Elon Musk's 'Scary Smart' Grok 3 Set for Release
- tech360.tv
- Feb 18, 2025
- 2 min read
Renowned tech entrepreneur Elon Musk's artificial intelligence company, xAI, is gearing up to launch Grok 3 on February 18 at 12 p.m. Musk himself has described Grok 3 as "scary smart," boasting enhanced reasoning, computational power, and adaptability compared to its predecessor.
The development of Grok 3 was significantly accelerated with the assistance of xAI's Colossus supercomputer, which utilises 100,000 Nvidia H100 GPUs, providing a tenfold increase in computational resources compared to Grok 2. This substantial boost enables more efficient data processing, leading to reduced training times and improved accuracy.
Apart from the impressive computational power, Grok 3 introduces several key enhancements in its training methodology. These include the incorporation of synthetic datasets to simulate diverse scenarios, self-correction mechanisms to refine errors, and reinforcement learning to enhance decision-making through trial and error.
xAI asserts that these updates will minimise incorrect responses, known as hallucinations, by implementing validation steps and ensuring better logical accuracy. The model also integrates human feedback loops, where reviewers evaluate the relevance and accuracy of its outputs, and contextual training, enabling the AI to adjust its responses based on past interactions and user intent.
Preliminary tests have shown that Grok 3 surpasses competitors such as OpenAI's ChatGPT and Google's DeepMind Gemini in handling complex reasoning tasks. During a recent interview at The World Governments Summit in Dubai, Musk expressed his confidence in Grok 3, stating, "This might be the last time that an AI is better than Grok," underscoring the model's advanced capabilities.
Elon Musk's xAI to release Grok 3 with enhanced reasoning and adaptability
Grok 3 developed with the aid of Colossus supercomputer and 100,000 Nvidia H100 GPUs
New features include synthetic datasets, self-correction mechanisms, and reinforcement learning
Source: CGTN
