09/24 2024 531
"The greatest imagination of AI lies not in the mobile phone screen, but in taking over the digital world and transforming the physical world."
At the opening ceremony of the 2024 Cloud Town Conference on September 19th, Wu Yongming, Alibaba's earliest programmer and usually reserved individual, occasionally adjusted his glasses and offered a slightly bold prediction on industry trends.
For him, standing center stage, the conference marked three important milestones: the one-year anniversary of the filing of the base model Tongyi Qianwen; the one-year anniversary of the announcement of the "User-First, AI-Driven" strategy; and the upcoming one-year anniversary of Alibaba Cloud's "AI-Driven, Public Cloud Priority" strategy.
Therefore, this Cloud Town Conference was of great significance. From keynote speeches to exhibition setups, the AI-infused venue subtly hinted at the upcoming themes of his speech. The core tasks were twofold: reviewing the implementation of strategies, summing up achievements, and looking into the future.
Alibaba, driven by AI, has formed a "AI+Cloud" business model and seen new growth potential fueled by AI. Additionally, AI has transformed various existing business scenarios, including core e-commerce (Taobao and Tmall), international expansion (AIDC), and enterprise collaboration (DingTalk), unleashing its future potential.
Passionate about AI, Wu Yongming, returning to his comfort zone, also offered his technical perspective. He believes that AI's greatest potential lies in gradually infiltrating and taking over the digital world, with most things in the physical world acquiring AI capabilities, leading to the creation of next-generation products and synergies with the cloud-based, AI-driven digital world.
In a year's time, Alibaba, the giant ship, has found its course.
The wheels of technology and the market continue to roll forward, with AI newcomers battling in the red ocean of capital and established players scrambling for niches and actively seeking change. As GPT-o1 declares a new technical paradigm, where will Alibaba sail towards in its AGI future?
Shift: From Data to Large Model Capabilities
The pivot of giants from the previous era stems from the fundamental shift in the landscape.
During the internet era, the mantra was "data supremacy," with various internet companies leveraging data from social media, e-commerce, search engines, entertainment, and other platforms to complete the first stage of their development.
However, generative AI has shattered the traditional data-driven logic. Instead of solely relying on existing data for learning and decision-making, it generates new content and solutions autonomously through deep learning algorithms and massive computing power. As the technology matures, the generalization capabilities of large models will continue to strengthen, with the ideal scenario being a universal large model capable of solving all problems. Consequently, applications and capabilities derived from foundational large models will become the core competitiveness at this stage.
Large models and cloud computing form a "golden couple" in the AI 2.0 era, their bond extending from IaaS, PaaS, SaaS to the present MaaS (Model as a Service), where models become services. Underlying this development is a transition from informatization to digitization and, ultimately, to intelligentization. Cloud computing and large models naturally address the fear of digital transformation and intelligentization among users, who risk falling behind.
The emergence of the MaaS concept signifies the provision of independent tools and products akin to SaaS. Once scalability is achieved, standardized profitability becomes feasible, explaining why Alibaba views the "AI+Cloud" business model as viable. This, in turn, objectively enriches the value proposition of cloud computing by integrating AI.
Data indicates that Alibaba's "AI+Cloud" strategy is already yielding tangible results. In the second quarter of 2024, Alibaba Cloud's public cloud revenue grew by double digits, with AI-related product revenue surging by triple digits. Alibaba Cloud's quarterly adjusted EBITA grew by 155% year-over-year, and the number of paying users utilizing Alibaba Cloud's AI platform (Bailian) increased by over 200% quarter-over-quarter.
With the advent of large models, two categories of products have emerged: AI-native and AI+. The former is represented by chatbots, while the latter involves AI-driven transformations of existing businesses. AI assistants, a current focus for AI companies, are still considered "work-in-progress," necessitating further user market exploration. In contrast, AI-enhanced existing businesses are more likely to demonstrate tangible benefits, manifesting in both toB and toC forms. In the short term, 2B business demands, closely tied to "cost reduction and efficiency enhancement" and "productivity," are stronger and can more quickly penetrate corporate scenarios.
DingTalk's AI transformation has progressed through an "+AI" phase, integrating AI across various scenarios and leveraging AI capabilities to enhance productivity for customers. To date, DingTalk boasts over 1 million AI ecosystem partners, 500,000 AI agents, and over 10 million daily AI invocations.
Agents may be a crucial catalyst for activating China's SaaS and office market. Results-driven, future services essentially sell services rather than software. The "Software + AI + Collaborative Production" perspective gives birth to countless digital employees within enterprises, unleashing a "snowballing" effect on productivity.
The more consumer-oriented "AI+E-commerce" segment, due to its vast scale and complexity, experiences a more protracted value release. Considering Alibaba and Tencent's previous integrations of information flow (e.g., WeChat ads redirecting to Taobao and Tmall) and capital flow (e.g., Taobao and Tmall integrating WeChat Pay), the short-term AI-driven incremental growth in e-commerce may have been overshadowed.
In reality, the value proposition of "AI+E-commerce" unfolds in three layers: the first involves AGI-powered marketing creativity, leading to cost reduction and efficiency enhancement. The second focuses on after-sales service and repeat purchases, emphasizing the discovery of existing value, including smart customer service, logistics, and user operations. The third and ultimate layer represents the true value increment of AI in e-commerce, encompassing intelligent marketing strategies such as automated advertising and AI-driven business operations.
Notably, Alibaba's AI-driven omnichannel advertising product, "Omni-channel Promotion," was fully launched on Taobao and Tmall in August, with a performance release cycle of 6-12 months.
Alibaba's Commitment
Last year, the domestic "Hundred Models War" was already in full swing.
By the end of July 2023, China had witnessed the launch of 130 large models, with 64 of them unveiled in the first seven months alone. The so-called "AI Six Tigers" secured financing rounds, witnessing skyrocketing valuations and robust momentum. These signs point to the dawn of a new era.
Much like their international peers like Microsoft, Google, and Amazon, Alibaba must also grapple with the historical question of whether to bet on generative AI. This question carries particular weight for Alibaba.
Alibaba's exploration of AI unfolds along three paths: as a technology pioneer, probing the potential of machine intelligence to achieve AGI; as a provider of open-source models and leading cloud computing services, actively investing in large models to bolster its cloud business; and leveraging AI to drive its core businesses with the broadest range of application scenarios.
Alibaba appears unwilling to miss any opportunity. According to insiders, Wu Yongming possesses a keen sense for emerging technologies and acts decisively. Yuanjing Capital, of which he is closely associated, is known for its daring investments even before products materialize, thereby securing early market advantages.
Since the explosion of large models, three types of players have emerged: those catching up, those overtaking on the bend, and rising stars. This has led to distinct product offering models, such as cloud-based MaaS services, closed-source large model products, and open-source large model offerings. Among these, only Alibaba, with its cloud technology advantage, can offer all three, similar to Google globally.
Precisely because of its cloud capabilities, Alibaba has emerged as the only company that walks the dual paths of open sourcing and in-house development from the outset. Wu Yongming once explained during an earnings call that "open-source models gain an advantage when widely adopted by developers, aligning with Alibaba Cloud's AI infrastructure business model, as more developers will prefer Alibaba Cloud's AI products when deploying their applications."
Alibaba's exploration of large model implementations has forged an unparalleled "AI+Cloud" business model.
Many developers and startups praise Tongyi's open-source large model Qwen as "conscientious," offering usability in terms of model capabilities, size, and open weights. Increasingly, Qwen is replacing Llama-series models.
It is reported that the new generation of open-source flagship model Qwen 2.5-72B outperforms Llama 3.1-405B, reclaiming the throne as the world's top open-source large model. Tongyi's flagship model Qwen-Max has undergone comprehensive upgrades, approaching the performance of GPT-4o. Tongyi's open-source models have amassed over 40 million downloads, with the total number of native and derivative Tongyi models exceeding 50,000, positioning them as a world-class model group second only to the US-based Llama.
Over the past year, large model technology has achieved several milestone advancements. Last April, Alibaba Cloud unveiled its first large language model, Tongyi Qianwen, which has since expanded to encompass full-modality capabilities in language, image, video, and audio, underpinned by a comprehensive technology development system.
Currently, the two primary constraints on large models are computing power and inference costs. Benefiting from its integrated "AI+Cloud" strategy, Alibaba Cloud offers competitive pricing in the market, reducing both cloud and model usage costs while amplifying economies of scale.
In May, Alibaba spearheaded a significant price reduction in the large model industry, with the Tongyi Qianwen GPT-4-level flagship model Qwen-Long API experiencing a 97% price cut. At the Cloud Town Conference, Alibaba Cloud announced further substantial price reductions for three of its Tongyi Qianwen flagship models, with the highest reduction reaching 85%.
To date, no closed-loop business model for global large models has emerged, with cloud and advertising remaining the clearest and most established revenue streams. Coincidentally, Alibaba holds both of these cards. Betting on emerging AI companies resembles a "stone-throwing" exercise for large corporations, with the crucial aspect being how new technologies interact with businesses and ecosystems.
Alibaba's AGI Chips
Like most tech companies, Alibaba has set its sights on AGI (Artificial General Intelligence), envisioning "an ultimate large model that seamlessly integrates voice, text, images, and videos within a unified framework."
AGI, like a swaying blossom on the other shore, lures those on this side with its resplendent hues, while the treacherous waters and twists along the journey can only be navigated through exploration.
After 22 months of rapid technological advancements, large models have entered a temporary stage of consolidation.
Technologically, OpenAI's paradigm shift with the introduction of the GPTo series signals an attempt to seek a post-training Scaling Law, implicitly acknowledging bottlenecks in pre-training and data. OpenAI's "small test" offers the industry a new technical direction but also entails the risk of trial and error. At this early stage, no one can predict the path to AGI with certainty.
Following OpenAI's path, the barriers to entry for large models have risen, making it virtually impossible for new players to join the game. Companies that reach GPT-4o levels will secure early admission tickets to the second half, while those still in the game must continue to invest heavily to stay competitive.
Alibaba has already secured its admission ticket and aims to widen its winning margin. Alibaba and Tencent have invested in China's "AI Six Tigers," while abroad, Inflection AI joined Microsoft, and Character.AI aligned with Google. Large models are capital-intensive, making profitability challenging in the short term. Relying solely on capital, AI companies are better off with the active "blood transfusion" provided by large corporations. To date, Alibaba's AI commercialization roadmap has emerged, spanning from "Cloud+AI" to "AI+toB" and ultimately to "AI+E-commerce." With robust AI infrastructure and foundational models in place, the focus shifts to sequentially rolling out applications.
However, the industry faces challenges in its next phase of development.
ChatGPT's emergence has predisposed people to a particular form of AI assistant. The question remains whether it is necessary to confine all operational processes within a "box." For most companies, atomic capabilities such as voice, video, vision, and text-based dialogue are available. Rather than being constrained by form, exploring how to implement and combine these capabilities is more meaningful.
Furthermore, the integration of AI into specific businesses has largely been completed. While established players' original advantages remain intact in the AI era, deep integration is still in its infancy and far from sufficient.
Recent developments indicate deepening integration between endpoints. For instance, Quark has leveraged AI search to further bridge the gap between PC and mobile devices. Businesses like "AI+E-commerce," which permeate distribution, consumption, and ecosystems, offer vast potential for deeper integration and AI-driven transformations.
These unknowns represent infinite possibilities, perhaps explaining Wu Yongming's firm and bold predictions at the Cloud Town Conference.