2027: Humanity's Final Choice—To Allow AI to Self-Evolve?

The development of artificial intelligence is approaching a historic tipping point. Jared Kaplan, Chief Scientist at Anthropic, has issued a stark warning: humanity will face an "extremely high-stakes decision" between 2027 and 2030—whether to allow AI systems to autonomously research and train the next generation of AI.

Kaplan explains that the current paradigm of training AI on human data is hitting its limits. High-quality textual data on the internet is nearing exhaustion, and the returns from merely scaling up model size are diminishing. The next phase, therefore, inevitably points towards "recursive self-improvement," where AI uses its own generated data and self-play to create more powerful successors.

Once this process begins, it could trigger an "intelligence explosion." AI could rapidly surpass human intelligence, and its evolutionary path and goals may become incomprehensible and uncontrollable to humans. This is the core of Kaplan's concern: we risk creating an entity that is not only intellectually superior but also completely inscrutable in its reasoning.

This future warning coincides with profound changes happening today. An internal Anthropic report reveals that in professional development settings, AI can now handle nearly 80% of coding tasks. While this boosts productivity, it also raises alarms about "skill atrophy" among engineers and the collapse of the traditional apprenticeship model, as junior engineers lose crucial hands-on learning opportunities.

2027 is identified as a critical year because it marks the anticipated deployment of next-generation, ultra-large-scale computing clusters, providing the necessary hardware foundation for AI self-evolution. Humanity stands at a crossroads: to press the accelerator or attempt to install safeguards. https://www.anthropic.com/research/ho...
https://www.theguardian.com/technolog...