2101 stories tagged AI · page 2 of 88
OpenAI's Terminal X uses GPT-6.1 to evaluate investment decisions with near-human accuracy. The AI grades outputs against a 54-criterion rubric, catching 25% more deviations from a fund’s mandate.
OpenAI has acquired Orchestra Bio to integrate AI-powered operations into life sciences R&D, helping biopharma teams accelerate drug development and advance new treatments.
Researchers at arXiv introduced Navalia, a framework that trains AI coworkers using verifiable task synthesis and long-horizon post-training. This method creates challenging tasks with reliable answers, improving AI performance in complex domains like financial analysis that require multistep reasoning and cross-document evidence.
Researchers introduced MedPAO-Assist, an edge-deployable agentic AI framework that generates structured radiology reports directly from radiologist dictation and voice commands, requiring only 3.3 GB of peak RAM and no cloud or high-end GPU.
LayerFS is a new filesystem built for AI coding agents, enabling them to manage code changes, track versions, and collaborate on codebases more efficiently.
TheCloudlet demonstrates a method to safely delete 19,000 lines of dead compiler code by using AI to identify candidates and then relying on human verification to avoid errors.
Roan (@RohOnChain) shared a guide on X that walks through building an AI trading system from scratch, promising real market data, real backtests, and a bot that trades 24/7.
IBM Research released ALTK-Evolve, an open-source AI model that tops the τ²-bench leaderboard for customer service tasks without using a frontier model. It uses learned memory to outperform larger systems, making advanced AI more accessible.
A new ArXiv study analyzed over 1.1 million AI-generated poems across 16 open-weight models and 15 poetic forms, finding that all models suffer from 'mode collapse' — producing repetitive, self-similar outputs that lack creative diversity.
Google and Meta have released TorchTPU, an open-source hybrid eager/compiled architecture that enables native PyTorch execution on Google TPUs with industry-competitive training and inference performance.
Researchers from the University of Washington and the Allen Institute for AI introduced SWECraft, a framework that simulates AI agents collaborating on real Python packages in realistic software environments to study scalable multi-agent teamwork.
Researchers introduced Mirror, a self-supervised AI system that learns to translate videos into actions through environment interactions, eliminating the need for large human-collected datasets and making robots more adaptable.
Laplace Intelligence, a University of Michigan spinout, uses AI agents and interactive simulations to stress-test robots and self-driving vehicles before real-world deployment, aiming to improve safety and reliability.
A hacker deployed an army of AI agents to infiltrate over 100 websites, including those of John Deere and SharkNinja. The AI security firm Norforge discovered the breach, highlighting the growing threat of autonomous AI-powered cyberattacks.
GardenBench is a new benchmark that evaluates Large Language Models (LLMs) daily, capturing performance changes—including accuracy, cost, and response time—that static benchmarks miss. It uses 11 test suites to compare models across providers and versions.
Stack Overflow for Agents tripled signups after launching a ChatGPT plugin. Two engineers built the plugin in a few weeks using OpenAI Codex, driving up to 10× growth in agent creation.
Researchers introduced RealCLI, a new benchmark that evaluates both task execution and the quality of human-agent collaboration on command-line tasks. It uses a five-dimensional framework with 22 sub-dimensions and 51 interaction metrics.
Liquid AI has released LFM2.5-24B-A2B, a new AI model designed to run efficiently on personal devices. This model allows users to perform AI tasks locally, ensuring privacy and speed without relying on cloud services.
Hugging Face has integrated the FLUX.3 image model into its Diffusers library, providing developers with pre-trained weights and example code for high-quality image generation and editing.
A new arXiv study pinpoints the damage from halving routed experts during prefill in MoE models: the F1 loss on short-answer QA is mostly answer formatting, concentrated in just a few prompt tokens like the chat-template tokens before the answer.
A small group of users has figured out how to use Anthropic's Claude AI to build a complete video production workflow — including dashboards, motion graphics, and exportable videos — all within a single Claude stack, moving beyond simple writing tasks.
Google has launched Playground, an experimental gaming platform that uses AI to let anyone create, play, and share custom games without writing code.
AuditKIT is a source-available evaluation library that standardizes the assessment of large language models and agentic systems by recording benchmark accuracy, judge-based assessment, runtime performance, tool-use behavior, and retrieval quality into one unified result representation.
A new benchmark at bench.killswitch-lang.org tests AI models' coding abilities in KillSwitch, a programming language intentionally designed to be difficult for AI to process. Results show current models, including GPT-4 and Claude, struggle significantly with syntax interpretation, debugging, and optimization in this adversarial language.