▶ Article reader Listen + reading controls ⌄

LISTEN + READ YOUR WAY

Article reader

Preparing the reader…

▶ Play article ← → Stop

0:00 0:00

Speed 0.75× 1× 1.25× 1.5× 1.75×

Reading settings

Typeface Field Notes Atkinson Hyperlegible Lexend OpenDyslexic

Text size

A−

100%

A+

Reading surface Field Notes dark Paper light Warm sepia High contrast Text spacing Standard Comfortable

Reset reading preferences

The national laboratories are an unusually good test for AI

OpenAI's January 30 announcement makes its reasoning models available to scientists at Los Alamos, Lawrence Livermore, and Sandia national laboratories through a partnership using the Venado supercomputer.

The announcement covers an extraordinary range of possibilities: materials, energy, medicine, cybersecurity, mathematics, and nuclear security. Yet the laboratories are interesting not only because their problems are difficult. They are interesting because their knowledge practices are demanding.

A laboratory does not need an oracle

Scientists already work with models. They understand that a useful model can be wrong, that an elegant result can depend on a fragile assumption, and that evidence has to survive replication and challenge. An artificial intelligence (AI) system, including a large language model, should enter that culture as an instrument for forming and testing ideas—not as an authority that compresses uncertainty into fluent prose.

The most valuable applications may be upstream of a final scientific claim: searching a large technical literature, translating between specialties, suggesting hypotheses, writing exploratory code, finding inconsistencies, or proposing experiments that reduce uncertainty. Each can accelerate inquiry without eliminating the scientist's obligation to establish what is true.

That division of labor aligns with the idea of centaur or hybrid intelligence: people and machines contribute different strengths to a joint system. Dellermann and colleagues' hybrid-intelligence framework emphasizes configurations in which human and machine capabilities complement and improve one another. The design target is not machine substitution. It is a better knowledge-producing system.

Provenance is part of the result

In scientific and national-security work, a generated answer is not enough. Researchers need to know which sources, data, tools, and transformations support it. They need to distinguish a retrieved fact from a model inference and a hypothesis from an established result. They must be able to reproduce calculations and inspect code.

This makes provenance infrastructure central. Prompt histories alone will not suffice. An effective system should retain source identifiers, model and tool versions, data-access boundaries, intermediate artifacts, reviewer decisions, and the relationship between a suggestion and the evidence later gathered for it.

The partnership also highlights a tension between collaboration and compartmentation. Scientific insight often emerges by connecting fields. Security practice deliberately limits information flow. The system has to help researchers find lawful and useful connections without reconstructing or exposing sensitive context across boundaries.

A strong adoption pattern

National laboratories can model a disciplined path for other organizations:

1. Select questions where AI can expand exploration without becoming the final decision authority.
2. Pair domain experts with model and human-factors specialists.
3. Require traceable sources and reproducible tool use.
4. Test performance on local scientific work, not only general benchmarks.
5. Record negative results and failure modes in a shared evaluation library.
6. Separate research environments from operational or classified systems until evidence justifies movement.

The laboratories' advantage is not merely access to more compute. It is access to communities that know how to doubt a result productively.

If reasoning models can be integrated without weakening provenance, peer challenge, or security boundaries, the partnership will teach us something larger than whether AI can accelerate science. It will show whether organizations can absorb a powerful new instrument while preserving the practices that make knowledge reliable.

Sources and research trail

• OpenAI, “Strengthening America's AI Leadership with the U.S. National Laboratories” (January 30, 2025).
• Dellermann and colleagues, “Hybrid Intelligence” (2019).
• National Academies of Sciences, Engineering, and Medicine, Reproducibility and Replicability in Science (2019).
• Cohen and Levinthal, “Absorptive Capacity: A New Perspective on Learning and Innovation” (1990).