xAI Launches Grok-2 Beta as it Targets ‘World’s Most Powerful AI’
Ramping up training capacity is seen as crucial to xAI's plans for AI dominance. Credit: Chesnot / Getty Images
Key Takeaways
- xAI has beta launched Grok-2 in a full-sized and mini version.
- While Grok-2 mini outperforms many other small models, the large model leaderboard is still dominated by Google and OpenAI.
- To compete with Big Tech rivals, xAI has built a new training supercomputer.
xAI has beta launched its highly anticipated language model Grok-2, which beats Gemini Pro 1.5 and comes close to GPT-4o level performance on several benchmarks.
However, with ever-more sophisticated models setting new benchmarks faster than ever, the company’s ambitions for AI dominance rest on a new GPU supercluster that CEO Elon Musk hopes will give it a crucial edge over the competition.
How Grok 2 Compares to Similar Models
Compared to the largest models from OpenAI, Meta, Google and Anthropic, the full-sized Grok 2 scores lower on the all-important Massive Multitask Language Understanding (MMLU) benchmark, ranking below GPT-4o, Llama-3 405B, Gemini 1.5 Ultra and Claude 3.5 Sonnet.
However, Grok 2 Mini outperforms GPT-4o Mini, Llama-3 8B, Gemini 1.5 Nano and Claude 3 Opus on the MMLU benchmark (Claude 3.5 Opus has not yet been released.)
Across math-based and multimodal benchmarks, a similar pattern emerges. While Grok-2 ranks among the top 3 or 4 large models, the mini version consistently outperforms similar-sized rivals.
But if CEO Elon Musk is to realize his dream of building the world’s most powerful AI, leading the field of small models won’t be enough.
Training Capacity Crucial
Grok-2’s success in the small model arena is a powerful endorsement of xAI’s development strategy. But the company will need additional firepower if it is to develop models that can compete with GPT-4o and Gemini 1.5 Ultra.
According to most estimates , OpenAI’s largest model has something in the region of 1.8 trillion parameters. Amnd it is likely that Gemini Ultra also boasts over a trillion.
Training such behemoths requires computational resources smaller firms can only dream of and the biggest players in AI have built dedicated data centers equipped with hundreds of thousands of GPUs to carry out the work.
xAI Memphis Supercomputer
To compete with the huge training capacity of Google and the Microsoft-backed OpenAI, xAI built a new AI supercomputer—the Memphis Supercluster—consisting of 100,000 high-end Nvidia GPUs.
According to Musk, the Memphis supercluster, started training in July and he expects it to deliver “the world’s most powerful AI by every metric” by December.
However, xAI’s supercomputing edge may not last long.
Earlier this year, it was reported that Microsoft is investing $100 into a new AI training platform to replace the 10,000 GPU cluster it built for OpenAI in 2020. The new data center is expected to be completed by 2028 when it will be used to drive the next stage of Microsoft and OpenAI’s model training.
Elon Musk Claims New Supercomputer ‘Most Powerful’ AI Training System—But for How Long?
Irish Watchdog Puts Grok Under Fire, Takes Musk’s X to Court Over EU User Data
