Big Tech Calls to Slow AI Pace... Leeway Created by US-China Technology Gap [Click-e Industry]

Gap Widens from 7 Weeks to 3 Months After Evaluation Change

U.S. Focuses on Securing Time for Safety Research

Monetizing Technological Superiority Now Key

There are emerging calls within segments of the American big tech sector—which has led the race in artificial intelligence (AI) development—to slow down the pace of innovation. The aim is to allow safeguards and safety measures the time they need to keep up with technological advancement. Analysts say this proposal is possible, despite China’s pursuit, because the U.S. is considered to have a more pronounced technological superiority than many assume. In terms of actual work performance capabilities, the gap between the United States and China remains significant.

Dario Amodei, CEO of Anthropic, published an article this month suggesting that the speed at which AI models improve should be controlled, and that external evaluation panels be instituted for ongoing monitoring. Earlier in July, Demis Hassabis, CEO of Google DeepMind, also proposed a safety evaluation system and stated that leading companies could adjust their development speed if needed.

The reason behind these proposals is the technological gap with China. CEO Amodei explained that the extent to which American companies can afford to slow down depends on their technological lead over China. To maintain that gap, he also proposed measures such as restricting exports of advanced AI semiconductors and blocking technology leaks. The idea is to maintain their lead position while securing time for safety research.

Il Hyuck Kim, an analyst at KB Securities, highlighted in a report the recent revision of the evaluation index by the AI evaluation organization ArtificialAnalysis. He noted that while the previous evaluation method indicated a U.S.-China AI model gap of about 7 weeks, the change in evaluation methodology increased the gap to about 3 months. This suggests that the difference did not suddenly widen in just a few days; rather, the shift in yardstick made distinctions appear that were not previously apparent in the scores.

Earlier this month, ArtificialAnalysis revised its index to place greater emphasis on complex knowledge work and work automation assessments. The newly introduced “AA-Briefcase” evaluates tasks involved in multi-week projects. The assessment not only considers short-answer responses but also examines the ability to deliver requested outputs using available data.

Mechanisms to filter out models that only perform well on public exams have also been strengthened. In index v4.3, released on September 7, the weighting for evaluations using non-public assignments or answers was raised to 45%. The intention is to reduce the likelihood that optimization for test questions will exaggerate practical performance capabilities. KB Securities explained that these changes are what caused the U.S.-China gap to appear wider.

ArtificialAnalysis compared not only the cost per model use, but also the actual costs incurred in completing evaluation tasks. For enterprises considering AI adoption, it’s essential to calculate the total time and expense required to complete a task. Even a low-cost model may fall short of expectations if it requires repetitive processing and substantial human edits.

Ultimately, if companies maintain superiority in practical capabilities, they gain flexibility to adjust intervals between model releases and allocate more time for validation. The ability to carry out work safely becomes, itself, a competitive advantage in winning over corporate customers.

Analyst Kim stated, “As success rates, work time, and safety evaluations become increasingly rigorous, the gap between the U.S. and China could widen further,” adding, “America’s AI supremacy acts as a factor that prevents instability in U.S. assets, including the dollar.”

This content was produced with the assistance of AI translation services.