numerous “options” instead of blanket solutions. For example, while writing, if a user is unsatisfied with a particular segment, they can opt to have AI “polish” it. Rather than providing an all-encompassing solution, AI engages the user by asking, "Would you prefer a more serious, more lively, or a more official tone?" Once a choice is made, AI generates alternatives and queries, "How does this revision meet your requirements?" If satisfactory, users can accept the option; if not, AI can provide additional suggestions.

Currently, Kingsoft's development strategy involves critically analyzing how to integrate AI to reconstruct every feature. Users have become accustomed to existing functionalities, and when AI can enhance user experience and solve challenges effectively, it signals that the approach is sound. Thus, Kingsoft has delineated clear boundaries for the application of generative AI, akin to placing constraints on Sun Wukong's abilities, ensuring that each function adheres strictly to its designated tasks—be it document creation, imaging, or refinement, avoiding distractions.

“The best AI applications are those where users are not even aware that AI is a component of the product.” In reality, most users do not concern themselves with the underlying technology, focusing instead on whether the product effectively resolves their issues and its usability. Marketing expert Theodore Levitt poignantly articulated, “People don't buy a quarter-inch drill; they buy a quarter-inch hole.”

In the early days of smartphones, capabilities were limited; it is only today that users discern the essence of native mobile applications. Simultaneously, they are exploring the potential forms of AI-native applications. As of now, there is no definitive answer to this inquiry, as technology and human needs will continuously evolve, with time revealing the ultimate answers.