Elon Musk announced that his company's next-generation AI chatbot, Grok, might be released within weeks, describing it as "scarily smart" and claiming it has outperformed all other AI models in testing.

The CEO of xAI made these remarks at the Dubai World Government Summit on February 13.

"Sometimes, I think Grok-3 is a bit scarily smart," Musk said. "It proposes solutions you wouldn't expect—solutions that aren't obvious."

The developers of the chatbot employed a unique training method for Grok-3. Unlike ChatGPT, which uses real-world data, Grok-3 relies on synthetic data and employs a self-correction mechanism to maintain logical consistency. Musk claims that the system has become so accurate that even when presented with incorrect information, it reflects on the data and removes content inconsistent with reality.

The computational demands for training Grok-3 are enormous. It requires 200 million GPU hours, far exceeding the 2.7 million hours needed by its Chinese competitor DeepSeek-V3. It runs on xAI's Colossus supercluster, featuring 100,000 Nvidia H100 GPUs—a tenfold increase in computing power over its predecessor. Even without fine-tuning, Musk claims the base model performs better than Grok-2.

Grok-3's integration with X allows it to capture real-time social media applications rather than relying on web browsing. The system extracts live data from X and features what the company calls an "unrestrained mode"—described in xAI’s FAQ as being "intended to provoke offense, impropriety, and discomfort."

However, the system isn't entirely ready for deployment yet. Musk likened the remaining work to finishing a house: "The last 5% is doing drywall, painting, and landscaping—it may not be much work, but it changes the appearance of the house significantly."

Nevertheless, it may be released earlier than OpenAI's GPT-4.5. At least, Sam Altman indicated it could be launched within weeks or months.

"Grok-3 might be released in one or two weeks," Elon said. He did not specify whether the new version would be publicly released or require a subscription like Grok-2 initially did.

Competition in the AI field is intensifying. While ChatGPT dominated market share in 2024, China's open-source model DeepSeek-V3 emerged as a strong contender, performing better despite using far fewer resources than GPT-4o and Meta's Llama 3.1.

Grok was initially launched on X Premium, greatly limiting its availability. Later, it became free for all users of Musk's social media platform and launched a new independent website, now open to everyone.

xAI Enters the Reasoning AI Arena

Major AI players are shifting their focus to reasoning models, developing AIs capable of reflecting on specific problems and finding solutions after extensive chains of thought.

This idea was first explored by Matt Schumer with the announcement of Reflection 70b. The model was trained to incorporate chain-of-thought reasoning, and although it was just a fine-tuning of Llama 70b, it reportedly outperforms Claude 3.5 Sonnet in complex tasks.