China's innovative artificial intelligence company, Zhipu AI, faces increasing competition in the generative domain.
Article Remix:
Zhipu AI: China's AI Heavyweight and a Global Player to Watch
Let's dive into the tale of Zhipu AI, a Chinese AI champion by headcount, and one of the most highly-valued startups in the realm, boasting a cool $2.8 billion valuation as of September 2024. This powerhouse develops a motley crew of applications, from chatbots to advanced AI tools, all built around its General Language Model (GLM) series – large language models that can handle a wider range of tasks in both Mandarin and English.
Zhipu was one of China's first AI pioneers, sprouting from the fertile ground of Tsinghua University Science Park in 2019, led by professors Tang Jie and Li Juanzi. This duo has left quite a mark on China's AI landscape, with Juanzi directing Tsinghua's Knowledge Engineering Group (KEG), where Zhipu was born, and incubating other AI heavyweights like DeepLang and Moonshot AI. Tang, another Tsinghua professor, has mentored a legion of Chinese AI leaders, including two movers and shakers at Moonshot AI.
This early success set the bar high for China's AI industry, but the road ahead is far from smooth. Zhipu now faces stiff domestic competition and the challenge of closing the gap with global frontrunners like OpenAI. So far, they're keeping up the pace, fueled by aggressive pricing strategies, strategic hardware partnerships, and a diverse product portfolio. But as the market continues to evolve, they'll need to keep pumping out innovative ideas to stay on top.
Zhipu's Mission
Zhipu AI is on a mission to lead the global AI charge, setting its sights on artificial general intelligence (AGI) – a level of intelligence surpassing human capabilities. Led by CEO Zhang Peng, their ambition lies in creating "super-cognitive intelligence" that surpasses our own. While they share this ambition with giants like OpenAI and Google DeepMind, Zhang insists that their unique path lies in independent research and development.
Zhang sees their journey as a shift from quantitative change to qualitative change, driven by breakthroughs in large model development. With AGI on the horizon, Zhang envisions systems capable of self-explanation, self-assessment, and self-supervision, positioning Zhipu to redefine the global AGI benchmark.
In July 2024, Zhipu unleashed technology akin to OpenAI's Sora, a text-to-video generator hailed for its multimodal capabilities. While they acknowledged the technological gap earlier in May 2024, Sora served as a benchmark rather than a competitor to copycat. The quest for Sora-like systems, combined with their broader AGI goals, highlights their focus on multimodal AI as a crucial step towards AGI. Zhang carries a vision that extends beyond domestic domination, aiming to establish Zhipu as a global AI powerhouse.
What Sets Zhipu's LLMs Apart
At the core of Zhipu AI's unique approach lies their language model, GLM. While BERT from Google and GPT from OpenAI also rely on the transformer model architecture, Zhipu brings fresh ideas to the table by introducing a model that excels at handling both Mandarin and English with precision. To truly appreciate GLM's distinct qualities, let's compare it to its competitors:
- BERT favors "masked language modeling," a method that hides some words and trains the model to predict the missing words based on context from both directions (before and after).
- GPT utilizes an autoregressive model, where each token is predicted based on the preceding tokens, relying on a unidirectional approach that looks only at prior tokens to predict the next word.
- GLM combines aspects of both techniques by predicting the missing parts of text in sequence while also taking context into account like BERT, and generating text like GPT. Moreover, GLM is specifically designed to handle Chinese text effectively due to its ability to perform contextual analysis and adapt to the language's unique script.
While Zhipu's GLM excels in processing Chinese-language input, it seems to trail behind US models when it comes to advanced reasoning, decision-making, and adaptability. US leaders like OpenAI and Anthropic outperform GLM in benchmarks like SuperCLUE, a test that evaluates reasoning, logic, and versatility across a range of Chinese-language tasks. This difference may stem from US models' stronger capabilities in higher-order logic, creative problem-solving, and general adaptability.
Business Model
Zhipu primarily thrives on a Model-as-a-Service (MaaS) platform, allowing companies across various industries to fine-tune their language models for specific needs. In keeping with this strategy, they've selectively open-sourced some models, maintaining proprietary rights over others to protect their competitive edge. Notable open-source offerings include their 9B and multimodal 9B models, designed for text and visual tasks, as well as their end-to-end speech model, GLM-4-Voice, built for natural human-computer interaction.
To keep up with rivals in China's flourishing language model market, Zhipu has slashed prices for its proprietary models. In June 2024, they cut prices by more than half, with CEO Zhang Peng attributing the move to innovations in their core technologies that boosted efficiency and lowered costs. Partnerships with tech giants like Huawei and Qualcomm also play a crucial role in integrating their models with advanced chipsets and infrastructure, expand their market reach, and comply with China's data sovereignty rules.
Beyond models, Zhipu is expanding its footprint with applications tailored for business and consumer use. ChatGLM and AutoGLM are two examples, catering to users seeking efficient customer support and personal assistance in daily tasks like scheduling appointments or placing orders. In addition to providing powerful conversational AI capabilities, Zhipu has ventured into content creation with Ying, a text-to-video tool enabling users to generate engaging video clips from text prompts. These diverse solutions reflect Zhipu's larger goal of positioning itself as a key player in China's growing AI market.
In 2024, Saudi Arabia's Prosperity7 became the first foreign investor in Zhipu AI, injecting around $400 million during a funding round. As a result, many view Zhipu as an emerging threat to OpenAI's dominance in the global AI scene.
Conclusion
Zhipu AI's meteoric rise from academic roots to a dominant force in China's AI landscape highlights the potential of China's research-driven startup ecosystem. Zhipu has leveraged its academic ties to attract top talent, forge crucial alliances, and drive rapid innovation in the AI field. Its ambitious pursuit of AGI, combined with a versatile business model, positions the company as a formidable force in China's AI ecosystem. However, intense competition calls for the company to bridge technological gaps and adapt to the evolving demands of the global AI market. In light of these challenges, Zhipu emerges as a critical player in China's AI ambitions, poised to shake up the global AI landscape and challenge Silicon Valley's supremacy.
- Zhipu AI's aggressive research and development in artificial intelligence focuses on achieving artificial general intelligence (AGI), a level of intelligence surpassing human capabilities, as part of its mission to lead the global AI charge.
- The General Language Model (GLM) series developed by Zhipu AI sets it apart in the AI industry, as it excels at handling both Mandarin and English with precision, offering a unique combination of masked language modeling, autoregressive models, and contextual analysis that makes it effective for Chinese text.
- Zhipu AI's business model thrives on a Model-as-a-Service (MaaS) platform, allowing companies to fine-tune their language models for specific needs, with some models selectively open-sourced and others kept proprietary to protect its competitive edge.
- To expand its market reach and comply with China's data sovereignty rules, Zhipu AI has strategic partnerships with tech giants like Huawei and Qualcomm, aiming to establish itself as a global AI powerhouse in the face of stiff domestic competition and the challenge of closing the gap with global frontrunners.
