AI Research news and analysis on Web Pulse: 21730 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
A developer argues that benchmark saturation and shifting evaluation methods make it increasingly difficult to judge how much better new AI models like GPT-6 Astra actually are, citing Astra's system card examples of ret…
A challenge called Goo AI Arena has opened, asking participants to use AI to build the tallest standing tower in World of Goo using all 300 balls through legal in-game building steps. The first entry, built by the author…
NVIDIA Transformer Engine is now integrated with JAX to accelerate dropless Mixture of Experts (MoE) training, targeting the conditional computation bottleneck in architectures used by DeepSeek, Qwen, and Mixtral. The in…
IBM and NASA released the NASA-IBM Lunar Foundation Model as open source on Hugging Face under the Prithvi family, cutting root-mean-square error in flagging likely lunar ice deposits by up to 22% against a SwinV2-B base…
NVIDIA reported that combining its Transformer Engine library with the JAX Python library raised DeepSeek-V3 MoE training throughput on NVIDIA GB200 GPUs from an unoptimized baseline of 103 TFLOPS/GPU to 1,068 TFLOPS/GPU…
A developer released neural_texture_es2, a self-contained C++ experiment that encodes an RGB image or up to four PBR material textures as a shared low-resolution latent texture plus a tiny MLP decoder, trained entirely w…
Researchers Florian Hahlbohm, Jorge Condor, Linus Franke, Martin Eisemann and Marcus Magnor published an arXiv paper introducing a compact neural appearance model for 3D Gaussian Splatting that reduces the per-primitive …
StackBlitz launched Forge on September 14, 2026, a new mode for its Bolt.new AI app builder that runs exclusively on open-source models and gives Pro subscribers up to 50 times more usage at no additional cost through Oc…
Researchers at MIT developed an algorithm called HardFlow that forced generative AI outputs to satisfy strict, non-negotiable safety constraints in simulated robotics and imaging tests, meeting the required rule every ti…
A new paper, OpenAgentFlow, proposes a control-plane architecture that enforces safety boundaries across heterogeneous AI agent fleets rather than at the individual model or tool level. The design splits an Action Plane,…
Richard Socher, founder of You.com, AIX Ventures and now Recursive, said Recursive has raised a $4.65B seed round and assembled open-endedness and self-improving-agent researchers to build an AI system that automates AI …
A healthy LLM application must consistently generate useful results while staying within acceptable limits for accuracy, safety, speed, reliability, and cost, according to a technical explainer on LLM evaluation. The pie…
A single 2B-parameter vision-language model placed first in the ≤2B parameter division of the EgoLongQA track of the Wearable-AI Challenge at ECCV 2026, scoring 0.8279 on the held-out test set. The system, described by i…
A submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge ranked first in the large-model division and second in the <=2B division, according to the team behind the entry. The task requires a wearable…
A technical textbook titled "Foundation Model Engineering: From Theory to Production" is being released for AI engineers and research-oriented readers, covering architectures, training pipelines, inference systems, retri…
The UK and US governments are set to sign two linked agreements on September 14, 2026, at the Global Fusion Summit in London to accelerate AI and advanced computing in fusion research, centered on a supercomputing partne…
TechCrunch Disrupt 2026 will host a session on how artificial intelligence, genetics, and computational biology are being applied to conservation research, including the work of a billion-dollar startup pursuing species …
Developer ljedrz published nachalnik on GitHub, a set of verbatim transcripts from Claude Opus 5 sessions in which the agent's context could be read and rewritten by an outside party. In the recorded sessions the model l…
A new method called TRACE (Trajectory-robust Admission with Evidence Ordering) addresses the growing inference latency and memory costs of GUI agents that accumulate high-resolution screenshots as their trajectories unfo…
Abacus.AI released three open-weight Smaug models — Smaug Flash on DeepSeek V4 Flash, Smaug Mini on Qwen3.8 27B, and Smaug Agentic on Kimi K3 — tuned for long-running agent loops and published on Hugging Face. Smaug Flas…