check out this podcast on #chips and #AI with chipstockinvestor and dr shalini sharma. the next gen of chips will not be designed the way the last ones were. youtu.be/EZIHlk-EdqA?is=k1gl…

AI agents can be very beneficial, but they can also wreak havoc with cybersecurity and token budgets.   Security teams are having difficulty auditing what agents are doing. CFOs cannot tie agent spend to clear outcomes. When an unmonitored agent leaks customer data, executes unverified code, or racks up unbudgeted compute costs across thousands of API calls, it becomes a major liability. That is why we expanded the SandboxAQ Switch platform with Switch Trust and Switch ROI. With Switch Trust, enterprise teams can discover every agent running across their environment, run adversarial tests against prompt injection and data leaks before deployment, and enforce real-time guardrails in production With Switch ROI, organizations can tie agent compute spending directly to measurable business outcomes rather than counting raw tokens, allowing teams to dynamically optimize model performance without sacrificing security. Whether deploying medical software tools in clinical settings or streamlining energy logistics, putting autonomous AI agents to work requires continuous governance and clear accountability. Switch is open source and free to use. Check it out here: switchagents.ai/

The biggest breakthrough in biopharma AI will be eliminating weak drug candidates before companies spend years developing them. In 2027 leading biopharma companies will begin using quantitative models to make earlier decisions about which molecules deserve investment. The result will be a fundamental shift in how the industry measures R&D productivity.

The physical world has zero margin for error. When a patient is rolled into the ER with chest pain, a surgeon needs diagnostic certainty within minutes. When an energy company spends $1 billion deploying battery storage, that chemistry must be flawless. For an AI output to be useful in the real world, it must be driven by reality. Large Language Models (LLMs) are incredible conversational interfaces, but the physical world requires Large Quantitative Models (LQMs) running the physics behind the scenes. LQMs run high-scale simulations constrained by real-world variables such as shelf stability, and manufacturability. Without those physical constraints, we risk sending engineers down a year-long R&D path that works in theory but fails in a factory. This is why we partnered with @AnthropicAI to bring Claude together with @SandboxAQ's LQMs: Combining intuitive conversational AI with science-grounded physics.

We had a productive time at @TechCrunch Disrupt this week with @StacieCT of NVIDIA. Our conversation centered on where AI goes beyond Large Language Models. Language models serve as a great interface. Now they can link to domain-specific models to do the science. Two examples are Claude linking to SandboxAQ's Large Quantitative Models and NVIDIA’s BioNeMo model. SandboxAQ and NVIDIA have collaborated on six papers across biopharma and material science that drive new quantitative models to the frontier! Thank you to the TechCrunch team for hosting us yesterday.

The physical world is not on the web. That's why we’re building a different kind of AI model at @SandboxAQ : Large Quantitative Models. Language models learn from the web. The words, the images, the YouTube videos… They're great resources, and they're why models like ChatGPT, Claude and Gemini are so good with language. But if you want to predict which compound will work as a drug, or how a material will behave under load, you need measurements from biology, chemistry and physics. Most of that data to train these new models doesn't exist in usable form. So we build our own data sets. We generate the measurements and validate them, sometimes over several years. Then we use these proprietary data sets to train quantitative models. These models go beyond the transformer architecture and incorporate new methods. A language model predicts the next word. A quantitative model predicts a physical property. How a molecule binds. How a novel material can make a vehicle stronger and lighter. How a battery chemistry can leapfrog current storage limits. Customers use our models to test thousands of ideas on a computer before running a single lab experiment. Then they test prototypes in the lab and feed that data back into the quantitative models to optimize the result. Over the next few months we will detail what that looks like in drug discovery, materials, energy and other fields.

Very proud of the AQVolt team, who won Battery Software and Digital Innovation of the Year this week! AQVolt is the SandboxAQ LQM that simulates battery materials down to the atomic level, using models trained on the equations of physics and chemistry. That lets the team screen new chemistries before anyone builds a cell. AQVolt is trained on our dataset of 322,656 r2 SCAN single-point calculations for lithium halides, generated via high-temperature configurational sampling across ∼5K structures.

check out Switch - our new tool to collaborate with agents right in the software you use every day such as Slack, Teams, Discord, etc it’s open source and free - give it a try

check out Switch - it’s #1 on product hunt today - add Switch to your collab tools such as Slack, Teams, Discord etc. it’s free

Tribute to Nobel Prize winner Ada Yonath. Significant contribution to #biochemistry — specifically the structure and function of the ribosome . Despite a tough early life she dedicated herself to science and we are all the beneficiaries of her work.

check out this new diagnostic med tech - could replace endoscopies over time

Today’s White House Executive Order 14409 marks a new era in America’s cyber defense posture:  The transition to post-quantum cryptography is now an imperative The standards exist. The challenge now is implementation.  Governments, financial institutions, energy providers, telecommunications networks, and other critical infrastructure operators must inventory and upgrade cryptographic systems before quantum computers are capable of breaking today’s encryption. At SandboxAQ we are already working with federal agencies to implement the NIST standards and now with this EO we can accelerate the work.   aqtiveguard.com/blogs/q-day-…

US Dept of Commerce Awards $500 million to @SandboxAQ for key technologies for semiconductor manufacturing. This program will focus on scaling semi production in the US through breakthroughs in: PFAS mitigation, catalysts, magnets and batteries. Let’s go! nist.gov/news-events/news/20…

Batteries are essential to the future of energy storage, AI infrastructure, transportation, and national security, but current battery chemistries depend on critical materials with supply chain chokepoints and geopolitical bottlenecks.  AQVolt26 is the latest frontier model from @SandboxAQ which combines high-fidelity Density Functional Theory (DFT) simulations with machine learning, to accelerate the discovery of novel battery chemistries at scale. Built on more than 320,000 quantum mechanical simulations and over 200 million atomic configurations, AQVolt26 enables robust atomistic modeling of solid-state battery materials under realistic operating conditions, including high-temperature and highly distorted environments where many existing models fail. To learn more about this crucial step forward for grid tech and defense, see our new paper here: arxiv.org/abs/2604.02524

Proud of the team for their new paper in the peer-reviewed journal Nature Computational Materials. The @SandboxAQ team presents a breakthrough in catalyst discovery and computational chemistry. Catalysis drives the global economy, from the fuels that power our world to the materials that shape it. With our AQCat model, industries can now simulate, screen, and optimize catalysts with physics-based accuracy, unlocking performance and sustainability breakthroughs at unprecedented scale. The team built a dataset containing 13.5 million Density Functional Theory (DFT) calculations and found that training a model on specialized chemistry improved performance on the target domain but often caused forgetting of knowledge learned elsewhere.   To address this challenge, the team trained on multiple datasets and provided the model with additional information about the underlying physics.  Using this approach, the team improved performance without sacrificing generalization. As AI moves deeper into chemistry, materials science, energy systems, and drug discovery, success will depend on more than larger models and larger datasets. It will depend on how effectively we incorporate scientific knowledge while preserving the ability to generalize across domains. Check out the paper: pub.sandboxaq.com/publicatio…