Zoom announces that its AI has achieved cutting-edge benchmarks, but experts criticize it as a 'hotspot' of other companies' AI



Zoom, a major video conferencing service, announced on its official blog that its AI feature, Zoom AI , achieved a cutting-edge score of 48.1% on the Humanity's Last Exam (HLE) , one of the world's most challenging AI benchmarks. However, experts have criticized the company, saying that it merely improved the benchmark by connecting multiple AIs together without directly delivering value to customers.

Zoom AI Achieves New Cutting-Edge Benchmark in Humanity's Last Exam (HLE) Benchmark | Zoom
https://www.zoom.com/ja/blog/humanitys-last-exam-zoom-ai-breakthrough/



Zoom says it aced AI's hardest exam. Critics say it copied off its neighbors. | VentureBeat
https://venturebeat.com/ai/zoom-says-it-aced-ais-hardest-exam-critics-say-it-copied-off-its-neighbors

On December 10, 2025, Zoom announced that Zoom AI achieved a state-of-the-art (SOTA) score of 48.1% on the HLE benchmark. This is 2.3 points higher than the previous highest score of 45.8% held by Google Gemini 3 Pro. Below is a table comparing Zoom AI's HLE benchmark with major AI models presented by Zoom.



According to Zoom, the key to achieving the SOTA score is the development of a 'federated AI' approach, which is completely different from traditional single-model-centric AI strategies. The federated AI approach integrates Zoom's own compact language models with advanced open- and closed-source models to link multiple language models and interactively collaborate to mutually generate, validate, and improve inference accuracy.

Zoom describes this achievement as 'the path to AI Companion 3.0.' AI Companion is Zoom's assistant AI. AI Companion 1.0 provided basic assistance functions such as summarizing meetings, while AI Companion 2.0 enabled cross-platform use, including external data integration with Gmail and Outlook and web search functionality through integration with Perplexity. AI Companion 3.0 advances a federated AI approach with agent capabilities, including information retrieval, text generation, and workflow automation, enabling unprecedented performance in complex reasoning tasks.

Zoom announces a number of new AI-powered features, including the ability to create videos using your own AI-generated avatar - GIGAZINE



In response to Zoom's announcement, AI engineer Max Rumpf criticized the company's claim of SOTA by combining API calls from Gemini, GPT, and Claude to slightly improve its benchmarks. While Rumpf acknowledged the usefulness of a federated AI approach that combines multiple models, he pointed out that 'Zoom does not train its own models, but in its announcement it obscures this fact and makes it appear as if it is taking credit for the work of others.'



Ryan Pream, founder of Exoria AI, an American AI startup, also expressed a negative opinion, saying, 'Zoom is just building another LLM-based harness and reporting it as an achievement. It's just noise in the HLE.'

Meanwhile, technology media outlet VentureBeat pointed out that Zoom's federated AI approach represents a vision for enterprise AI that is fundamentally different from the model-centric strategies pursued by OpenAI, Anthropic, and Google. As AI technology rapidly develops and the latest AI models emerge one after another, it becomes difficult for users to determine which model is optimal. However, by building an infrastructure that allows switching between providers, Zoom AI can theoretically provide 'customers with the best AI for any task.' It remains to be seen whether Zoom's federated AI approach will truly revolutionize the AI industry, or whether it will merely be a temporary combination of cutting-edge AI HLE scores.

Related Posts:

in AI, Posted by log1e_dh