Grok 3, Xai's Next-Generation Primary AI Model, Is Not yet Available

Last summer, xAI founder and CEO Elon Musk announced that xAI would launch its next-generation AI model, Grok 3, by the end of 2024, planning to compete with models like OpenAI’s GPT-4 and Google’s Gemini. Grok 3 will feature image analysis and Q&A functions and support technical upgrades to Musk’s social platform, X.

In July 2023, Musk said that after being trained on 100,000 H100 GPUs, Grok 3 would be a major technological leap. But by January 2025, Grok 3 still hadn’t been released, and there were no signs it was coming soon. In an interview with podcast host Lex Fridman, Musk mentioned that Grok 3 would hopefully be released in 2024 “if we’re lucky,” showing his uncertainty about the timing.

The delay of Grok 3 is not unique. The AI startup Anthropic also failed to release its upgraded model, Claude 3.5 Opus, on time. Although the training was finished, they canceled the release due to financial reasons. Google and OpenAI have also faced similar challenges recently.

This trend may show the limits of current AI scaling methods. In the past, AI model performance improved significantly by increasing computing resources and data sets. But with each new generation, the performance improvements have become smaller, forcing companies to look for new technical paths. Musk admitted in the interview that Grok 3 may not fully meet expectations, which reflects this issue.

Also, xAI’s smaller team size could be one reason for the delay. However, this delay further suggests that traditional AI training methods may be reaching a bottleneck. In the future, AI companies and researchers may need to explore more efficient training techniques and innovative model designs to overcome these limits. At the same time, industry leaders should be more realistic when setting goals, considering factors like technology and the market, and avoiding being overly optimistic.

In 2025, the development of AI technology companies may give us some clues from the delayed releases of models, but it remains unknown when the speed of AI model development will hit a bottleneck.