Zoom: AI Prodigy or Copycat? Acing the Hardest Exam but Facing Accusations of Cheating

Zoom Video Communications, a company renowned for facilitating remote work connectivity during the pandemic, made headlines last week with a groundbreaking achievement in the realm of artificial intelligence. The San Jose-based company proudly announced that its AI system had achieved an unprecedented score on the challenging Humanity’s Last Exam, surpassing the previous record set by Google’s Gemini 3 Pro.

According to Zoom’s Chief Technology Officer, Xuedong Huang, the company’s AI system scored an impressive 48.1 percent on the Humanity’s Last Exam benchmark, marking a significant improvement over the previous state-of-the-art result. This achievement sparked both surprise and skepticism within the technology industry, prompting questions about how a video conferencing company managed to outperform established AI giants like Google, OpenAI, and Anthropic on such a demanding benchmark.

Zoom’s success can be attributed to its innovative approach to AI development. Rather than training its own large language model, the company adopted a federated AI approach. This approach involves routing queries to multiple existing models from reputable sources like OpenAI, Google, and Anthropic, and then using proprietary software to select, combine, and refine their outputs. At the core of this system lies Zoom’s Z-scorer, a mechanism that evaluates responses from different models and selects the most suitable one for each task.

By leveraging a federated approach and combining diverse models, Zoom effectively built an advanced traffic controller for AI, demonstrating a unique method of utilizing existing systems to achieve exceptional results. This approach sets Zoom apart in an industry where the focus often revolves around building the most powerful individual model.

The response to Zoom’s achievement within the AI community was diverse. While some praised the company’s innovative approach to model federation, others raised concerns about taking credit for work done by other AI laboratories. Despite the debate, Zoom’s success highlights the importance of collaboration and integration of multiple AI capabilities in driving technological advancements.

Xuedong Huang’s background as a veteran AI expert from Microsoft underscores Zoom’s commitment to AI innovation. His strategic leadership at Zoom signals a shift towards a more collaborative and integrative approach to enterprise AI development, focusing on maximizing the capabilities of existing models rather than solely focusing on building proprietary models.

The Humanity’s Last Exam benchmark, designed to challenge the world’s smartest machines, has set a high bar for AI performance. Zoom’s achievement on this benchmark reflects the company’s dedication to pushing the boundaries of machine intelligence and exploring new frontiers in AI development.

Zoom’s approach to AI orchestration carries significant implications for the future of enterprise AI. By positioning itself as an integrator of the best AI capabilities from various providers, Zoom aims to provide customers with the most advanced AI solutions without being tied to a single model. This strategy mitigates the uncertainty of rapidly evolving AI technologies and offers customers access to the best available AI solutions for their specific needs.

Ultimately, the success of Zoom’s AI strategy will be determined by the real-world impact on its users. As AI Companion 3.0 is rolled out to Zoom’s extensive user base, the effectiveness of the AI in enhancing meeting experiences and productivity will be the ultimate test of Zoom’s innovative approach to AI integration.

In conclusion, Zoom’s pioneering achievement in AI highlights the importance of collaboration, integration, and practical application of AI technologies. By leveraging existing AI capabilities and developing sophisticated orchestration systems, Zoom has positioned itself at the forefront of enterprise AI innovation, setting a new standard for how AI technologies can be effectively utilized in real-world scenarios.