📈 Monday data: For science the bottleneck isn’t intelligence
AI already improves AI. Whether that compounds across the sciences depends on the field.
In recent research, AI systems have demonstrated the ability to improve upon themselves, challenging the notion that intelligence is the primary bottleneck in scientific progress. This self-improvement, known as recursive self-improvement (RSI), has been observed in various fields, including coding and mathematical research.
Anthropic co-founder Jack Clark recently stated that the singularity, a point at which AI systems drastically accelerate scientific and technical progress, may have already begun. This idea is not new, as mathematician Irving John Good proposed decades ago that a machine smarter than humans could lead to an "intelligence explosion" and surpass human intelligence.
Recent studies have shown that AI agents can iteratively improve their own code, resulting in significant gains on coding benchmarks. For instance, researchers at Sakana AI and the University of British Columbia found that their agents achieved improvements of 30 pp on SWE-bench and more than 15 pp on Polyglot, testing coding skills beyond Python.
Moreover, an AI-built agent at Google reportedly proved as good as a human-built one in improving other AI agents, though the result was inconclusive due to greater differences between runs. However, the chatter around RSI in AI continues to grow, with major tech companies like Google acknowledging the potential of this technology.
While some experts argue that the loop of AI self-improvement is closing in on other scientific fields, others remain skeptical. Nonetheless, the evidence presented in these studies suggests that the bottleneck in science may no longer be human intelligence, but rather the challenges in creating and managing AI systems that can improve themselves.
Written by urgent.news from Exponential View's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
