AI Is Now Helping Build the Next AI, and Progress Has Nearly Doubled in Speed

Why AI has gotten smarter has been fairly simple until now: put in more chips, train on more data, build bigger models.

A new variable has appeared. AI that already exists has begun to raise the productivity of the researchers building the next generation of AI.

On Nov. 5, METR, an independent AI evaluation group, published its analysis of Anthropic's internal research materials. Anthropic's AI models were improving on key performance benchmarks about 1.9 times faster than before. That means the past year of gains matches roughly two years of earlier progress.

More striking, the acceleration is not explained only by faster growth in training compute. AI helping to develop AI is changing the pace of research and development itself.

What actually got close to twice as fast

METR spent about 11 weeks, from Aug. 12 to Oct. 29, 2026, reviewing Anthropic's internal model performance, training compute and R&D materials, and interviewing researchers.

The measurement does not cover all of AI researchers' working time. It covers the rate of improvement in AECI, Anthropic's composite model performance index.

Since October 2025, the index has risen at an estimated 1.9 times the earlier trend. The result mainly concerns tasks that are clearly measurable. It does not mean progress in all intellectual work has doubled.

Key chart: pace of AI improvement, then and now

This is a relative comparison with the earlier pace set at 1. It does not mean model accuracy rose 90%. It means gains are arriving faster.

Faster experiments, not more GPUs

Traditional AI development repeated a cycle: researchers form a hypothesis, write code, run an experiment and analyze the results.

AI now assists with much of the coding, experiment setup, data analysis and troubleshooting. One researcher can try more things in the same amount of time.

METR found internal evidence that does not support the explanation that faster growth in training compute drove the acceleration. It also suggested that a substantial part of the current acceleration may come from the use of AI. It could not pin down the specific contribution to research productivity.

• A stronger AI model arrives
• Researchers become more productive at coding and experiments
• More research and more model improvements follow
• The next generation of AI is developed faster

This is the path by which recursive self-improvement could begin in practice. For now, human researchers lead the core research, and there is little evidence that AI performs the major development steps fully on its own.

What will AI competitiveness depend on?

OpenAI, Google DeepMind and Meta have the same incentive to chase these productivity gains. But using the same AI tools does not produce the same research results.

Differences in development speed can emerge where strong research talent, proprietary experiment environments, accumulated data and evaluation systems, and an organization's ability to improve and ship models quickly come together.

Conversely, if AI development tools spread widely, barriers to entry for latecomers could fall. These results alone cannot predict a long-term lead for Anthropic or a rise in any one company's market share.

Good or bad for Nvidia?

If AI shortens research time, companies can run more experiments in the same period. In that case, GPU usage and demand for data center investment could rise.

But gains in algorithms and experimental efficiency could also reduce the compute needed to reach the same performance. Whether falling unit costs from efficiency or growth in total usage proves the bigger effect has not yet been settled.

For Nvidia, then, what matters is less the research acceleration itself than actual customer GPU purchases, cloud usage, gross margin and return on investment. For AI developers, the signals to watch are inference costs, paying customer counts and revenue per customer, alongside the pace of model improvement.

What to watch

Insight Times Editorial Desk