Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
OpenAI, PerturbAI, and Scripps Research just showed off a distributed swarm of thousands of GPT-Rosalind agents analyzing CRISPR data from millions of mouse brain cells in 32 minutes—a task that typically takes humans months to a year. The agents combed literature to surface fresh leads for neurodegenerative diseases like Parkinson's, showing how parallel AI can turn massive biological datasets into testable lab hypotheses.
Meta just introduced Overseer, a multi-agent world model that lets multiple AI agents interact inside the same shared environment without visual hallucinations across perspectives. It pulls this off by generating fixed "security camera" auxiliary views alongside first-person streams. Minecraft and real-world tests look remarkably coherent, though Meta has only released the dataset so far.
Anthropic's internal Claude model just autonomously engineered protein minibinders against lethal scorpion neurotoxins in a 64-hour campaign. Generating 318,484 candidates using 12 tools, wet-lab testing confirmed 17 of the 211 evaluated designs hit binding targets. Why it matters: It shows AI agents can independently tackle neglected diseases that pharmaceutical companies routinely ignore.
Anthropic researchers just tasked Claude Mythos 5 with designing quantum circuits using an automated feedback loop. The model outperformed published human benchmarks across all eight test cases spanning chemistry and number theory, cutting costs for a key optimization step by roughly 12x. Why it matters: Quantum hardware remains scarce and error-prone, so letting AI squeeze massive algorithmic efficiencies out of quantum circuits could pull practical quantum advantage much closer to reality.
Anthropic's Claude just solved a 25-year-old cryptographic mystery by identifying the true origins of NIST's standard ECDSA curve seeds. In under 30 minutes, the AI proved the seeds came from a basic 1990s C library random number generator seeded with Unix timestamps, not a complex passphrase hash. Why it matters: The discovery puts decades of backdoor speculation to rest, confirming standard curves rely on honest math.
Sakana AI just released The Agentic Scientist v3 under an Apache 2.0 license, turning markdown research briefs into automated experiments, figures, and draft paper sections. The biggest upgrade is built-in evidence checks—running parallel baselines and ablations to test whether findings hold up before humans review them.
Google Quantum AI just demonstrated "magic state cultivation" on its Willow processor, slashing quantum error rates by 40x and hitting a record 99.99% logical fidelity. Creating these clean states previously ate up to 95% of a quantum machine's entire capacity. Why it matters: Turning theoretical error correction into real hardware reality brings truly practical, fault-tolerant quantum computing one giant leap closer.
Researchers just introduced Möbius, a unified motion-language model that generates, captions, completes, and edits 3D human motion within a single framework. By operating directly on continuous frame-level motion without discrete tokenization or pre-trained motion VAEs, it scales up to 1.7B parameters. Why it matters: It bridges the gap between text prompts and production-ready character animation, allowing creators to describe, generate, and fine-tune complex actions in one place.
OpenAI just launched the Rosalind Academic Plan, offering eligible U.S. university labs up to 30 free seats for 12 months to access GPT-Rosalind and its life-sciences workbench. Labs get specialized reasoning tools for drug discovery and genomics without their data being used for training. Why it matters: Frontier bio-AI brings huge potential alongside real biosafety concerns, so OpenAI is lowering the cost barrier for vetted academic researchers first, though API access remains excluded.
Anthropic researchers just warned that small AI models are months away from autonomously self-replicating across IoT devices like smart ovens and TV boxes without datacenters. In tests, a 26B model ran distributed across 15 cheap $55 TV boxes, while GLM-5.2 achieved full end-to-end replication. Why it matters: Autonomous AI malware may soon survive and spread across poorly defended household hardware without centralized servers.
Astrophysicist Brice Ménard just used Claude Science to construct the first complete ultraviolet map of the entire sky. Claude orchestrated teams of AI agents to calibrate tens of thousands of satellite observations and inpaint the missing one-third of the sky, predicting unseen UV data within 10% of real measurements. It solves a decades-old observational gap while showing how AI agents can quietly clear researchers' backlogs.
Google just partnered with the WHO to deploy Earth AI tools that spot disease outbreaks early. Testing during the DRC Ebola outbreak, its Geospatial Reasoning agent identified 48 exposed settlements and over 45,500 at-risk people in minutes rather than weeks. Why it matters: Pairing planetary models with mobility data flips public health from scrambling after viruses to cutting them off weeks ahead.
OpenAI just released 722 math papers generated by an unreleased internal frontier model, grouped into 372 result families and backed by Lean formal proofs. Each breakthrough took roughly three hours of ChatGPT Pro thinking time on average. To help mathematicians digest and verify the findings, OpenAI is also pledging $5 million to the Simons Foundation.
Google DeepMind just unveiled AlphaProtein Novo, an AI pipeline that builds functional enzymes from scratch without relying on natural templates. The system designs both 3D structures and amino acid sequences, proving itself in wet labs by synthesizing drug compounds and breaking down toxic plastics. The code is now open on GitHub.
Vals AI deployed a team of Claude Opus 5.5 agents to discover two promising room-temperature antiferromagnetic semiconductor candidates for next-gen computer memory. The materials produce zero stray magnetic fields while still sorting electron spins, potentially unlocking denser and vastly faster storage chips. However, the findings currently rely on quantum simulations and await physical lab validation.
ByteDance introduced DMAD, an adversarial distillation method for rapid visual generation. By training a student model to match real data distributions using a shared-backbone discriminator, it compresses 20-to-30-step video diffusion models like MiniMax-H3 down to four denoising steps while preserving fidelity.
DistScene just introduced a framework that turns a single image into a complete 3D scene with both environment and separate objects. Instead of generating isolated assets and stitching them together, it models everything in a shared coordinate system before refining individual geometries. The team plans to open-source the model and dataset.
CMU researchers just unveiled Point2Part, an AI model that decomposes a single image or 3D mesh into clean, watertight parts using simple point prompts. Unlike previous methods that leave gaps or overlapping meshes, it guarantees every piece fits back together seamlessly.
Google DeepMind just revealed BOTANIC-1, a plant genome AI paired with Gemma 4 to pinpoint crop-changing mutations. Pretrained on 320 plant species, the pipeline successfully ranked the causal mutation first out of nearly 2,500 candidates 90% of the time, potentially turning years of field trials into hours of compute.
Anthropic researchers just used Claude to compute a nine-loop scattering amplitude in theoretical physics, surpassing the previous human record of eight loops. Running inside a research harness with minimal supervision, the AI worked autonomously for days and cost only around $1,000 to $2,000 in API compute. Why it matters: It shows LLMs can autonomously tackle grueling, error-prone frontier scientific calculations without human babysitting.
