Japan's NIMS team experimentally confirmed that ultra-thin RuO2 films behave as altermagnets, a third class of magnetic order distinct from ferromagnets and antiferromagnets, with follow-up neutron-scattering work in early 2026 clarifying the spin structure. Altermagnets pair antiferromagnetic robustness against stray fields with ferromagnet-like fast electrical readout — exactly what's needed to break memory bandwidth and latency bottlenecks in large-scale AI training. Devices built from this material could enable ultra-fast, ultra-dense memory tailored to AI data centers.
Emory University physicists trained a custom neural network on 3D laser-tomography tracking of charged dust particles in plasma and uncovered previously unknown non-reciprocal force laws to >99% accuracy, published in PNAS. The AI revealed that classic theory tying particle charge to size is incomplete — charge actually depends on plasma density and temperature in ways no analytical model had captured. It is one of the rare cases of AI not predicting from data but inducing entirely new physical laws.
Mathematician Ernest Ryu used GPT-5 over three evenings (~12 hours) to resolve a four-decade-old question about why Nesterov's Accelerated Gradient method speeds up dramatically without losing stability — a phenomenon researchers had observed since 1983 but never fully explained. The model proposed creative routes drawn from adjacent fields, which Ryu pruned and verified to assemble a working proof. The episode is being cited as the clearest current evidence of LLMs functioning as research-level mathematical collaborators, not just retrieval engines.
USC Viterbi researchers showed an LLM with almost no exposure to the dependently-typed language Idris could go from 39% to 96% program correctness simply by being fed compiler error messages and asked to repair its own code — no extra training data required. The result suggests AI systems can push far past what training data alone would allow when an external 'oracle' (the compiler) reveals exactly what failed. It reframes how niche-domain capability gets unlocked: not more data, but tighter feedback.
Goodfire released Silico, claimed to be the first commercial mechanistic interpretability platform that lets developers inspect and adjust individual neuron activations within language models throughout training, not just after deployment. The tool automates interpretability work previously locked inside major AI labs, enabling smaller teams to diagnose and fix behaviors like hallucinations or ethical drift without bespoke in-house expertise. MIT Technology Review named mechanistic interpretability one of its 10 Breakthrough Technologies for 2026, and Silico is the first attempt to productize these techniques at scale.
Researchers at Okinawa Institute of Science and Technology trained AI agents to generate hidden internal mumbling signals, language-like self-talk invisible to users, while working through tasks, published in Neural Computation. This internal rehearsal mechanism, combined with enhanced working memory, let the AI adapt to new tasks, switch goals mid-stream, and multitask with far less data than standard training. The approach mimics how humans use inner speech to organize thought, and consumer-grade mumbling robotic agents are expected by late 2026.
Insilico Medicine AI identified an unknown protein driving idiopathic pulmonary fibrosis, then autonomously designed rentosertib to block it, potentially the first drug where AI handled both target discovery and molecular design end-to-end. Phase 2 clinical trials showed safety and effectiveness, with regulatory submission underway. This collapses what was previously a decade-long sequential pipeline into a unified AI-driven workflow.
Isomorphic Labs used AlphaFold variants to study proteins long considered undruggable because they lacked accessible binding pockets. Their AI identified small molecules that cause these stubborn proteins to physically change shape and open new therapeutic attachment points, revealing targets pharmaceutical researchers had written off entirely. The approach could unlock a new class of drug targets across cancer, neurodegeneration, and other diseases.
UC San Diego and the Allen Institute for AI developed Spherical DYffusion, a diffusion-based climate model that projects 100 years of global climate patterns in just 25 hours, 25x faster than conventional methods that require massive supercomputer clusters. The model uses a sphere-aware diffusion architecture trained on historical climate data while maintaining physical consistency across multi-decadal projections. It could democratize high-resolution climate modeling for institutions without supercomputer access.
UCL researchers combined a 20-qubit quantum computer with classical AI to predict chaotic spatiotemporal systems, achieving roughly 20 percent better accuracy and requiring hundreds of times less memory than standard approaches. The key insight is to use quantum processing just once offline to extract stable statistical patterns, then feed those patterns into classical ML training, sidestepping the noise that plagues repeated quantum-classical data exchanges. Published in Science Advances, this is among the first demonstrations of practical quantum advantage integrated directly into a machine learning pipeline.
When physicist Alex Lupsasca prompted ChatGPT with a warm-up black hole problem, the AI independently derived the same event-horizon symmetries he had discovered but via a previously unpublished mathematical route. The alternate derivation was novel enough to be incorporated into the final paper, making ChatGPT an effective co-contributor to original theoretical physics. The episode challenges the assumption that LLMs merely recombine training data rather than exhibiting genuine mathematical creativity.
Chalmers University physicists built a quantum refrigerator published in Nature Communications that uses quantum noise, the very disruption that normally destroys qubits, as the driving force for cooling. The device realizes long-theorized Brownian refrigeration, where carefully steered random thermal fluctuations produce a net cooling effect at quantum scales. The same device can also act as a heat engine or energy amplifier, turning quantum noise from a liability into a versatile resource.