On February 18, Elon Musk's Xai announced the Grok 3, marking the startup's entry into an advanced AI reasoning contest led by players such as Google (GOOGL), OpenAI and China's DeepSeek. In live streaming on X, Musk described Grok 3 as “AI of the greatest truth,” even when challenging political correctness and “a significant leap in reasoning and computing efficiency”, it prioritizes accuracy.
Musk claims that Grok 3 outperforms AI’s top AI models, including GPT-4O, Gemini 2 Pro and DeepSeek-V3, scored over 1,400 points on LMARENA’s CHATBOT ARENA, an open source developed by UC Berkeley Skylab AI Bench Marked Leaderboard.
Grok 3 introduces new “thinking mode” functionality for real-time problem solving and “brain mode” to calculate heavy tasks. Outstanding feature is DeepSearch, an AI-powered research tool designed to compete with Google search and AI search alternatives, including OpenAi's Deep Research, DeepSeek's search mode, and Pro Search that confuses AI.
The AI model is an elite team of former large-scale technical researchers and engineers at XAI. Key members include Jimmy BA, former student at AI pioneer Geoffrey Hinton, and Yuhua “Tony” Wu, a former researcher at Google DeepMind. Leading the engineering job was Igor Babuschkin, a former engineer at Openai, who Musk personally recruited him to build a Chatgpt competitor. Babuschkin previously served as Senior Director of Engineering X.
Grok 3 is currently exclusive to members of X Premium+ and in the U.S., it costs $40 a month, and access to advanced features such as DeepSearch and Think Mode reasoning costs $30 a month.
What's behind the impressive features of the Grok 3'?
According to Xai's blog, Grok 3 uses test time calculation (TTCS) (TTCS) (specific test time scaling) to power inference. This machine learning strategy enables AI models to dynamically tune computing resources, ensuring higher accuracy for complex queries while maintaining simpler tasks.
Jeetu Patel, chief product officer of Cisco, said in LinkedIn Post that Grok 3's TTC approach could unlock breakthrough discoveries, including potential methods for lung cancer combined with dedicated computing clusters. “This is the first public sharing of a test-time computing cluster, whose inference model is multi-modal and can consume real-time data,” he wrote.
Additionally, Grok 3's development projects involve Colossus, a large supercomputer cluster built by XAI in Memphis, Tennessee. The system houses 200,000 NVIDIA H100 H100 GPU accelerator. Xai uses 100,000 NVIDIA H100 GPUs from Colossus to Train Grok 3, which offers 200 million GPU hours, a tenfold increase over the training setup of the Grok 2. Musk said the next generation of Colossus training clusters will increase by five times last year.
Some are skeptical of the sustainability of Xai's rapid progress, Musk's lofty claims about Grok 3 and the ability to compete with OpenAI's AI-Search competitors, such as Openai's in-depth research.
“In the long run, expanding computing power increases costs and makes model deployment too expensive, which is an unsustainable business approach.”
“The model absorbs everything they've been trained, including bias. I don't think AI will ever be truly neutral. She notes that if XAI continues to accelerate training, Grok 3 can be related to Google searches. “On the professional topics,” she said. On the one hand, Google Search has no advantage on domain-specific AI models (Groks) and trains on focused data. ”
Similarly, Andrej Karpathy, a founding member of OpenAI and former senior director of Tesla AI, conducted early tests on Grok 3, suggesting that while DeepSearch is better than Google's Gemini Models, it occasionally cited the hallucinations of references and URLs. “My impression of DeepSearch is about being confused about the in-depth research product, but not yet at the level of Openai's recently released “DeepSearch” which still feels more thorough and reliable,” he wrote in his X post.
