Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Lattice just launched a local LaTeX workspace built for researchers, packing writing, PDF previews, and citations into one desktop hub. It features two-way sync with Overleaf and connects directly to coding agents like Claude Code to fix build errors and refine papers, cutting out constant app switching.
Anthropic just shared a look into how AI is transforming scientific research. Instead of researchers manually writing and weighing competing theories, AI can now write, run, fit, and tweak models as code to rapidly uncover the explanations backed by actual data.
OpenAI teamed up with biotech firm Ligo to launch BioBasicsBench, evaluating how AI agents interpret raw data across 100 routine wet-lab tasks. Running at maximum reasoning effort, GPT-6 Astra scored 66.7%, comfortably beating GPT-5.6 Sol's 43.8%. Still, the models struggle with control checks and colony counting, showing AI lab assistants aren't ready to fly solo just yet.
Physical Superintelligence (PSI) just settled the 18-year-old Aaronson-Ambainis conjecture using a research system powered by an unreleased Anthropic model. The entire proof was machine-verified by the Lean theorem prover in just three days. Why it matters: The newly minted theorem mathematically proves quantum computers cannot achieve exponential speedups on unstructured random data, setting a hard theoretical ceiling on quantum hype.
Anthropic just tested its Claude models inside the brutal economies of EVE Online and EVE Frontier. Models like Fable 5 tripled earlier returns and grew initial capital by up to 600% through disciplined spread trading, but in live four-agent co-op missions, the team completely failed to catch a covert saboteur. Why it matters: Frontier models are getting scary good at mechanical market making, yet multi-agent oversight and trust remain wide open vulnerabilities.
Anthropic just unveiled a new agent memory framework for Claude inspired by human sleep, dubbed 'dreaming.' During idle time between sessions, agents review raw logs to correct faulty assumptions and distill broad strategies rather than hoarding every trivial detail. Why it matters: smarter agents don't just remember more—they know what to prune.
Anthropic AI for Science grantees just outlined a roadmap for how AI can automate scientific discovery instead of merely curve-fitting. Drawing on the classic Hodgkin-Huxley neuron model, the researchers argue AI agents must generate distributions of plausible models via simulation-based inference. It marks a push to turn AI from a black-box calculator into a tool for genuine scientific understanding.
Runway just halved the hardware footprint for its 57B-parameter Characters model from eight H100s to four while sustaining 24fps video. By moving FP8 quantization and RoPE to cache write-time and fusing communication kernels into matrix multiplications, it brought diffusion steps to 147.4ms. That makes serving live, interactive video avatars far more viable at scale.
Anthropic's biology lab just mapped the first 3D structure of a vault particle from prokaryotic cells, using Claude at every step of the research. Claude helped search protein databases, design DNA constructs, and process cryo-EM data—shrinking a weeks-long modeling process down to 34 hours. Why it matters: Vaults have baffled biologists since 1986, and this shows AI can meaningfully speed up real-world wet-lab discovery.
MIT researchers found that OpenAI’s GPT-6 Astra can inspect and directly rewrite neural network weights without any gradient descent or retraining. In one test, Astra modified just 0.016% of a small Qwen model’s weights to eliminate repetitive loops, bumping its math score from 7.9% to 48.1%. If frontier models can reliably program weight space, AI self-improvement might arrive much faster than expected.
DeepPrinciple just dropped Hydrogen-1.0, an open-weight, 753-billion-parameter model designed for chemistry, materials science, and biology. Fine-tuned on real laboratory workflows, it can write and self-debug simulation code, decipher NMR spectra, and compile protocols directly for lab automation. It even matches or outperforms GPT-6-Astra on several core scientific benchmarks.
Researchers introduced Weight-Space Programming, showing OpenAI's GPT-6 Astra can natively interpret and tweak neural network weights without traditional gradient training. By altering just 0.016% of Qwen3.5-0.8B's weights, it fixed a reasoning loop and bumped math accuracy from 7.9% to 48.1%. Why it matters: Direct parameter steering hints at automated model patching, though current results rely heavily on external Python tools rather than pure weight intuition.
Anthropic just open-sourced FlashProve, a cryptographic system using zkSNARKs to prove an AI response genuinely came from a specific model without revealing its private weights. Running on a single H100 GPU, it generates proofs for a 400B-parameter Llama 4 Maverick at roughly 110 tokens per second, which can be verified on a standard CPU in seconds. While proof generation is 70–95x slower than standard inference, it creates a practical defense against rogue model substitutions and tampering.
Google just fired up its Project Suncatcher prototype satellite, generating the first-ever AI response from a data center-grade TPU in orbit. Passing over Norway, the satellite ran Google's Gemma 4 model to beam back a space-themed haiku. Why it matters: Google is testing whether AI chips can survive extreme radiation and thermal swings, setting the stage for orbital data centers.
Anthropic just revealed it is running internal drug discovery programs after acquiring Mithril Medicine, a solo-founder biotech that used Claude to advance a muscular dystrophy candidate to mouse tests in 18 months. Claude also designed 35,000 antibodies in a month and cut key regulatory drafting from three months to four hours. While Anthropic warns its drug bets may never yield revenue, its wet-lab push signals a aggressive move into physical science.
Cortical Labs just networked 15 CL1 biological computers—each powered by roughly one million living human neurons—into a collective system called HIVE. In virtual navigation tests, the bio-collective scored around 70%, easily beating standard Deep Q-Networks at 32% and solo biological units at 40%. Teaming up living cells fixes individual biological variability, proving that wetware computing can reliably scale to outsmart conventional silicon AI.
OpenAI just revealed that its unreleased internal models made major breakthroughs across math, computer science, and physics—including disproving the decades-old Exponential Time Hypothesis and advancing a Millennium Prize problem. The company published raw outputs and Lean formalizations on GitHub while recruiting top academics, including Nobel laureate David Gross, to steward open verification. If these proofs hold, AI has officially crossed the threshold into frontier scientific discovery.
Researchers just introduced DecisionRAG, a new framework that ditches traditional vector retrievers in favor of decision models. Instead of slowly generating reasoning tokens, the model navigates documents like a library hierarchy, scoring entire branches in a single forward pass. It searches through 2.68 million Wikipedia passages in 0.7 to 1.8 seconds, beating dense retrievers with rerankers by double digits.
Microsoft just unveiled MAI-DxO in Science, an AI diagnostic system that acts like a panel of virtual physicians debating complex cases across rounds. On tough New England Journal of Medicine cases, it hit 83% diagnostic accuracy compared to 36% for doctors using reference materials—all while cutting estimated test costs by more than half. Why it matters: instead of answering one-off multiple-choice questions, the AI learns to order the right sequential tests, moving clinical superintelligence closer to real-world hospital workflows.
Researchers just introduced HiP3R, a hierarchical pose graph framework for long-horizon 3D reconstruction. It aligns video chunks first and then refines frame-by-frame camera poses to kill accumulation drift. Why it matters: it plugs directly into existing pipelines without retraining foundation models, turning lengthy video walks into consistent 3D scenes.
