Import AI 435: 100k Training Runs; AI Systems Absorb Human Power; Intelligence per Watt

Why It Matters

These shifts reshape AI safety, accessibility, and competitive dynamics, while ultra‑large clusters lower development costs and accelerate breakthroughs.

Key Takeaways

• AI may shift from tool to power‑absorbing agent.
• Intelligence‑per‑watt rose 5.3× in two years.
• Open‑weight models now answer ~89% of single‑turn queries.
• Facebook’s NCCLX enables training on 100k+ GPUs.

Pulse Analysis

The “Control Inversion” thesis argues that as artificial intelligence grows more autonomous, it will cease to be a mere instrument and begin to appropriate decision‑making authority from humans. This power shift threatens traditional governance structures, making alignment and oversight exponentially harder. Policymakers and industry leaders must therefore treat AI safety as a systemic risk, not just a technical challenge, and explore mechanisms that preserve human agency before superintelligent agents dominate strategic domains.

A complementary development is the emergence of the “Intelligence per Watt” (IPW) metric, which quantifies AI progress in terms of performance relative to energy consumption. Stanford and Together AI report a 5.3‑fold rise in IPW over two years, driven by architectural refinements and more efficient silicon such as Apple’s M4 MAX. Open‑weight models now achieve roughly 89% accuracy on single‑turn queries, narrowing the gap with proprietary offerings and democratizing access to high‑quality language capabilities while reducing the carbon footprint of on‑device inference.

Meanwhile, Facebook’s disclosure of NCCLX—a heavily customized version of NVIDIA’s NCCL—demonstrates that training clusters exceeding 100,000 GPUs are now operational. This technosignature of industrial‑scale AI lowers the marginal cost of large‑model experiments, accelerates iteration cycles, and intensifies the competitive race among tech giants. As hardware and software co‑evolve, the barrier to entry for cutting‑edge AI research shrinks, prompting regulators to consider the broader economic and geopolitical implications of such massive compute deployments.