Discover how a supercomputer and innovative training methods enabled the launch of a new version that could reshape the artificial intelligence landscape
xAI, founded by Elon Musk, unveiled the launch of the third version of its Grok chatbot on Monday, February 17. Musk describes this version as “terrifyingly smart,” pointing to the massive leap in analysis and reasoning capabilities. This development comes at a time when the world is witnessing an artificial intelligence race among major companies such as OpenAI and Google, making xAI’s announcement a prominent event in the tech industry.
Acceleration via the Colossus supercomputer
xAI confirms that the development of Grok 3 was made possible thanks to the supercomputer named Colossus. Built in just eight months, this computer relies on 100,000 Nvidia H100 processing units. According to the company, the system provided approximately 200 million GPU processing hours to train the robot—ten times what was available for the previous version, Grok 2. This leap in resources allowed for accelerating the model’s deep learning processes and reducing guesswork errors, thereby improving accuracy and increasing response speed.
Read also: Elon Musk’s company raises $6 billion to boost competition against OpenAI
Improvements in training curricula
The developments were not limited to the hardware side, but also included the training methods and techniques themselves. xAI introduced new curricula such as:
1. Synthetic data: Instead of relying solely on real data that may be limited or incomplete, xAI generates artificial data covering various scenarios to broaden the model’s learning scope.
2. Self-correction mechanisms: This feature allows the model to recognize its errors and compare them with stored correct answers, then autonomously improve its accuracy level through continuous learning.
3. Reinforcement learning: Based on the principle of reward and punishment, the algorithm receives a “reward” when providing correct answers and is penalized when errors occur, motivating it to gradually improve its performance.
Reducing hallucinations and enhancing accuracy
Minimizing what is known as “hallucination” represents a fundamental goal for Grok 3, as this term refers to incorrect answers or fabricated information produced by the model. xAI emphasizes that the current version uses multiple verification steps compared against trusted sources, which helps limit the proportion of incorrect information. In addition, human evaluation methods and context-guided training are added to ensure that Grok 3’s answers are as close to the truth as possible.
Read also: Tech giants: How Elon Musk and Sam Altman’s battle ignites competition in the AI world
Adopting human feedback and contextual training
xAI introduced a human feedback feature, where the model receives direct input from specialists to evaluate the quality of answers and their relevance to the context. On the other hand, Grok 3 takes into account previous information and background in the conversation, contributing to the generation of more detailed and realistic answers. These two steps aim to improve human-machine interaction and achieve a smoother, more useful experience for users.
Surpassing competitors?
Early tests indicate that Grok 3 may outperform competitors such as ChatGPT and DeepMind’s Gemini in complex tasks, especially in the areas of logic and reasoning. In an interview at the World Government Summit in Dubai, Elon Musk stated that “this might be the last time an artificial intelligence is better than Grok,” in a clear indication of his ambition to make Grok 3 the gold standard in the world of artificial intelligence.
The launch of Grok 3 marks an important milestone in the race to develop smart chatbots and reflects the technological acceleration witnessed in the AI world. While other companies strive to impress users with the capabilities of their models, xAI stands out by delivering a system that combines hardware power with advanced training curricula. The most important question remains how to effectively leverage these extraordinary capabilities without ignoring responsibilities of accuracy, credibility, and safety.
