Advanced models are held by 5,000 lab insiders while broader science waits

Ethan Mollick highlighted a major imbalance in how state-of-the-art artificial intelligence is deployed today. Roughly 5,000 people across elite AI laboratories hold access to advanced frontier models. Almost all of their focus remains locked into three narrow domains: cybersecurity, pure mathematics, and accelerating AI research itself.

Frontier intelligence primarily serves its own development

The primary signal emerged from discussion on X, where Mollick outlined the current reality inside leading AI organizations. While researchers and engineers hold models significantly more capable than public releases, daily deployment remains trapped in a narrow corridor. Outside this perimeter sit dozens of scientific fields where these systems could solve pressing problems, yet domain specialists rarely get access.

Focusing compute on internal AI R&D makes immediate economic sense for lab operators. When a frontier model optimizes training architectures or debugs infrastructure, the efficiency directly benefits the next training run. However, this closed loop creates a false impression that model capabilities are stalling, when in truth they have only been tested against a tiny slice of human inquiry.

The evaluation bubble reinforces narrow applications

Laboratories naturally optimize benchmarks for challenges they understand and can grade automatically. A mathematical proof or a code vulnerability yields an unambiguous programmatic outcome. Experimental biology, materials discovery, and field sciences instead require laborious empirical validation and deep domain context that fast-paced software labs rarely accommodate.

This selection bias skews our collective understanding of frontier model utility. When 5,000 software engineers evaluate models strictly through the lens of their own daily workflow, the resulting systems specialize in replicating their creators rather than unlocking scientific breakthroughs.

Unrestricted lab access cannot replace domain expertise

A model solving difficult formal equations does not automatically accelerate therapeutic discovery. Lab leadership frequently assumes that raw reasoning power will effortlessly spill into outside disciplines the moment an API endpoint opens. In practice, a massive methodological gulf separates raw token generation from domain scientific rigor.

Even 5,000 privileged insiders cannot substitute for thousands of dedicated domain specialists. As long as advanced systems are handled solely by the people building them, evaluations merely reflect familiar assumptions and shared blind spots. Rigorous security audits are vital, but they do not constitute a broader scientific revolution.

Real progress hinges on opening access to domain specialists

Meaningful acceleration will not come from shaving seconds off internal lab scripts, but from placing advanced models directly into the hands of external researchers. Whether frontier labs can deliver interfaces suited to biochemists, epidemiologists, and material scientists will determine their lasting impact.

A key metric to track over coming quarters is the ratio of published interdisciplinary findings compared to internal safety and capability whitepapers. If advanced reasoning systems remain confined to engineering the next generation of models, their promised scientific transformation will remain largely theoretical.

Lilith's verdict

Laboratories kept their sharpest models locked in a hothouse, teaching machines to solve internal homework. A genuine breakthrough will only arrive when this machinery is pointed at problems for which Silicon Valley has no tidy spreadsheet.