Introduction: On February 18, 2023, at noon Beijing time, the Grok-3 model developed by Elon Musk’s AI startup xAI was officially released. Previously, Musk described Grok-3 as “the smartest AI on Earth”. He stated on the X platform: “I spent the entire weekend refining the product with the team.” It is reported that on the backdrop of the Grok-3 launch event, it was written “our mission is to understand the universe”. Musk has stated that xAI’s goal is to “understand the universe”. So far, the live broadcast has attracted over 3 million viewers.
Image: xAI live broadcast on the X platform with Musk present (Source: X platform)
During the live broadcast, Musk stated that the name Grok comes from Heinlein’s novel “Stranger in a Strange Land”, where the protagonist is a human raised on Mars, and the word Grok represents a deep and comprehensive understanding of things.
The xAI team introduced that Grok-3’s performance has improved by an order of magnitude compared to Grok-2, and chatting with Grok-3 is very interesting.
The xAI team demonstrated during the live broadcast that Grok-3 and Grok-3 mini scored higher than or comparable to competitors like Gemini and ChatGPT in multiple testing metrics. During the training of Grok, the team built a large computing cluster and overcame challenges related to heat dissipation and power supply, taking 122 days to get the first batch of 100,000 GPUs operational and now they plan to double the size of the cluster.
The team also demonstrated generating code with Grok-3 and running this code. The screen displayed an animation of a spaceship traveling back and forth between Earth and Mars. Subsequently, Grok-3 demonstrated creating a game similar to Tetris. Musk stated that xAI will launch an AI game studio.
Image: Screenshot of the animation of the spaceship traveling between Earth and Mars during the live broadcast (Source: X platform)
Additionally, xAI announced the launch of a smart search engine called Deepsearch based on Grok-3. The name is quite similar to the recently popular Deepseek.
What Makes Grok-3 Smart?
During the previous World Government Summit, Musk stated via video call that Grok-3 is “the smartest artificial intelligence on Earth”. He mentioned that Grok-3 will have significant improvements in reasoning, programming capabilities, and multimodal abilities such as text and image analysis, and its performance “surpasses all currently released products”.
Musk emphasized that Grok-3 can reduce AI hallucinations by checking data back and forth and attempting to achieve logical consistency. He also revealed that the computing power used to train Grok-3 is far greater than that of previous versions and utilized a large amount of synthetic data.
Unlike the algorithm optimization path of DeepSeek (DeepSeek-V3 used 2048 H800 GPUs for 2788 thousand hours of training), xAI revealed that Grok-3’s development benefited from the Colossus supercomputer built over 8 months, powered by 100,000 NVIDIA H100 GPUs, providing 200 million GPU hours for training, more than ten times that of Grok-2.
In July 2023, Musk founded xAI. In November 2023, xAI released the first large model Grok-1, which had 314 billion parameters, becoming the largest open-source large language model at the time; Grok-2 released in August 2024 competed with the latest model of ChatGPT in performance. This series of large models can access real-time information via the internet and browse and use information on the X platform (formerly Twitter), making them timely in information acquisition.
AI Large Models Competing
Musk’s xAI has become a strong competitor in the AI large model field, competing with OpenAI, Google, Anthropic, and the recently globally noticed Chinese startup DeepSeek.
OpenAI recently announced that the company will launch the next generation of AI models GPT-5 and GPT-4.5 in the coming months. It is said that GPT-5 will integrate several core technologies from OpenAI, including the o3 reasoning model.
OpenAI’s CEO and co-founder Sam Altman stated early this morning on social media: “For high-demand testers, the experience of trying GPT-4.5 is far deeper than I expected!” Industry insiders expect this indicates that GPT-4.5 has entered the testing phase and is very close to official release.
In early February, after the release of DeepSeek’s new model, Google released the Gemini 2.0 series of models, which enhanced coding and reasoning capabilities, made them fully open for use, and reduced costs.
Additionally, it is reported that Anthropic plans to release a new hybrid large model Claude 4 in the coming weeks, allowing users to control the reasoning costs during use.
This week on Monday, Mistral, based in Paris, released a custom large model named Mistral Saba, characterized by high accuracy in Arabic interaction.
Since this year’s Spring Festival, the Chinese startup DeepSeek has released a new model DeepSeek-R1, which surpasses OpenAI’s benchmark model with extremely low training and usage costs, stirring up competition in the AI large model sector and initiating a wave of large model integration across various domestic industries.
According to incomplete statistics, hundreds of companies have officially integrated the large model developed by DeepSeek. This includes the three major domestic telecom operators, over 15 chip manufacturers, and more than 200 enterprises in various sectors including cloud services, internet technology, finance, several smartphone manufacturers and automotive companies, local government administrative systems, as well as world-class cloud computing giants like Microsoft, NVIDIA, and Amazon.
Tencent Group confirmed that WeChat began gray testing the integration of the DeepSeek-R1 model to enhance its search functionality starting February 15. This news caused Tencent’s stock price to surge. On the evening of February 16, Baidu Search announced it would fully integrate the latest deep search capabilities of DeepSeek and Wenxin large models.
Smart, Free
Large Models Accelerate Into Daily Life
It is noteworthy that driven by DeepSeek, large models are moving towards becoming increasingly smart and free, which also accelerates their integration into daily life.
Baidu officially announced that Wenxin Yiyan will be fully free starting April 1, allowing all PC and APP users to experience its latest model, including ultra-long document processing, enhanced professional search, advanced AI painting, and multilingual dialogue functions. In the early hours of February 13, OpenAI also announced the latest news about GPT-5, stating that it will launch the GPT-5 model in the coming months, allowing free version ChatGPT to use GPT-5 for conversation without restrictions under standard intelligent settings. Additionally, Google also announced earlier that its latest AI model suite Gemini 2.0 is officially open for use by all users.
Engineer Gong Zheng from the Institute of Technology and Standards at the China Academy of Information and Communications Technology stated that the maturity of AI technology is rewriting business models, and the rise of open-source models like DeepSeek is reconstructing industry ecology. OpenAI’s CEO Sam Altman predicted that the cost of using AI will decrease tenfold every 12 months.
How will free large models make money? An investment service professional focused on AI innovation told reporters:“AI large models have actually started a price reduction trend since last year, and large model companies have not yet completed commercialization.” He stated, “For AI large model companies to make money, they can provide solutions aimed at enterprises, and the valuation of large model companies is not based on the model but rather on their ecosystem.”
The aforementioned industry insider stated that for large model companies, future valuations will no longer solely rely on the model itself, but will focus more on ecosystem construction, user scale, data quality, and the profitability of value-added services. Enterprises with a large user base and a well-established ecosystem will have an advantage in future market competition.
The Industry Chain Welcomes More Opportunities
Guoxin Securities stated that the three major telecom operators will successively integrate DeepSeek after the New Year, and the vast data of telecom operators will help provide rich material for the training and optimization of DeepSeek’s models. The integration of telecom operators with DeepSeek will help develop new AI-driven businesses, and the cloud platforms of telecom operators are expected to achieve deep integration of AI capabilities, accelerating cloud business growth and promoting the second curve of growth for telecom operators.
In the content creation field, large models can help enterprises quickly generate copy, images, videos, etc., improving creative efficiency. In the intelligent customer service field, large models can achieve smarter interactions, enhancing customer satisfaction. In the financial sector, large models can be used for risk assessment, investment decision-making, etc., improving the operational efficiency and risk management capabilities of financial institutions.
Professor Liang Zheng from the School of Public Management at Tsinghua University, who is also the deputy director of the Tsinghua University AI International Governance Research Institute, mentioned in a recent interview that the future development of AI will trend towards terminalization and lightweight development. With advancements in multimodal and reinforcement learning technologies, large-scale deployment of service robots, autonomous vehicles, and drones will become possible.
With the rapid development of AI large models, more opportunities are emerging in the related industry chain. Robeco believes that in the short term, the semiconductor industry will face significant fluctuations, as the market reassesses the potential impact of DeepSeek’s technological breakthroughs on the broader AI ecosystem; but in the medium term, with the rise of proxy AI, the demand for advanced reasoning algorithms and their next-generation chips will continue to grow. Morgan Asset Management stated that it will focus on technology industries driven by artificial intelligence, new energy industries, high-end manufacturing, and healthcare industries mainly focused on innovative drugs.
Source: China Securities Report by Zheng Cuiying
Disclaimer: The content is sourced from the internet, WeChat public accounts, and other public channels. We maintain a neutral stance on the viewpoints expressed in the text and provide this for reference and non-commercial purposes. The copyright of reprinted articles belongs to the original authors and institutions. If there is any infringement, please contact us for deletion.
