A genuine AI capabilities explosion won't happen just because models can write novel research papers. The bottleneck is the full automation of the R&D loop, which includes a long tail of "messy" real-world tasks like fixing failing GPUs in a data center or managing facility cooling. This physical and logistical grounding is often overlooked.

As AI, like Anthropic's Claude, generates vast amounts of code, the primary constraint on development speed shifts. The bottleneck is no longer code creation but the capacity of human engineers to review, validate, and integrate it. This is a real-world example of Amdahl's Law applied to organizational workflows.

A slowdown in compute growth may have a squared negative effect on AI progress. It not only reduces resources for training larger models but also stifles the discovery of new algorithms, as breakthroughs like the Transformer required immense compute for experimentation. This double impact could significantly delay major capabilities milestones.

Previously, implementing a new algorithm could take weeks, leaving compute idle. With advanced coding assistants, ideas can be prototyped in hours, making the availability of compute resources to run experiments the primary limiting factor for progress again.

A "software-only singularity," where AI recursively improves itself, is unlikely. Progress is fundamentally tied to large-scale, costly physical experiments (i.e., compute). The massive spending on experimental compute over pure researcher salaries indicates that physical experimentation, not just algorithms, remains the primary driver of breakthroughs.

Even if AI perfects software engineering, automating AI R&D will be limited by non-coding tasks, as AI companies aren't just software engineers. Furthermore, AI assistance might only be enough to maintain the current rate of progress as 'low-hanging fruit' disappears, rather than accelerate it.

The AI's ability to learn and scale is constrained by the slow feedback loops of the human economy. Because it takes time for a human client to pay for a job, the AI's reward signal is delayed. This human latency, not computational power, is the primary blocker to faster learning and optimization.

Braintrust's CEO argues that developer productivity is already 'tapped out.' Even if AI models become 5% better at writing code, it won't dramatically increase output because the true bottleneck is the human capacity to manage, test, deploy, and respond to user feedback—not the speed of code generation itself.

The shift from simple query-based AI to agentic AI, where AI calls itself recursively to solve complex tasks, increases compute demand by orders of magnitude. Most people, especially non-coders, fail to grasp this exponential shift, leading them to consistently underestimate the scale and duration of the AI infrastructure build-out.