Musk’s xAI Unveils Grok-3: The Dawn of AI’s "Hyper-Convergence" Era

Elon Musk’s artificial intelligence startup, xAI, founded in July 2023, launched its Grok chatbot and Grok 1.5 model six months later. By August 2024, Grok 2 was released, and on February 18 (Beijing time), the company officially introduced its next-generation AI system, "Grok-3". The name "Grok," derived from the science fiction novel *Stranger in a Strange Land*, signifies profound comprehension. Emblazoned on the backdrop of the launch event was xAI’s mission—“Understand the Universe”—echoing Musk’s vision of AI expanding the boundaries of human knowledge.  

During the livestreamed launch, Musk revealed unprecedented details about Grok-3’s development costs, disclosing that its training consumed "200,000 NVIDIA GPUs". Dubbed by Musk as the smartest AI on Earth, Grok-3’s breakthrough lies in its “chain-of-thought” reasoning" and "multimodal capabilities", enabling it to decompose complex tasks step-by-step like humans, thereby enhancing logical coherence and problem-solving proficiency. Across benchmarks in mathematics, scientific reasoning, and code generation, Grok-3 outperformed leading models such as "DeepSeek-V3", "GPT-4o", and "Gemini-2 Pro". For instance, in the "2025 AIME competition", Grok-3 achieved a record-breaking score of "93 points", showcasing its "superhuman" mastery of interdisciplinary knowledge.

Grok-3 was trained on "Colossus", a supercomputer developed by xAI over eight months, powered by "100,000 NVIDIA H100 GPUs" and consuming "200 million GPU hours". The model leverages synthetic data training, self-correction mechanisms, and reinforcement learning, optimized through human feedback loops to minimize AI “hallucinations.” Musk stated at the event, “Grok-3 surpassed Grok 2 in an extraordinarily short time—we believe it is an order of magnitude more powerful.” xAI engineers further noted that Grok-3 underwent "10 times the training scale" of its predecessor.

Live demonstrations highlighted Grok-3’s versatility: designing an original game merging "Tetris" and "Bejeweled", calculating optimal time windows for Mars missions, and excelling in programming, creative tasks, and scientific simulations. The newly introduced "Grok-DeepSearch" tool enhances information retrieval and fact-checking capabilities. Notably, in the "LMSYS Chatbot Arena"—a blind-testing platform—Grok-3 ranked first in code generation, instruction response, and reasoning tasks.  

Two critical distinctions set Grok-3 apart: its benchmark comparisons targeted "DeepSeek R1" (the latest iteration of DeepSeek) rather than older versions, and even the compact "Grok-3 mini" demonstrated remarkable performance.  

Grok-3’s launch transcends technological advancement—it is a manifesto for AI’s future trajectory. As OpenAI pivots toward open-source models, DeepSeek advances China’s AI ambitions, and xAI pursues its “cosmic-scale” vision, the AI race has evolved from singular model performance to a multidimensional contest encompassing technology, ecosystem development, and commercialization. In Musk’s words: “We can finally learn from each other”—a sentiment positioning Grok-3 as the starting point for humanity and AI to collaboratively explore the unknown.