ResearchStartup Recursive demonstrates automated-research system improving language-model training and GPU kernelsRecursive, a newly founded AI research startup, demonstrated that its automated research system achieved new best results on three benchmarks: NanoChat Autoresearch, NanoGPT Speedrun, and SOL-ExecBench.2 min read
ResearchDeepMind outlines possible routes from AGI to superintelligenceResearchers at Google DeepMind published a paper mapping ways the field could move from human-level artificial general intelligence (AGI) to artificial superintelligence (ASI).4 min read
ResearchHumanoid robot progress will shape timelines for self-sustaining AIForecasters argue that whether artificial intelligence becomes self-sustaining depends largely on the development of humanoid robots and physical infrastructure.3 min read
ResearchStudy finds advanced AI more persuasive than human experts in text conversationsResearchers from the University of Oxford, the UK AI Security Institute, Stanford University and the London School of Economics ran four experiments with 6,923 participants and 18,978 conversations and found that current AI systems outperform humans in text-based persuasion, including real-money charitable donations.4 min read
ResearchNASA's ERNEST rover prototype demonstrates new suspension, gaits and autonomy for lunar and Mars explorationNASA's Jet Propulsion Laboratory has tested the ERNEST experimental rover in the Colorado desert, where it traveled about 26 km with minimal human intervention and reached roughly 1 km/h.3 min read
ResearchUC Davis develops hardware‑agnostic neural decoder that translates attempted speech into wordsResearchers at the University of California, Davis published a Nature Medicine paper describing BRAND, a machine‑learning platform that decodes attempted‑speech brain activity into phonemes and words.2 min read
ResearchAI-powered breath analyser to test causes of dyspnoea using exhaled VOCsAinos, an AI and biotechnology company, will partner with National Taiwan University (NTU) on a one-year study beginning in July to assess whether its Smell AI platform can diagnose causes of dyspnoea by analysing volatile organic compounds (VOCs) in patients’ exhaled breath.3 min read
ResearchThree New Benchmarks Aim to Replace SWE-bench for Evaluating Agentic Software EngineeringThree recently released benchmarks — DeepSWE, ProgramBench and ITBench-AA — propose more demanding evaluations of agent-driven software engineering than the SWE-bench family.5 min read
ResearchUsing Privileged Hints Improves RL Training on Competition Math ProblemsResearchers at Carnegie Mellon University introduced Privileged On-Policy Exploration (POPE), a training method that pairs the on-policy RL algorithm GRPO with datasets that include the initial steps of solutions as hints.3 min read
ResearchUN warns AI could consume significant electricity and freshwater by 2030, worsening inequalitiesA new UN environmental report projects that by 2030 artificial intelligence could use about 3% of global electricity and consume more freshwater annually than the world’s population drinks.3 min read
ResearchInitial step toward more useful and aligned artificial intelligenceResearchers and developers are working on training models so that the models' useful properties can be transferred to new situations; this is the first step to ensure that as AI capabilities grow it…1 min read
ResearchNew research: Training AI models for broad and lasting usefulnessThe research team developed new methods to enable artificial intelligence models to exhibit useful and safe behavior in novel domains beyond their training, and to maintain this under stressful, high-stakes tasks.1 min read
