AI in Science News

Collection offers 144 ready-to-use scientific agent skills

Paper: Human oversight key for AI social science reliability

New paper argues that human oversight is critical for reliable AI-assisted social science research. LLMs are increasingly used for hypothesis generation, specification choice, and drafting conclusions, but reliability depends on human judgment and validation.

Podcast explores AI's ability to invent general relativity

Adam Brown discusses why inventing general relativity is a crucial test for AI, covering challenges and implications. The conversation delves into how current AI systems compare to human scientific reasoning.

Google Research head discusses AI accelerating scientific progress

Monkey brain controls robot in Japan via thought, 2008

Two papers advance AI bird call monitoring

A paper introduces a method for time-frequency localization of bird calls in dense soundscapes. Another releases a strongly annotated passive acoustic dataset for tropical bird monitoring. Both aim to improve large-scale biodiversity assessment via passive acoustic monitoring.

AI research deciphers cell behavior behind varied cancer drug response

Ai2 introduces ACE2S-SHiELD+, a climate emulator

GPT-1.5 used to translate 23,000+ ChinaRxiv papers

Podcast explores how OpenAI model disproved 80-year-old Erdős conjecture

OpenAI researchers Alexander Wei, Hongxun Wu, and Lijie Chen discuss how their model found a counterexample to the Erdős unit distance conjecture, a problem unsolved for 80 years. Mathematician Timothy Gowers called it a 'major open problem' solved by AI.

How Terry Tao Became an Evangelist for AI in Math

Quanta Magazine profiles Terry Tao's advocacy for AI in mathematics, adapted from Kevin Hartnett's book 'The Proof in the Code'. Highlights include his participation in a 2014 panel on AI and mathematical discovery.

Bloomberg explores Chinese brain-chip startup operations

The video report examines the development and technical infrastructure of a brain-chip startup based in China. It provides an inside look at the company's research facilities and current progress in neural interface technology.

The weather and climate science AI revolution isn't revolutionary

An Ars Technica analysis argues that the hype around AI in weather and climate science exceeds actual advancements, noting that traditional physics-based models still outperform AI in many areas. The article examines specific limitations of current machine learning approaches for forecasting and climate modeling.

Multi-Scale Feature Attention Network for polymer classification via THz spectroscopy

The network uses attention mechanisms across multiple scales to classify polymers from THz spectral data. It is designed to improve sorting accuracy for recycled plastics.

Making Claude a chemist

Anthropic's David Kamber tested Claude on NMR spectrum analysis, a standard chemistry task. The company is collaborating with chemists to improve Claude's chemistry skills; the CAS registry contains over 290 million substances.

Biomazon dataset aids 3D forest structure modeling in Amazon

Biomazon provides multimodal 3D forest structure data for the Amazon Basin, enabling accurate carbon accounting. It goes beyond traditional canopy-top height proxies by offering spatially explicit characterization.

Language model hidden state dynamics predict human processing costs beyond surprisal

Study finds that the trajectory of hidden state changes in language models explains human reading times better than surprisal alone. The paper introduces Trajectory Dynamics as a new predictor of cognitive processing during language comprehension.

EpiEvolve uses self-evolving agents for pandemic forecasting

The system handles streaming data with label arrival delays and disease regime shifts, improving over static forecasting approaches. It employs LLM-based agents that continuously adapt to new data.

Interpretable AI for osteoarthritis structure-pain studies from OAI

Framework combines deep learning-based MOAKS prediction with interpretable statistical modeling to analyze structure-pain relationships. Uses longitudinal data from the Osteoarthritis Initiative (OAI) to scale the study.

Horse eye blink detection for pain assessment

Paper on automated detection of equine facial action units (half and full blinks) as indicators of pain and stress. Uses computer vision to analyze micro-expressions.

Anthropic shows Claude can interpret NMR spectra like a chemist

Claude matches and on some tasks beats dedicated NMR interpretation software, Anthropic reports. The model was tested by chemist David Kamber on the standard analytical input of NMR spectra. The work is part of a broader effort to improve Claude's chemistry capabilities.

DataDIVER discovers concise computational models from data

Podcast features AI tools for solopreneurs, therapy, and science

The Cognitive Revolution podcast interviews founders of AI tools for solopreneurs, mental health, and scientific research. Guests include Hooman Radfar (Collective), Taras Pohrebniak (ElomiaHealth), and Peter Jansen (Allen AI).

Jeff Bezos funds Flourish's $500M hunt for brain's 'core algorithm'

Flourish, a neuroscience startup, has raised $500 million from Jeff Bezos at a $2.5 billion valuation to reverse-engineer the brain's core algorithm and reinvent AI. The company aims to build AI by studying real neurons under the microscope.

Paper examines LLMs in computational conceptual history

An arXiv paper explores how large language models can be used to study the history and philosophy of scientific concepts, building on earlier digital methods. The authors examine what LLMs add to existing computational approaches.

Derivative Informed Learning of Exchange-Correlation Functionals

Paper proposes a machine-learned approach to exchange-correlation functionals that uses derivative information to improve accuracy. The method aims to consistently outperform traditional O(N^4)-scaling density functional approximations.

Physics-Informed Machine Learning for Short-Term Flood Prediction

Proposes physics-informed ML integrating physical constraints for flood forecasting in data-scarce environments. Aims to improve accuracy by embedding hydrological principles into model training.

LLMs for scientific reasoning in simulation-driven decisions

Paper proposes a framework integrating LLMs with scientific simulators for high-stakes decision-making. Treats LLMs as reasoning engines that simulate, reason, and decide, extending beyond generation or calibration tasks.

Tabular RL method for fair metro network expansion proposed

Researchers introduce a tabular reinforcement learning approach for the Metro Network Expansion Problem (MNEP), aiming to satisfy travel demand while considering fairness. The method is evaluated on benchmark instances, showing competitive performance against traditional exact and heuristic methods.

SCI-PRM: tool-aware process reward model for scientific reasoning verification

Proposes SCI-PRM, a process reward model adapted for scientific reasoning in biology, chemistry, and physics. Incorporates tool awareness to verify domain-specific reasoning steps.

Stein kernelized MD method improves active learning of interatomic potentials

Introduces Stein kernelized molecular dynamics (SKMD) for enhanced sampling. Method uses a kernel-based drift force to efficiently explore configuration space for training data.

Channel-Oriented Design for EEG-to-Music Reconstruction

The paper proposes a channel-oriented design for reconstructing music from EEG signals, a far less explored setting compared to vision and language. The method aims to decode naturalistic music stimuli from brain signals.

Metric-Aware Hybrid Forecasting for the CTF4Science Lorenz Challenge

The paper describes a hybrid forecasting approach that combines short-horizon prediction, long-time distribution matching, and trajectory reconstruction for the CTF4Science Lorenz challenge. The key discovery is that no single model family excels at all task pairs, necessitating a hybrid solution.

Biohub releases world model of protein biology

The model ESMC was trained on 2.8 billion protein sequences. Lab-validated binders designed by ESMFold2 achieved high affinity in days.

Researcher designs full-atom peptides using geometric latent diffusion

Podcast revisits Axiom's perfect Putnam score in 2025

Seven-month-old startup Axiom solved all 12 Putnam problems, scoring 8/12 within the time limit, outperforming top undergraduates (110/120) and DeepSeek (103/120). The interview with Carina Hong discusses how Axiom's approach scales beyond informal AI.

Video explores AI co-scientist capabilities

Two Minute Papers video discusses an AI system positioned as a 'co-scientist' for research. No specific details about the tool are given in the source.

Stem-cell microrobots reconnect severed spinal cords in mice

OpenAI introduces new GPT-Rosalind capabilities

GPT-Rosalind gains enhanced biological reasoning, medicinal chemistry, genomics analysis, and experimental workflow capabilities for life sciences research. The update aims to accelerate drug discovery and genomic analysis.

Paper proposes auditable climate risk AI from ESG data

The paper introduces deterministic orchestration and imbalance-aware learning for validating Scope 1-3 emissions and climate data, emphasizing provenance-aware auditability and reproducibility. It addresses fragmented ESG reporting with a machine learning pipeline.

Conditional Hypothesis Generation for LLM Text Analysis with Covariates

The paper introduces a method for generating interpretable hypotheses from text using large language models, conditioned on researcher-specified covariates to discover language differences across outcomes like political affiliation. It targets computational social science applications.

GTBench evaluates LLMs as math research assistants in graph theory

GTBench is a curriculum-grounded benchmark testing LLMs as mathematical reasoning assistants in graph theory. It provides structured tasks to assess reliability and problem-solving capabilities.

Scalable uncertainty quantification for extreme weather forecasting via NTK

Paper introduces a scalable uncertainty quantification method using empirical neural tangent kernels for deep learning weather models. The approach addresses the critical lack of uncertainty estimates in deterministic AI weather forecasts for high-stakes extreme weather events.

Transformer and LSTM frameworks compared for prediction in ungauged basins

The paper evaluates both frameworks on ungauged basins, comparing performance and identifying key factors for accurate prediction. It highlights the challenge of missing direct observations in such basins.

AdaWeather adaptively mixes probabilistic weather forecasts

The paper proposes AdaWeather, a method that adaptively combines probabilistic weather forecasts to address the issue that no single model consistently dominates spatio-temporally. It achieves logarithmic regret in mixing multiple forecast sources.

Google unveils DolphinGemma AI for dolphin communication

DolphinGemma is a Google AI model designed to understand dolphin sounds and behavior, as detailed on the Google blog. The Reddit community eagerly awaits its release, expressing impatience over delayed availability.

Tool automates research with 23,000 AI agent skills

DeepMind launches Co-Scientist, a multi-agent system for science

Co-Scientist is a new Gemini-based multi-agent system that generates, debates, and evolves novel hypotheses for complex scientific problems. It aims to serve as a dedicated research partner to accelerate scientific discovery.

PDE solver efficiently scales to 100M geometries

Two Minute Papers discusses possibility of second Nobel Prize for AlphaFold

Over 3 million researchers already use AlphaFold, spurring discussion on whether AI-driven science could earn a second Nobel Prize. The video explores expert opinions on the historic impact of DeepMind's protein-folding AI.

Probe Before You Edit: LLM agents for drug design

Paper introduces a probing-guided framework using LLM agents to optimize molecular ligands in structure-based drug design, balancing binding affinity and druggability. The method iteratively refines molecules via a probe-edit loop.

Interactive map shows AI-assisted academic research lifecycle

CorticalLabs grows human neurons to run computers

WindBorne's WeatherMesh-6 AI model beats government forecasts by days

WeatherMesh-6 is as accurate five days out as a traditional forecast is the day before on surface temperature. It produces forecasts every hour at 3 km resolution in the continental US, besting the European Centre for Medium-Range Weather Forecasts.

Opus 4.8 leads new Singularity Gate benchmark for predicting future discoveries

Benchmark tests frontier AI models' ability to predict paradigm-breaking scientific discoveries published after their training cutoff. Opus 4.8 currently tops the leaderboard.

Brett Kagan teaches brain cells to play Pong faster than AI

Terence Tao on how AI is changing mathematics

Fields medalist Terence Tao and OpenAI Chief Research Officer Mark Chen discuss how AI is transforming mathematical research, enabling easier experimentation and collaboration on difficult problems. They explore the evolving role of AI in mathematical discovery and reasoning.

Aleph Prover formalizes OpenAI's disproof of Erdős planar problem

Multi-agent system automates end-to-end scientific research lifecycle

Largest protein data collection released on Hugging Face

AI tool generates atlas of over 1 billion protein structures

ESMFold2 protein structure prediction model released

Blog: GPT5.5Pro solves Erdős's Unit Distance Problem

Scott Aaronson's blog reports that GPT5.5Pro has solved Paul Erdős's 1946 Unit Distance Problem, a central open problem in discrete geometry. The model reportedly proved the conjecture that at most n^(1+o(1)) pairs of n points can be unit distance. The post reflects on the implications for AI-driven mathematics.

Latent Space podcast traces the story of ESMFold2 and BioHub

Alex Rives, Head of Science at BioHub, discusses the development of ESMFold2 and related models. The podcast covers the history from ESM-1 to ESMFold2, including BioHub's acquisition of EvoScale.

Researchers train AI model using IBM quantum computer, outperforming base model

Researchers trained an AI model on an IBM quantum computer, achieving better performance than a classical baseline. The quantum-trained model answered questions correctly that the base model could not.

DeepMind's AlphaProof Nexus solves 9 open math problems

AlphaProof Nexus autonomously solved 9 Erdős problems unsolved for up to 56 years at ~$200 each. It also proved 44 OEIS conjectures and resolved a 15-year-old algebraic geometry question.

Chart shows unsolved math problems solved by AI

A Reddit chart lists recent unsolved math problems solved by AI, highlighting growing AI mathematical reasoning capabilities. The post has garnered community discussion around AI's progress in mathematics.

Daedalus journal publishes special issue on AI and Science

Toward a science of intelligence: unifying physics, neuroscience and AI

Tool converts unstructured research to submission-ready LaTeX papers

University of Tokyo builds chip component 1000x faster, 1/100th power