On June 4, Anthropic published a pivotal blog post titled 'When AI Builds Itself,' co-authored by co-founder Jack Clark and head of internal research Marina Favaro, revealing previously undisclosed operational data that challenges the current trajectory of artificial intelligence. Valued at nearly $1T and preparing for an initial public offering, the company issued a rare public plea for a global pause on cutting-edge AI development. This strategic pivot aims to allow social structures and alignment research to synchronize with rapid technological acceleration, a move that has sparked intense debate between critics viewing it as a regulatory maneuver and supporters seeing it as a genuine safety warning. The core of this argument rests on the concept of recursive self-improvement, which Anthropic assesses could materialize within 2 years, fundamentally altering the dynamics of AI creation.

Data compiled by Woofun AI indicates that the acceleration of AI development is not merely theoretical but empirically observable within Anthropic's own workflows. As of May 2026, over 80% of the code merged into the company's codebase was authored by its AI models, a stark contrast to the single-digit percentages recorded before the February 2025 release of Claude Code.

This shift has driven a dramatic increase in engineering productivity; by the second quarter of 2026, the typical engineer merged 8 times more code per day compared to 2024 levels. In an internal survey conducted in March 2026 covering 130 research employees, the median estimate suggested that output using the latest Mythos Preview model was approximately 4 times higher than when working without AI tools, signaling a compounding effect on R&D velocity.

The technical capabilities driving this acceleration are evident in both engineering and research benchmarks. In software engineering, the SWE-bench, which tests models on real-world open-source bug fixes, has seen scores evolve from single digits to near saturation over a 2-year period. Similarly, the CORE-Bench, which evaluates a model's ability to reproduce existing research results, saw success rates climb from 20% in 2024 to near saturation within 15 months. Woofun AI notes that the Mythos Preview model demonstrated the ability to work continuously for over 16 hours, identifying more than 10,000 high-risk software vulnerabilities in critical global systems during initial tests. This capability has already shifted the bottleneck in network defense from vulnerability discovery to the rapid patching of identified flaws.

In the realm of experimental research, the divergence between human and AI performance is widening rapidly. When tasked with optimizing code execution speed, Claude Opus 4 achieved a 3x speedup in May 2025, whereas by April 2026, Mythos Preview reached a 52x speedup, far surpassing the 4x improvement a skilled human researcher could achieve in 4 to 8 hours.

Furthermore, in open-ended research scenarios where agents were tasked with solving AI safety problems, autonomous agents bridged 97% of the performance gap in roughly 800 hours of computation time, a task that took two human researchers a week to bridge only 23% of. These metrics suggest that the cost of human hours at the execution level is approaching zero, leaving only the strategic direction-setting as a human domain.

The implications of these trends point toward three potential futures, with the most concerning being a scenario where AI systems autonomously design and improve their successors. Woofun AI analysis suggests that if the current trajectory holds, the bottleneck in AI development will shift from execution to human review and strategic judgment. Currently, human comparative advantage lies in grasping the big picture and determining which problems are worth solving, but as AI models like Mythos Preview demonstrate the ability to propose and execute experimental plans with increasing autonomy, this gap is narrowing. The code written by Claude, which was slightly below human quality by the end of 2025, is now roughly on par, with expectations of full superiority within the next year.

Anthropic argues that a unilateral pause by a single lab would be ineffective, merely shifting the competitive lead rather than addressing the systemic risk. Instead, the company advocates for a verifiable, coordinated global slowdown, acknowledging the immense difficulty of verifying compliance in an environment where training runs are easier to conceal than missile silos. The incentive to defect is high, as entities that continue advancing while others pause could inherit a decisive technological lead. To address this, Anthropic proposes the development of institutional frameworks that enable leading developers to verify if other global entities have genuinely halted or slowed development, ensuring that no malicious actors advance covertly.

The urgency of this situation is underscored by the potential for recursive self-improvement to outpace human oversight mechanisms. If AI systems can fully build their own successors, the methods for protecting, monitoring, and shaping their behavior will become critically important. The company warns that without a global coordination mechanism, national governments and corporations will be forced to make difficult security decisions under intense geopolitical pressure, potentially leading to a less secure outcome for all. The window for establishing such a framework is narrow, as the pace of AI advancement continues to accelerate, with the cycle for models to reliably complete tasks independently shortening from 7 months to approximately 4 months.

Ultimately, the decision to pause or slow down hinges on the ability to balance technological progress with societal readiness. While the benefits of AI in science and healthcare are immense, the risk of losing control over autonomous systems that can self-improve poses an existential threat. Anthropic's call to action is not merely a request for a delay but a strategic imperative to build the necessary infrastructure for a trusted deceleration mechanism. As the industry stands on the precipice of a new era where AI may soon design its own evolution, the choices made in the coming months will define the trajectory of human-AI interaction for decades to come.