xAI
Grok 5 training starts in a few weeks: xAI CEO
xAI Founder and CEO, Elon Musk, has announced that the Grok 5 AI model will start training in a few weeks, which could bring major upgrades to the performance.
Musk made this comment while referencing the ARC-AGI leaderboard update, highlighting xAI’s rapid development with Grok 4 topping the charts with 79.6% and 29.4% on ARC-AGI V1 and V2. These performances have surpassed models such as DeepSeek R1, which scored 57.7% two weeks ago.
The ARC-AGI benchmark tests general intelligence through efficiency and adaptability, showing a surprising gap in current AI models, as even high-performing LLMs struggle with human-like puzzle-solving, a finding supported by Chollet’s 2019 paper on the limitations of scale in deep learning.
A few months ago, xAI released Grok 4, its new frontier model with powerful reasoning and a multi-agent version for big queries and research. The company has used Colossus 1 with 200,000 GPUs to run reinforcement learning training to refine Grok’s reasoning capabilities at pretraining scale.
It’s revealed that the new improvements in this model training were a result of new infrastructure and a better algorithm that increased the computational efficiency of training by six times.
However, the next version is likely to surpass the current computational capabilities with the Colossus 2 supercomputer cluster. According to reports, xAI has already completed many milestones in the facility’s development, and it will employ more than 250,000 new GPUs, taking the total GPU count to 550,000 units. This new capacity will enable a more robust training and reasoning of the Grok 5 model.
For now, xAI is expected to launch a fine-tuned version of Grok 4, such as Grok 4.1 or Grok 4.2. However, Elon Musk has yet to press the launch button on this new model update.
(source)
