Ⓙⓐⓒⓚ Ⓗⓘⓓⓐⓡⓨ

Less than 2% of U.S. grid power can be stored. Across most of the world, that number is even lower. If we are going to transform the grid and meet surging power demand, we need scalable battery storage that can operate safely at industrial scale. Solid lithium halides could replace the flammable liquid electrolytes inside today's batteries, but finding the right material requires knowing how quickly ions travel through the crystal lattice. Measuring that on a computer means running molecular dynamics simulations above 1,000 Kelvin to speed up rare ion hops. At those temperatures, halide lattices become dynamically soft, distorting far from equilibrium. While foundational models handle stable crystals well, their energy predictions break down in these distorted states. In collaboration with NVIDIA, we built AQVolt26: 322,656 r²SCAN density functional theory calculations, selectively sampled from more than 200 million high-temperature lithium halide configurations. This is a Large Quantitative Model (LQM) that has peer-reviewed published benchmarks and outperforms other models. On strained halide structures, energy error drops to 15 to 24 meV/atom when co-training the eSEN architecture on AQVolt26, compared to 52 to 85 meV/atom without it. The models also become more conservative on ionic conductivity, cutting down on false positives before a single material is synthesized in the physical lab. That is quantitative AI at work. Train on the physics of the problem, and the LQM holds up where discovery actually happens.

A list of ingredients won't tell you how to bake a cake. You need the complete recipe to understand how to combine all the ingredients, in what order, at what temperature etc. Industrial catalysis and chemical discovery work in a similar way. Catalysts are fundamental to more than 80% of the manufacturing and industrial world. They speed up reactions, reduce the energy needed, and then remain to be recharged and reused for the next reaction cycle. To develop a new catalyst, it is not enough just to specify a bunch of atoms that comprise the catalyst molecule. We have to account for magnetic spin, electronic fidelity, surface restructuring, and more parameters, all of which have an impact on the ability of a catalyst to do its job. Each of these variables shifts binding energies and activation barriers, ultimately deciding whether a catalyst performs efficiently or fails under stress. This matters most for earth-abundant metals like iron, cobalt, and nickel, which drive massive industrial processes like ammonia and Fischer-Tropsch synthesis. Their strong spin polarization is exactly what most large-scale catalysis datasets leave out, simply because spin-polarized density functional theory (DFT) calculations are too computationally expensive. Our core team at @SandboxAQ addressed that gap with AQCat25, introducing a dataset of 13.5 million DFT single-point calculations alongside spin-aware machine learning interatomic potentials. By explicitly training the models on magnetic spin and computational fidelity, we improved accuracy on these complex magnetic systems without sacrificing generalizability.

We are now entering the reality of AI fighting AI. At the time of the @OpenAI breach at Hugging Face in July, we warned that AI cyber attacks were only going to get worse. That prediction is playing out across the enterprise the past few weeks. Security researchers just discovered threat actors deploying open-source autonomous agents to relentlessly probe and breach more than 1,200 corporate networks, impacting major manufacturers like @JohnDeere and @SharkNinja @QuestDX also confirmed a cybersecurity breach that compromised sensitive patient records. Employees are granting unsanctioned AI tools unchecked access to corporate environments, and a major AI aggregator breach this week exposed terabytes of confidential enterprise data uploaded outside of IT oversight. Organizations must move past legacy perimeter security and establish military-grade air gapping alongside dynamic, model-driven defenses. We cannot just wait for the next incident to happen. Recognizing that containment has broken down is the first step toward building true resilience.