Key AI papers, new benchmarks and methods, and what AI is discovering in biology, medicine, materials and other sciences.
Researchers just showed that rentosertib, an AI-designed drug candidate, significantly rolled back patients' biological age during a phase 2a clinical trial. By testing six proteomic aging clocks against serum samples, the team found treated patients shed an average of 2 to 3 years in biological age by week 4. The study proves standard disease trials can successfully double as anti-aging readouts.
OpenAI just revealed that its coding agents generate 3.1 agent-workdays for every human workday inside its lab, with the median researcher running through over $600 in daily inference. The company says it hit its target of creating an "automated research intern" and aims for a full AI researcher by March 2028. Why it matters: AI is genuinely speeding up its own development, though recent safety halts prove human control is still catching up.
Anthropic just open-sourced the first complete, machine-checked proof of Fermat's Last Theorem in Lean 4, generated entirely by AI agents. Spanning over 13 million lines of code and 29,511 theorems, the massive artifact successfully passed checks by two independent verification kernels without extra axioms. It marks a historic milestone for AI-driven mathematics.
Google DeepMind just launched WeatherNext 3, its flagship weather model now powering Google Search, Maps, and Gemini. By ingesting live geostationary satellite mosaics and ground station data, it delivers hourly forecasts at up to 5-kilometer resolution with up to 60% sharper precipitation tracking. Why it matters: hyper-local, real-time weather predictions are now available globally without traditional supercomputing lags.
Google Research and HHMI Janelia just unveiled the complete wiring diagram of a male fruit fly's brain and central nervous system. By combining AI reconstruction algorithms with human verification, the team mapped over 166,000 neurons and 125 million synapses—the largest connectome to date. It fully traces neural pathways from sensory perception to motor control, and researchers can explore the whole model right now via Neuroglancer.
Anthropic just used its Claude model to solve a decades-old holy grail problem in probability theory. Guided by researchers, the AI proved that percolation phase transitions are continuous in 3 to 10 dimensions, backed by machine-verified Lean code. Why it matters: AI is moving beyond competition math to cracking unsolved conjectures at the frontier of pure mathematics.
Google just paired its Antigravity Teamwork framework with Gemini 3.7 Flash, allowing autonomous agent teams to solve seven open math and computer science problems, including Knuth's Cycles Conjecture verified in Lean. The multi-agent setup also built an out-of-order RISC-V CPU simulator from scratch that boots xv6. Why it matters: Multi-agent systems are moving past simple coding scripts into genuine scientific discovery and complex systems engineering.
ByteDance researchers just introduced Lucida, a pipeline that turns indoor video captures into interactive 3D simulations. Instead of outputting a single frozen mesh, it builds clean 3D assets for each detected object and uses a vision-language agent to place them precisely back into position. Why it matters: It gives robotics and embodied AI fully manipulable digital twins of real-world spaces.
Anthropic just showed that Claude can autonomously research and patch 10 categories of alignment failures, including deception and sycophancy. In tests, Claude beat 28 human safety researchers by 20% on fixing deception, and Sonnet 5 aligned an unreleased Opus 4.8 checkpoint in 60 hours with 15,000 times fewer training examples. Anthropic has open-sourced the research harness.
Google Research just unveiled the Planetary Prediction Engine (PPE), an autonomous Earth AI capability that turns natural-language prompts into end-to-end geospatial predictions. Instead of weeks of manual data engineering, PPE retrieves planetary datasets, trains models, and generates reports in minutes, outperforming expert baselines in food security and disease tracking. Why it matters: It could massively accelerate real-time response during humanitarian and climate crises.
Anthropic just opened 10,000 free and discounted Claude subscriptions for researchers worldwide through a new team plan. Academic lab leads get free standard seats or $15 monthly premium tiers, plus access to up to $50,000 in compute grants across multiple scientific fields.
Bocconi University and OpenAI just ran a randomized trial on over 1,000 undergraduates and found that ChatGPT (GPT-4o) boosted rubric scores by nearly a full point on a five-point scale, making student work look expert-level. However, only critical-thinking exercises drove unique, outside-the-box ideas. AI polishes the answers, but human reasoning supplies the spark—meaning assignments must evolve to reward originality rather than tidy execution.
Surgeons at University College London Hospital just used a live AI assistant during a high-stakes brain tumor surgery for the first time. Built with Google and NIHR funding, the tool tracks surgical instruments in real time and highlights hidden blood vessels and critical nerves directly over the endoscope feed. The team is now prepping for larger trials.
Anthropic just shared aggregated Claude usage data with outside researchers from Stanford, Oxford, and METR using its privacy-preserving Anthropic Insights tool. Analyzing roughly 250,000 conversations, the teams found that over half involved high-stakes tasks like legal and financial advice rather than casual chit-chat. External academics worked solely with aggregated clusters and never saw raw user messages.
Anthropic just launched a $5 million grant program to fund independent research into how AI chatbots affect user wellbeing and mental health. Grantees will receive funding, model access, and technical support to build open-source benchmarks for high-stakes multi-turn conversations. Applications are open until September 21.
Alibaba DAMO Academy just unveiled DAMO LiON, a liver cancer diagnostic AI model published in Nature Medicine. In a real-world trial of over 10,000 patients, the tool cut physician reading times by 27%, boosted malignancy detection sensitivity by 11.5%, and caught 51 previously overlooked lesions averaging just 1 cm. Why it matters: it helps junior doctors match senior specialists in spotting tricky, early-stage cancers.
NVIDIA just revealed that its general-purpose agent architecture AVO achieved a 100% score on the ARC-AGI-3 public benchmark, solving all 183 levels across 25 environments. Built with persistent memory and a supervisor module, AVO required 12% fewer actions than rival harness VISTA. The experiment underscores that sustained agent performance relies on the harness as much as the frontier model itself.
Anthropic just revealed that Claude can design custom protein binders from scratch, successfully hitting 14 out of 15 biological targets with hit rates up to 35%—more than double typical industry campaigns. It also tackled analytical chemistry, parsing raw lab data in under 25 minutes. Why it matters: Frontier AI is moving past code and chat into physical wet-lab drug discovery, turning months of specialized orchestration into hours.
Researchers just introduced GeoWeaver, a framework that reconstructs complete and consistent 3D scenes from standard long videos. By breaking footage into chunks and stitching them with long-range geometric matching, it stops camera drift and sets the lowest average error on benchmarks. The project is currently available as a research paper.
Inherent Labs just unveiled Faraday, a 27B agent designed to replicate academic papers. Instead of writing code itself, it decouples scientific hypothesis from implementation by directing GPT-5.5 Codex as its tool, outperforming Claude Opus 4.8 and GPT-5.5. It shows that small models steering frontier tools can outmaneuver massive systems in scientific research.
