Meta has described Intelligent Delivery, its machine-learning system for directing limited serving capacity toward ad-ranking work expected to benefit most. Predictive models guide allocation across requests, hardware and ranking stages, while AI agents help investigate changes that engineers review before deployment. This is an account of an existing …
Google has introduced the Gemini Decisions API and Gemini Decision 1, an AI model that scores supplied options instead of generating an open-ended answer. The endpoint supports routing, classification and rubric-based evaluation over shared multimodal input. Google’s speed and calibration claims are developer-reported, not guarantees for every workload.
A new Google Arts & Culture retrospective of South African artist Esther Mahlangu combines more than 45 curated stories with archival material and AI-generated visualizations of her early memories. Curator Nontobeko Ntombela describes a process built around interviews and workshops with the artist. The generated scenes illustrate recollections; …
Google DeepMind has introduced AlphaGenome 2, an AI model that predicts how DNA may affect gene regulation in a specified cell type or biological state. It is available for non-commercial research through the AlphaGenome API and Colab tutorials; commercial access is a private preview. DeepMind explicitly says the …
Hugging Face has published a training example for compressing the token vectors used by AI retrieval models without losing as much retrieval quality. Its matched experiment adds a pooled-document loss during training, rather than applying compression only after training. The results are author-reported benchmark measurements, not independently replicated …
Anthropic has launched Claude Sheets and Whiteboards in beta and made the renamed Claude Data generally available. The AI assistant can build and edit shared spreadsheets, boards and data presentations. Access still depends on paid-plan and organization settings, and Anthropic’s current primary announcement describes Claude Meetings as coming …
Hugging Face now offers Google Cloud Marketplace as a payment route for organization subscriptions and AI services, including model inference and GPU compute. Linking an account is not enough to activate it: Hugging Face reviews each subscription, and existing card-paying organizations switch at their next billing period.
Google has announced new Gemini-powered Android capabilities and a staged expansion of Gemini Intelligence to more devices and regions. Some updates are available with the 10 November Pixel Drop, while custom widgets and proactive help across other screens are still coming; access varies by feature, device, country, language …
Tether AI Research has introduced BotPsy, a 0.8-billion-parameter language model finetuned from Qwen3.5-0.8B for function calling and agent workflows. The 11 November research announcement targets low-to-medium-difficulty tasks; its model card explicitly excludes high-stakes decisions and applications involving money, personal information, health or other sensitive topics.
Google has outlined planned Adobe integrations spanning Gemini, Gemini Enterprise and AI Mode in Search, alongside exploratory image-model work with Google DeepMind for Photoshop. The 10 November announcement describes upcoming capabilities, not a fully available package.
Meta has detailed how Triton Low-Level Extensions (TLX) let engineers coordinate GPU work for recommendation systems and language-model training. The 10 November account describes tools already used internally for months, not a first-day release or a general speedup guarantee.
Meta has described how Your algorithm uses large language models to turn people’s stated interests into Facebook Feed and Reels recommendations. Its 10 November engineering account explains a shared preference profile, task-specific models and changes to retrieval, ranking and cached content; it does not establish a new worldwide …
Google DeepMind has introduced Gemini Robotics 2.5, a vision-language-action AI model that turns visual observations and instructions into robot actions. Google says it is rolling out to robotics partners and reports a Mercedes-Benz automotive-parts kitting pilot using Apptronik's Apollo 2, not general availability or proof of universal reliability.
Google DeepMind says AlphaProof Code, a multi-model AI-agent framework, proved specified security properties of code implementing Signal's cryptographic protocol. Humans reviewed the mathematical specifications, while AI agents generated proofs checked by Lean. This is a claim about selected code properties, not proof that the entire messaging application is …
Google and Artificial Analysis have introduced Multimodal Knowledge Agent (MKA), a benchmark for AI agents working across visually complex files. It tests projects such as interpreting drawings and producing reports or spreadsheets, rather than treating professional work as text-only question answering.
NVIDIA says it is contributing OpenShell, a runtime that applies deterministic controls to AI agents' tool use, inference access and workloads, to the Cloud Native Computing Foundation. Its KubeCon announcement also introduces Kubernetes deployment previews and checkpoint restoration for AI inference.
Google is widening access to Gemini 4 Carbon, its AI model for complex coding, knowledge work and scientific reasoning, after an earlier restricted cybersecurity rollout. The November 10 announcement extends paid API and enterprise access and begins an Ultra-subscriber rollout, rather than unveiling the model for the first …
An Amazon-authored explainer published on 9 November describes HOPE, a method for shrinking mixture-of-experts language models by considering which experts work together. These models route each token through only some of their neural-network components, but retain the wider expert pool in memory. HOPE uses pairwise activity statistics to …
NVIDIA and Elith are presenting Nemotron-Personas-Finance-Japan as synthetic profiles for training and evaluating Japanese-language financial AI agents. A 9 November post describes a pipeline combining statistical graphical models with Gemma-generated Japanese text, giving developers artificial household contexts for conversations and user simulations rather than real customer histories. The …
NVIDIA introduced GR00T 2 and Isaac Lab 3.0 at CoRL on 9 November, pairing a learned robot-action model with simulation tools for training dexterous skills. GR00T 2 is trained on robot and human data to help robots learn manipulation tasks; Isaac Lab supplies environments for training and testing …
NVIDIA announced an investment in Point2, which builds interconnects for moving data between chips and racks in AI computing clusters. Its 9 November statement separately says investments in Motif Technologies, RLWRLD and Upstage are planned; it gives no amounts or terms.
NVIDIA has announced Dynamo deployment recipes for four Korean open AI models, optimized for B200 GPUs. The recipes address distributed inference: coordinating requests and communication when serving a model across multiple GPUs. This is deployment tooling, not a new model launch.
TikTok’s For You feed uses machine learning, a form of AI, to predict and rank videos for each user. The European Commission issued preliminary Digital Services Act findings on 6 November over the feed’s safeguards against repeated exposure to harmful themes. TikTok disputes the account. The case has …
Google Research's 4 November Agile Deliberation account describes AI image classifiers built with vision-language models. Models surface borderline images for human decisions, and an LLM-driven loop refines the classifier prompt as users clarify subjective categories. The initial labels are not treated as a fixed specification.
NVIDIA’s 6 November research roundup describes 5G testbeds linking AI-driven robot decisions, learned signal receivers and AI-assisted network power management. These are controlled research experiments, not evidence of a commercial 6G deployment or a new Aerial open-source release; NVIDIA says that release happened a year earlier.
OpenAI and Ligo introduced BioBasicsBench on 6 November: 100 tasks built from experimental evidence. The evaluation rewards analysis and recognizing insufficient evidence; it does not establish reliable autonomous laboratory work.
OpenAI's 6 November Hacktron partnership combines security researchers with frontier AI models to test its systems, especially third-party connections. Hacktron builds AI agents for code review, penetration testing and red teaming; expanded model access is through Daybreak Red. This is a research collaboration, not evidence every integration is …
The Open ASR Leaderboard, which evaluates automatic speech-recognition models, announced two Liva AI conversational-speech sets on 6 November, adding Egyptian- and Nigerian-accented English. The change expands model-evaluation coverage and accent-level reporting; it does not establish a new best model or representative performance for every speaker.
Ai2’s 6 November ToyBlocks release combines a tabletop robot simulator, automatic demonstration generation and evaluation tools. Its deliberately narrow block-manipulation setting lets researchers vary training data and separate planning failures from execution failures; it is not proof of general-purpose robot competence.
NVIDIA’s 5 November Vera CPU post examines AI agents that alternate between GPU model reasoning and CPU execution of searches, code and other tools. It argues that this infrastructure needs both quick execution and high capacity per watt, presenting NVIDIA and partner results across coding, orchestration and sandbox …