A launch built around a big ambition
On 18 February, xAI presented Grok 3 on X. The presentation opened with a statement of intent: the company says it wants to understand the universe. For a team led by Elon Musk, whose other ventures include SpaceX, that framing is coherent. It gives the model a longer-term mission rather than just a product launch.
The xAI team also put its people front and centre. Engineers and researchers took the stage, including one who had previously been at DeepMind and OpenAI. The message was direct: if you are talented, they want you. I found the emphasis on empathy particularly notable. The team said Grok 3 was designed to feel a little more human in conversation than other AI systems. That is a strong claim, but it is one I can test.
Scale, power and a very fast ramp-up
xAI explained that continuing to accelerate would require taking hardware and infrastructure into its own hands. The presentation cited cooling, power and GPU supply as the main bottlenecks. The team described building large clusters and using an amount of energy comparable to a small town. They also said they had added thousands of GPUs within a 90-day window.
I have seen similar energy debates in the Bitcoin mining space. Large-scale computation always collides with power availability. The claim of such a fast ramp-up is impressive, but I treat it as a company statement. The real test is whether that infrastructure translates into model quality over time.
Benchmarks: real progress, but watch the axis
xAI showed benchmark results, including MMLU, and said Grok 3 was progressing at unprecedented speed. They argued that even the smaller Grok 3 Mini reaches the frontier of competitor models. The presentation also described a blind test in which users compared answers from different chatbots without knowing which model produced them. According to xAI, Grok 3 came out first.
I pay attention to how charts are drawn. On the graph xAI displayed, the y-axis started around 1300 and went up to about 1400. If you start that axis at zero, the differences between the top models look like one or two percent, not a crushing lead. That does not mean Grok 3 is weak. It means the marketing visual exaggerates the gap. The progress is real, but the distance between the leaders is narrower than a first glance suggests.
DeepSearch and the move from chat to agents
One of the more interesting announcements was DeepSearch. xAI described it as a first-generation agent product aimed at engineers, researchers and anyone looking for answers. What I liked is the way DeepSearch exposes its reasoning steps. It shows the sources it consulted and the path it took, a little like reading a book of thought.
That transparency matters. When an AI gives me an answer, I want to know where the information came from and how the model reached its conclusion. I am also curious whether DeepSearch will be able to pull directly from X posts in a useful way. That would be a real differentiator, because X hosts a large amount of real-time discussion.
My practical test: Bitcoin, robots, energy and chips
I decided to test the models in Chinese, because that is the language I use with my audience. The first topic was Bitcoin. Grok 3 immediately opened a thought process and, within a few sentences, had checked around 50 web pages. That speed of retrieval was striking. On the Bitcoin price, one of the models gave a range from $92,000 to $105,000. I find that reasonable and rather conservative, given the market's usual volatility.
I also asked what could happen if Elon Musk integrated Grok 3 into robotics. Grok 3 read posts on X, discussed AI and robot technology, and brought up energy constraints. It mentioned robots already working in factories and the possibility of sending them to Mars. The answer was useful because it connected chips, energy and real deployment. That reminded me of the discussions I follow in the Bitcoin world, where energy and hardware are central.
DeepSeek and ChatGPT in the same test
DeepSeek answered the same questions correctly and also displayed a thought process. It was fast, but Grok 3 felt slightly quicker. On the AI chip question, DeepSeek gave a solid answer that looked similar in quality to Grok 3. I would not say DeepSeek lost; it was simply very close.
ChatGPT was more concise. Its response was shorter, but it included links. One detail impressed me: for the same Bitcoin question, ChatGPT used Chinese-language sources, while Grok 3 had used English-language sources. That made ChatGPT's answer more adapted to my Chinese-speaking audience. For Bitcoin-related questions, I found ChatGPT slightly stronger. Overall, Grok 3 still felt a small step ahead because of its speed and the depth of its reasoning display.
The current leader may not stay in front for long
The launch reinforces something I have said for a while: the AI race is accelerating in a way that resembles crypto cycles. Grok 3 is impressive, and DeepSearch is a genuine step toward more transparent agents. At the same time, the benchmark charts can flatter a model. When I look at the underlying differences, DeepSeek and ChatGPT are much closer than the presentation suggests.
I am not ready to declare a single winner. The practical advantage will come down to accuracy, speed and source transparency, and each model has different strengths on those fronts. What excites me most is the competition itself. xAI, DeepSeek and OpenAI are pushing one another, and that pressure should produce better tools for users quickly. The current leader may not hold that position for long, and that volatility is what makes this moment so interesting.