Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Sakana AI researchers just developed CycleQD, an evolutionary framework that merges specialized small models into multi-skilled AI swarms without expensive fine-tuning. By using model merging and SVD for crossover and mutation, it outpaced traditional fine-tuning on Llama 3-8B tests, offering a cheaper path to balanced AI agents.
Researchers just introduced BioEmu, a generative AI system that samples thousands of independent protein conformations per hour on a single GPU. Trained on molecular dynamics and stability data, it predicts free energy with errors near 1 kcal/mol. Basically, it brings months of supercomputing simulation down to a few cheap GPU hours.
Google DeepMind just published GenCast, an open-source diffusion ensemble model for 15-day global weather forecasts. Trained on ECMWF ERA5 data, it beat top operational system ENS on 97.2% of targets while running in just 8 minutes on a single TPU. It marks a huge leap for extreme weather and renewable energy forecasting.
Google DeepMind just introduced Genie 2, a foundation world model that turns a single prompt image into playable, action-controllable 3D environments using standard keyboard and mouse controls. It can simulate physics, lighting, and character interactions consistently for up to a minute. Why it matters: It cracks a major bottleneck in robotics and embodied AI by providing endless synthetic playgrounds to train and evaluate future agents.
Researchers from the Allen Institute for AI and UC San Diego just introduced Spherical DYffusion, cutting a 10-year global climate simulation from 78 hours to under 3 hours. The model adapts diffusion architectures directly to spherical geometry to emulate the physics-based FV3GFS model, trading higher polar biases for massive compute and energy savings.
Altera put up to 1,000 LLM-powered AI agents into Minecraft and let them loose with basic prompts. The agents spontaneously formed specialist jobs, voted on tax reforms, and even spread a parody religion across towns. Why it matters: it shows how modern language models can mirror complex human social dynamics at scale, though they remain clever mimics rather than genuinely sentient beings.
MIT researchers just unveiled the Earth Intelligence Engine, a tool combining generative AI with physics-based flood modeling to generate realistic satellite imagery of future flood scenarios. While standard GANs hallucinated water on dry hilltops, fusing them with physical simulations keeps the disaster projections accurate enough to guide actual evacuation choices.
OpenAI is funding Duke University researchers to build algorithms that can predict human moral judgments in complex legal, medical, and business scenarios. Part of a 3-year, $1 million grant, the project aims to turn AI into a "moral GPS." Why it matters: Morality is notoriously subjective, making ethical reasoning one of the hardest nuts for statistical models to crack.
Google DeepMind and Google Quantum AI just introduced AlphaQubit, a Transformer-based neural decoder targeting quantum error correction. Published in Nature, it cut error rates by 30% compared to leading scalable decoders on the Sycamore processor. It sets a new benchmark for decoding accuracy, though it is currently still too slow for microsecond real-time correction.
ChatGPT just outperformed human doctors in a diagnostic study, scoring 90% on medical cases compared to 74% for unassisted physicians. Why it matters: doctors paired with ChatGPT barely improved at 76%, largely because physicians refused to listen when the AI disagreed with them and struggled to leverage its capabilities.
NVIDIA, the Weizmann Institute, and Pheno.AI just unveiled GluFormer, a transformer-based generative AI model that forecasts glucose levels and health outcomes up to four years in advance. Trained on continuous glucose monitor data, it can also predict how a body reacts to specific foods, opening the door to precision preventive care for diabetes.
IBM just upgraded its Quantum Heron processor and Qiskit stack, supporting up to 5,000 two-qubit gate operations and slashing a 112-hour benchmark run down to just 2.2 hours. The platform also integrates IBM Granite models to generate quantum code and optimize circuits with AI. Why it matters: Bringing generative AI into quantum supercomputing speeds up algorithm development for physics and materials science.
Google DeepMind just released the software code for its protein-folding model AlphaFold 3 for non-commercial scientific research. The release follows six months of academic backlash over withholding the code, though accessing the actual model weights still requires an application request.
University of Tokyo researchers just demonstrated an AI system that screens for high blood pressure and diabetes using a brief 5-to-30-second video of your face and palm. It analyzes micro-changes in skin blood flow with up to 94% accuracy for hypertension and 75% for diabetes, completely skipping cuffs and needles. Why it matters: If miniaturized into smartphones, contactless daily health checkups could finally become a reality.
SingularityNET and the ASI Alliance just deployed AIRIS inside Minecraft, an AI agent that ditches pre-programmed instructions to create its own survival rules in real time. Running via the Vereya mod, it adapts pathfinding and obstacle navigation entirely on the fly to test autonomous learning for future robotics.
OpenAI just open-sourced SimpleQA, a factual accuracy benchmark packed with 4,326 short, single-answer questions built to expose hallucinations. Frontier models like GPT-4o struggle hard, scoring under 40% correct. Why it matters: Measuring how often models fabricate answers—and how persistently they overstate their confidence—is the first concrete step toward fixing AI's chronic guessing problem.
Biotech startup Iambic Therapeutics just unveiled Enchant, a multimodal transformer model designed to predict how drug candidates behave in humans from the earliest discovery phases. By pairing abundant lab data with tiny amounts of clinical trial data, it sets a new state-of-the-art benchmark for predicting human drug half-life and aims to eliminate years of risky preclinical R&D.
Meta's Reality Labs just introduced the Digital Twin Catalog (DTC), a dataset packed with over 2,400 physical objects modeled down to sub-millimeter precision. It provides researchers with ultra-accurate geometry and realistic material textures that were previously too expensive to produce, aiming to fast-track digital twin creation for AR and e-commerce.
Google just officially named and released pre-trained weights for AlphaChip, its reinforcement learning system that treats chip floorplanning like a board game. The AI generates superhuman circuit layouts in hours instead of months, and has already designed Google's last three TPU generations as well as MediaTek chips. Why it matters: AI is no longer just running on advanced silicon—it is actively redesigning the hardware that powers it.
Harvard Medical School researchers just launched TxGNN, the first AI model specifically built to repurpose existing medicines for rare and untreated conditions. It evaluated nearly 8,000 drugs across 17,080 diseases, beating rival models by roughly 50% while explaining its clinical reasoning. The tool is free for clinicians.
