Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
A paper titled "Read, Allocate, Route: Role-Coupled Tokenization for Pixel-Space Generative Transformers" proposes role-coupled tokenization to cut the cost of pixel-space generative Transformers, which conventionally us…
AiDroneTree, an AI-based software powered by TreeNet, automatically detects, maps and measures individual trees from 3D point clouds derived from LiDAR or photogrammetry, producing tree-level data on location, height, cr…
Samaya AI released Sterling, a post-trained Nemotron 3.5 Lightning model with 3B active parameters built for financial research, which scores 58.9% on the FrontierFinance benchmark at $0.61 per question. That beats Claud…
IFlytek unveiled its Spark X2.6 foundational model on October 29, 2026, at the 1024 Global Developer Festival in Beijing, built on a hybrid cloud-device architecture that splits computing between end-user hardware and cl…
Samsung has filed a patent for a method that trains vision-language models to hallucinate less by having a pretrained model generate several captions for each image, scoring them with CLIP via cosine similarity on embedd…
Developer RuiNelson built a custom tool called RageDecompiler that transpiles Motorola 68000 assembly from the SEGA Mega Drive version of Streets of Rage into C++ for a native PC build, using Claude, ChatGPT and Grok 4.5…
A new benchmark, Investor's Last Exam, evaluates whether AI agents can perform equity investment research using only decision-time evidence, with their forecasts and investment judgments scored against subsequent market …
A developer built Nishi Daak, an Arduino UNO Q project that detects a resident's return via a Bluetooth presence scanner and then spells the person's name twice on an LED matrix, mimicking the Bengali Nishi spirit that c…
The tt-decomp project has decompiled 3,413 of the 13,545 functions (25.2%) in the Xbox 360 version of Rockstar Presents Table Tennis as of April 2026, lifting PowerPC assembly into C++ through an MCP-enabled Claude workf…
A research paper proposes certificate-guided visual routing for diffusion vision-language models (DVLMs), which process their full input at every denoising step even though visual tokens account for 89-97% of that input …
Researchers describe an online dense reinforcement learning approach that trains a planning-grounding agent for UI navigation on live iOS apps, targeting the high inference costs that make frontier vision-language models…
Mistral AI released the weights of Mistral Large 4, a decoder-only language model with 1 trillion parameters, under an Apache 2.0-compatible open-use license that permits commercial use. The model uses grouped-query and …
Google's Gemini 4 Argon leads Artificial Analysis's new HLE-Diamond leaderboard with 69.1%, ahead of Claude Opus 5.5 (max with fallback) at 67.9% and GPT-6 Astra (max) at 63.8%, on a 1,000-question closed-ended benchmark…
An open research initiative called Multa Lumina proposes routing every external channel of an AI system under evaluation — mailbox, chat, and tool accounts — through infrastructure held by the evaluator, so that a contac…
A developer released korea-data-mcp, an open-source MCP server that wraps the official free APIs of OpenDART, data.go.kr and law.go.kr so AI agents return correct answers on Korean corporate filings, stock prices, apartm…
A behavioral check of the released CAEL 0.3.7 non-neural assistant found that all 15 CPU calls with fictional inputs completed without execution errors, and six history/question pairs exactly reproduced the saved answer …
The University of Arizona reported a 32-fold rise in outreach productivity during a pilot of AI agents built by Risely, the startup co-founded by Shahryar Abbasi and Sadia Saifuddin, according to Y Combinator's company l…
A developer documented the setup pitfalls of running MOSS-Image, OpenMOSS's 5B-parameter Apache 2.0 text-to-image diffusion transformer released 16 October 2026, in ComfyUI, finding that file placement, loader incompatib…
Google announced Gemini 4 Carbon on October 29, 2026, a model built for finding, proving and patching security flaws in real code, with reported scores of 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71% on CWE-bench.…
A blogger probing TypeSafe's Jev AI API found the model misidentifies its own origin, answering that it is more likely Qwen than Jev, and is highly sensitive to question wording. Asked to pick a random digit, Jev chose 5…