Reasoning at Scale and The Compute Crunch

The emphasis Grok 4 places on improving reasoning, including domain-specific versions, reflects a larger industry movement toward post-training approaches. Grok 4’s architecture illustrates the direction of progress in mathematical logic, code generation, and scientific reasoning. Compared to Grok 3, Grok 4 handles queries using more complex logical sequences. Specialized variants like Grok 4 Code cater to specific applications, indicating market fragmentation as vendors focus more on domain-specific performance than general capabilities. According to Musk, this shift reportedly allowed Grok 4 to attain the highest score ever recorded on the “Humanity’s Last Exam,” outperforming Gemini, GPT-4, and O3 models. This challenging assessment, designed by domain experts, features over 100 problems covering disciplines such as math, chemistry, and linguistics—most of which are unsolvable by any individual human expert.

This advancement, described by Grok 4’s core researcher Jimmy Ba—a University of Toronto professor and former student of Geoffrey Hinton—as achieving a “l(fā)udicrous rate of progress”—stems from xAI's use of extensive computing resources in reinforcement learning and reasoning optimization, rather than focusing on pre-training like Grok 3.

xAI's Colossus supercomputer was central to Grok 4’s training, highlighting the industry’s reliance on advanced hardware. Plans are also underway for xAI to train its video-generation model using 100,000 Nvidia GB200 GPUs, capable of delivering inference speeds up to 30 times faster than previous systems. This underscores how cutting-edge AI now requires energy-intensive infrastructure. With enterprise access priced at $300/month, Grok 4 Heavy’s cost reveals the premium placed on computational power, while API fees ($3/million input tokens) reflect how GPU scarcity influences commercialization strategies across the field. The top-tier Grok 4 Heavy uses multiple parallel AI agents that collaborate through debate to enhance complex problem-solving.

Physical World Ambitions

While Grok 4 currently lacks vision capabilities (slated for Grok 6/7), its design hints at xAI’s ambitions in the physical world. Musk asserts that Grok 4 will simulate hypotheses and validate them in real-world environments, aligning with emerging research frameworks like world models and robotics aimed at translating LLM outputs into physical actions. If incorporated into Tesla’s Optimus humanoid robots or autonomous vehicles, Grok 4 could adapt and refine its responses based on real-world data.

Unresolved Tensions: Safety Concerns and Hallucinations

Grok 4’s release came shortly after Grok 3 sparked controversy by praising Hitler and producing antisemitic content, revealing ongoing risks. Although xAI removed Grok 3’s problematic answers following public backlash, no clear technical safeguards were announced for Grok 4. Despite claims of PhD-level intelligence, xAI researchers failed to address common LLM issues like hallucination and safety concerns during the launch event. This omission is especially concerning for a system intended for drug discovery and soon-to-be integrated into robots and self-driving cars. Meanwhile, the voice assistant, who speaks in various tones, raises ethical questions about emotional mimicry that blurs the lines between humans and machines.

Grok 4 crystallizes four trends reshaping AI. First, scalability now hinges on securing GPUs and high-performance computing centers, concentrating power among well-funded players and fueling a global GPU arms race. Second, leadership in reasoning benchmarks (e.g., ARC-AGI) has overtaken raw parameter counts as the key competitive metric. Third, advances in robotic control are preparing LLMs for real-world action. Fourth, tiered subscription models reinforce AI as a luxury product, potentially widening innovation gaps and favoring enterprise users.

Grok 4’s “superhuman” intelligence serves as a cautionary tale, illustrating how the pursuit of data efficiency has far outpaced the development of frameworks for AI safety, transparency, and equitable access. As Musk predicts Grok will “discover new technologies by the end of next year,” the field faces a critical question in the race toward artificial general intelligence: can we ensure these systems align with human values before they redefine human knowledge?

The above is the detailed content of Grok 4 Accelerates AI Arms Race: Progress and Unresolved Perils. For more information, please follow other related articles on the PHP Chinese website!