Vals AI finds magnetic semiconductors with Claude agents

Vals AI published details on 5 October.

The company said a team of Claude Opus 5.5 agents identified two new magnetic semiconductor candidates. One is a novel compound designed by the AI. The other is a material created in 1999 that was previously unrecognized for these specific properties. Both are predicted to have zero net magnetism while sorting electrons by spin. This property is essential for next-generation computer memory, specifically for spintronics applications that store information based on electron spin orientation rather than magnetic polarity alone.

How antiferromagnets differ from fridge magnets

Common ferromagnets, like the magnet on a refrigerator door, have atomic magnets that all point in the same direction. This creates a macroscopic magnetic field that leaks out from the surface. While useful for holding notes, this external field is a liability in dense computer memory. It interferes with nearby materials and is difficult to control precisely. Switching ferromagnetic states is also relatively slow and power-hungry.

Antiferromagnets work differently. In these materials, neighboring atomic magnets point in opposite directions, effectively canceling out the external magnetic field. This lack of a macroscopic field allows engineers to pack storage devices much closer together, potentially enabling higher performance. However, ordinary antiferromagnets cannot distinguish between electrons with spin up and spin down. This makes it hard to read or store information using standard spintronics techniques. The candidates found by the Vals AI team sit in a unique middle ground: they offer the packing density and speed of antiferromagnets but retain the ability to sort electrons by spin.

Vals AI shared the full calculations, code, and a list of known caveats for both candidates. The company noted that antiferromagnetic materials are approximately a thousand times faster than their ferromagnetic counterparts for certain switching tasks. This speed advantage, combined with the ability to operate at room temperature, could significantly impact the architecture of future memory devices.

AI in the wet lab

While Vals AI focuses on inorganic materials, other AI systems are making waves in biological sciences. Recent industry news highlights a trend of AI-directed protein engineering. A cloud lab dedicated to AI-directed protein engineering recently received $20 million from the National Science Foundation. This funding aims to accelerate the design of new proteins using computational models. The connection between material science and biology is becoming clearer as AI models learn to predict physical properties across different domains.

In August 2026, Discovered Materials raised $9 million in seed funding to accelerate AI-powered semiconductor materials discovery. The company uses similar agentic approaches to hunt for new chip materials. This funding round follows a seed investment in July 2026, indicating sustained investor interest in AI-driven discovery pipelines. The parallel between Vals AI's recent release and Discovered Materials' fundraising suggests a maturing market for AI tools that can replace or augment traditional trial-and-error experimentation.

DeepMind has also entered this space with new watermarking technologies for proteins. In early October 2026, reports indicated that DeepMind watermarks proteins to enable lab verification. This move addresses a growing concern in AI-generated science: how do we verify that a predicted structure is real and not an artifact of the model? By embedding watermarks in protein designs, researchers can trace the origin of a prediction back to a specific AI model. This is a critical step for making AI-generated biological data trustworthy in regulatory and pharmaceutical contexts.

From prompting to autonomous discovery

The Vals AI project relied on Claude Opus 5.5 agents, a model known for its agentic capabilities. This means the AI could plan, execute, and verify steps in a workflow without constant human intervention. The company emphasized that the agents were tasked with designing and finding materials, not just retrieving existing data. This distinction is important. Retrieval systems find what is already known. Agentic discovery systems can propose novel hypotheses and test them computationally.

Another example of this agentic approach is the "Gauntlet Loop" prompting method described by a developer in October 2026. This method involves giving a lead agent a goal and a high-quality reference example. The agent then breaks the work into small pieces, assigning each to a builder and a critic. The critic compares the output against the reference, and if it falls short, it sends the work back with specific instructions for improvement. This loop continues until the result meets the high bar. While originally demonstrated for building a Call of Duty-style game in Three.js, the developer argues the method applies to research, code, and design. It is a practical example of how structured prompting can enhance AI output quality in complex tasks.

However, not all AI-driven discoveries are without controversy. A study led by UC Riverside computer scientists, published in September 2026, found that AI models can give wrong answers with high confidence. The researchers identified internal features in large language models that are associated with confidence and correctness separately. This challenges the assumption that a confident model is a correct model. For AI-driven scientific discovery, this is a significant risk. If an AI agent confidently predicts a material property that does not exist in the physical world, the error could propagate through the research pipeline. The UCR team suggested that adjusting these internal features could make models more cautious when they are likely to be wrong, a step toward more reliable AI scientists.

The field is also grappling with the issue of "AI slop" in science. As more papers and predictions are generated by AI, distinguishing between novel insights and hallucinations becomes harder. The Vals AI team addressed this by publishing their full code and calculations, allowing independent verification. This transparency is becoming a standard expectation for AI-generated scientific claims. It is no longer enough to state a result; the underlying logic and data must be inspectable.

For now, the discovery of two room-temperature antiferromagnetic semiconductor candidates is a step toward practical applications. These materials could lead to faster, denser, and more energy-efficient memory. But the broader implication is that AI agents are becoming capable partners in the discovery process. They can navigate vast chemical spaces, identify overlooked materials, and propose new compounds. The challenge now is to integrate these tools into standard scientific workflows while maintaining the rigor and verification that define good science.