===== Claimed to be the "smartest AI on Earth"! The large model Grok3 developed by Elon Musk's xAI has officially been released. =====

Wallstreetcn
2025.02.18 06:45
portai
I'm PortAI, I can summarize articles.

Chatbot Arena has the first model in history to break 1400 points, surpassing or matching competitors such as Gemini, DeepSeek, and ChatGPT. Grok3 excels in mathematical reasoning and scientific logical reasoning, and Grok3mini's reasoning performance currently outperforms Grok3. Musk stated that the Grok3 version will be open-sourced once debugging is mature. In addition, xAI has doubled its data center capacity in just three months and reached a $5 billion agreement with Dell to support Grok3 training

On February 18th at noon, Elon Musk's XAI held the Grok 3 launch event, which was watched by over 1 million people online. Musk praised it as "the smartest artificial intelligence on Earth."

The demonstration at the event showed that Grok 3 and Grok-3 mini outperformed or matched competitors like Gemini, DeepSeek, and ChatGPT in various performance aspects, including mathematical reasoning and scientific logical reasoning. Additionally, xAI introduced the more powerful Grok3 Thinking reasoning mode.

In terms of computational ability, Grok 3 also excelled in understanding and creativity. The demonstration included real-time solutions to complex physics problems, such as plotting interplanetary trajectories and conceptualizing video games...

Even more surprising, Musk revealed that the best experience of Grok-3 would be available "about a week later," with voice interaction features still in development, allowing users to have a conversational experience in a week.

Musk stated during the event that Premium Plus users on X would be the first group to gain access, while users could also subscribe to the Super Grok service separately.

Furthermore, to support Grok 3 training, xAI doubled its data center training cluster to 200,000 GPUs in just three months, and xAI reached a $5 billion agreement with Dell, which may deliver servers equipped with NVIDIA GB200 chips to xAI this year.

"The Smartest AI on Earth"

According to reports, Grok-3 achieved better results than DeepSeek-v3, GPT-4o, and Gemini-2 pro in multiple benchmark tests regarding capabilities in mathematical reasoning, scientific logical reasoning, and code writing. It is known that Grok 3 has been running internally at xAI for two weeks.

xAI engineers explained that although Grok started later, it has quickly caught up to ChatGPT in MMLU scores.

Musk and his team stated that Grok 3 will also have reasoning capabilities similar to DeepSeek R1 and OpenAI o3 Mini.

Musk's team pointed out that the pre-training of Grok 3 was completed about a month ago, and since then, they have been working hard to integrate reasoning capabilities into the Grok 3 model. The training time for the Grok 3 reasoning mini version is longer, and it performs slightly better than the Grok 3 reasoning model, indicating that the Grok 3 reasoning model has enormous potential

From various indicators, Grok 3 has surpassed all models and ranked first in the world, featuring reasoning patterns and deep research capabilities. Last week, Musk announced the launch of Grok 3 during a video call at the World Government Summit in Dubai, stating that the chatbot possesses "very powerful reasoning abilities" and is "the smartest artificial intelligence on Earth."

Grok 3 is ten times faster than Grok 2, with large-scale installations of more computing power allowing it to run large datasets in a shorter time while providing higher accuracy.

In a specific demonstration, Grok 3 generated animated 3D graphics for a space launch on-site, showcasing its ability to understand complex physical knowledge.

The Musk team input prompts requesting Grok-3 to generate code on-site, and after running the code, an animation of a spacecraft traveling back and forth between Earth and Mars appeared on the screen.

In another demonstration, Grok-3 created a game similar to Tetris and Bejeweled, showcasing its outstanding creativity.

Musk also revealed that xAI will launch an artificial intelligence game studio. If you are interested in joining us to develop AI-driven games, please join us.

xAI has achieved better functionality for Grok3 by modifying its training process (not just hardware improvements). The updated model implements synthetic datasets, self-correction, and reinforcement learning to enhance its performance.

Regarding subscription trials, Musk stated that the Grok3 beta is now open to X Premium users, and the SuperGrok subscription service has been launched.

Additionally, xAI plans to open-source previous versions of its Grok model immediately after the latest version matures, with Musk expecting the transition for Grok-3 to be completed within a few months.

xAI reaches a $5 billion agreement with Dell to double data center capacity in three months

It is worth mentioning that xAI plans to reach a $5 billion agreement with Dell to provide AI server support for Grok 3. xAI is about to finalize an agreement with Dell Tech to acquire AI-optimized servers equipped with NVIDIA GB200 chips worth over $5 billion, aimed at meeting the growing computing demands of Grok 3 and other AI applications Previous articles pointed out that insiders said if the deal is reached, Dell will deliver servers equipped with NVIDIA's GB200 chips to xAI this year to optimize AI work.

At the press conference, Musk's team also revealed that xAI doubled its data center capacity in just three months, using 200,000 NVIDIA H100 GPUs to create the best AI.

Musk's team stated that last April, Musk decided that for xAI to succeed and create the best AI, the only way was to build its own data center. It took us 122 days to get the first batch of 100,000 GPUs up and running. We quickly realized that to build the AI we envisioned, we needed to double the cluster size. Therefore, we initiated another phase, and in just 92 days, we doubled the capacity.

Netizens' Hot Comments: The First Model in History to Break 1400 Points, Cost-Performance Ratio Exceeds Gemini

Netizens are also excited about the release of Grok-3, with AI guru Andrej Karpathy praising it after trying it out, saying that Grok3+Thinking is roughly equivalent to OpenAI's strongest model (like o1-pro, with a monthly fee of $200), slightly better than DeepSeek-R1 and Gemini 2.0 Flash Thinking.

Some netizens praised that Grok-3 is the first model in history to break 1400 points in the Chatbot Arena, outperforming OpenAI and Google's best public inference models. xAI was established 13 years later than DeepMind and 8 years later than OpenAI, and is now ahead of both.

Some netizens expressed strong confidence in AI, stating that Grok3 seems very powerful, which well proves that the law of scaling has not actually ended, and they are very optimistic about the future of artificial intelligence.

Some netizens also pointed out its shortcomings, stating that Grok3 is not good at coding.

In addition, more netizens speculated whether this would trigger a new round of price wars among large models?