Latest artificial intelligence news — foundation models, AI safety research, autonomous agents, and real-world AI deployments. Updated every 30 minutes from top AI publications.
Meituan released LongCat-2.5 on November 2, 2026, a multimodal AI model with 1.6 trillion total parameters and roughly 48 billion active parameters per token, available through an API and built for long-horizon agentic t…
OpenAI said on October 30 that it plans to honor a shared robots.txt signal letting site operators state a preference on whether their content is used for foundational generative-AI training, with Google announcing suppo…
Cisco introduced AI Transit, a capability that lets existing Cisco SD-WAN customers connect AI workloads running in NeoClouds, private and sovereign AI clouds, and enterprise AI factories onto their existing SD-WAN fabri…
A community-maintained mainline Debian port for the Retroid Pocket Duo Lite is missing working microphone capture because upstream Linux never wired the VA macro's DMIC clock enable into the TX macro, according to a tech…
Companies are losing more than $100,000 per day to illicit LLM charges as attackers steal API tokens and resell or abuse access to large language models, according to a report in The Register. The piece frames token thef…
Hume AI on Monday launched the Hume Audio Decision API, built on its audio-native decision model, which answers many structured questions about raw audio in parallel with confidence levels rather than predicting the next…
Developer divagr18's Qwenputer project compiles a complete 32-bit RV32IM RISC-V processor into the weights of Alibaba's Qwen3.5-4B as 11 transformer layers whose values are calculated by a compiler rather than trained, a…
Chinese memory maker CXMT is preparing to begin limited HBM3E production within two months, with estimated yields of only 20–30% versus roughly 70% for SK hynix, according to a weekly China tech newsletter. CXMT's perfor…
Schmidt Sciences awarded $732,574 to a research team including Brown University archaeologist Parker VanValkenburgh and Vanderbilt University researchers Steven Wernke and Yuankai Huo for a project titled "Scaling Archae…
Google's Gemini 4 Carbon ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59 and leads 15 of 17 benchmarks in Google's own comparison table, according to figures published by Google in its laun…
In a two-part test, Claude Fable 5.1 caught 59 of 60 deliberately planted mistakes in AI-generated business plans while DeepSeek 4.1 caught 50, but on 31 anonymized real small businesses whose later problems were known, …
In experiments run by Y Combinator Labs and Olam Labs, 50 GPT-6.1 Sol agents built a peaceful civilization in Minecraft and responded to a harsh in-game night by inventing chicken handoff protocols, a Fertilizer Officer …
A review by Kapoor and Narayanan found that 294 published research studies across 17 domains produced overly optimistic results because of data leakage, a problem the article examines in the context of k-fold cross-valid…
SemiAnalysis's new GDPx benchmark, spanning 206 tasks across finance, materials science, cybersecurity, software engineering, legal and biology, ranks Anthropic's Claude Opus 5.5 first, with Google's Gemini 4 Carbon clos…
DeepSeek published dsh-superagent, an MIT-licensed research prototype coding agent for a single GitHub repository that accepts requests from GitHub and optionally Feishu and replies with pull requests. DeepSeek says the …
A paper submitted to arXiv on 30 Oct 2026 proposes Latent Sparse Attention (LSA), an attention method that hierarchically clusters keys in a learned latent space so queries score clusters instead of every key, keeping ac…
DeepSeek unveiled Creator Mode inside DeepSeek Harness, the open-source agent framework it shipped as a developer preview on August 13, letting users describe a plugin in plain language and have Harness build and install…
Meta FAIR and Meta Superintelligence Labs researchers published an arXiv paper on October 29, 2026 introducing TextSeal2, a watermarking technique that detected 89% of 200-token AI-generated outputs at a 0.1% false-posit…
Cybersecurity leader Sean Ofodile outlined a practical approach to securing AI agents against indirect prompt injection and excessive agency, arguing that system-prompt warnings are not a security boundary. His patterns …
LinearB launched Agentic Studio on November 2, 2026, a unified context layer that exposes engineering delivery data — Git repositories, project trackers, sprint history, code ownership maps, rework rates and incident log…