AI Research news and analysis on Web Pulse: 22675 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
Cognition released SWE-2 on September 10, a coding model it says approaches Fable 5.1's performance at 64% lower mean rollout cost, according to the company's own FrontierCode benchmark. SWE-2 scored 50.0% on the Main se…
OpenAI reported substantial progress on the Navier-Stokes existence and smoothness problem, one of the Clay Mathematics Institute's seven Millennium Prize Problems, according to a New York Times report published Septembe…
Skild AI launched S1, a robot foundation model that learns previously unseen, long-horizon tasks from a single video demonstration without retraining, built and researched on NVIDIA AI infrastructure. In Skild's tests on…
Jacob Coxon, a former Anthropic researcher who resigned this week after three years across Anthropic and OpenAI, told NBC News' Tom Llamas on Wednesday that "there is a substantial probability that this technology could …
NVIDIA's BioNeMo Inference Runtime (BioIR) reduces biomolecular structure prediction latency by using CUDA Graphs to record and replay GPU execution paths, bypassing standard PyTorch kernel-launch overhead. The runtime t…
Anthropic released an interactive economic model projecting that artificial intelligence could lift U.S. GDP by as much as 32% through 2030 while putting knowledge jobs at risk. The model examines AI's potential effects …
OpenAI announced that an unreleased model resolved the Navier–Stokes existence and smoothness problem, one of the seven Millennium Prize Problems, with agents exchanging 2.7 million messages and consuming roughly 130 bil…
OpenAI said in a statement Wednesday night, reported by the New York Times, that it has made "substantial progress on another Millennium Prize problem" since completing its Navier-Stokes work, without naming the problem.…
A new benchmark called SchemeArena factorizes stress testing of scheming in LLM agents, examining how covert pursuit of misaligned goals arises from the interaction of instrumental goals, environmental affordances, overs…
César de la Fuente's lab is using OpenAI's Codex and ChatGPT to search living and extinct genomes for antimicrobial candidates to fight drug-resistant infections, according to OpenAI's blog. The researcher's work applies…
A new Agent Evaluation Metric (AEM) decomposes multi-turn agent quality into named, separately measurable sub-metrics evaluated turn by turn, starting with correctness through two sub-metrics: truthfulness and completene…
Edge0-AI released edge0, an open-source streaming MoE inference framework that runs a 35B-parameter language model on Apple Silicon iPhones at roughly 2.9 GB peak active memory for the edge0-35b tier and about 1.0 GB for…
Feyn released MultiMatte, a promptable image background-removal model built on Meta's SAM 3 that uses low-rank fine-tuning to modify 19.49M of SAM 3's 860M parameters, or 2.27% of the model weights. MultiMatte scores 0.9…
Developers built game experiments on MaleCNS v1.0, a connectome of an adult male fruit fly's brain and central nervous system released publicly in June and described in Cell in September, with researchers from Google Res…
DeepSeek introduced its DeepSeek-V4.1-Flash model into the Agent Arena for testing on complex real-world tasks, evaluated via a causal tracing methodology with scores expected soon. Prediction-market pricing puts an 87.5…
In the second essay of his AGI series, developer KunYuan argues that the word "General" in Artificial General Intelligence is dangerously underspecified, with researchers, labs, and the public using it to mean anything f…
Unconventional AI announced the first cohort of its 2026 Unconventional Grant, awarding $100,000 each to five projects selected from nearly 100 submissions. The winning projects are based at Stanford, UT Austin, Yale, MI…
Cognition launched SWE-2, a coding model that scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 while costing 64% less, and is available starting today in Devin Desktop and CLI. SWE-2 is post-trained f…
Arena.ai analyzed tens of thousands of benchmark responses and found that Anthropic's Claude Fable 5.1 writes in a more matter-of-fact but more verbose style than its predecessor, Claude Fable 5, with the newer model's l…
A technical breakdown of how tokenizer training data skewed toward English — with English and code making up 95% of Llama 3's training data — creates measurable cost, context-window, and accuracy penalties for other lang…