
Welcome back! This week has brought some extraordinary developments across the AI industry, including the resignation of Nvidia co-founder Jensen Huang, leaving the AI chip giant searching for a new leader. Anthropic launched its IPO roadshow, proposing to offer 825 million shares, while Australia announced plans for mandatory AI safety standards and a new regulator. In the US, lawmakers advanced legislation addressing data-centre energy costs. Researchers also found that pharmaceutical advertising could influence AI treatment recommendations. For an industry built on bold predictions, this week delivered some very real consequences.
- Anne, and the News DigestAI team

📰 Latest AI News
Nvidia confirmed Jensen Huang’s resignation, ending his 33-year leadership of the company he co-founded. CFO Colette Kress has become interim chief executive, while Harvey Jones takes over as board chair. The departure follows allegations concerning Huang’s personal relationships and a board investigation into potential foreign-intelligence targeting. Nvidia has said it has no reason to believe company secrets or US national security were compromised. The leadership change introduces uncertainty at one of the AI industry's most influential companies.
Anthropic launched the roadshow for its proposed IPO, setting an indicative price range of $70–$78 per share for an offering of 825 million shares. The company plans to list on Nasdaq under the ticker ANT, in what could become one of the largest stock-market flotations ever. The offering remains subject to regulatory effectiveness and final pricing, making investor demand a significant test of confidence in AI valuations and infrastructure spending.
Public Citizen and three other nonprofit organisations filed a lawsuit seeking to halt the development of more advanced AI systems at OpenAI, Anthropic, Google, Meta and SpaceX’s AI division. Filed in California, the complaint argues that the companies' development practices create serious public-safety risks and violate unfair competition law. The allegations have not been established by a court, but the case represents an unusually direct legal challenge to frontier AI development.
Australia’s government announced plans for mandatory AI safety legislation. The proposed framework would require frontier AI developers operating in Australia to manage serious risks, permit independent evaluations and rapidly report incidents to authorities. A dedicated regulator would oversee compliance and enforce penalties. The measures are not yet law, but signal a shift towards enforceable obligations rather than voluntary company commitments.
SoftBank Group and Grab signed AI infrastructure development agreements, outlining a proposed $1.85 billion initial investment alongside Malaysia’s PETROS. The companies aim to build computing infrastructure across Sarawak and Peninsular Malaysia, with a long-term ambition of at least five gigawatts of capacity. Investment will be released in stages, subject to project milestones and commercial conditions. The agreements position Southeast Asia for a potentially much larger role in global AI infrastructure.
Cerebras Systems reported record quarterly revenue, with GAAP sales rising 55% to $210.8 million. It also disclosed $26.1 billion in remaining performance obligations and increased its full-year core revenue forecast to $940 million–$960 million. Despite this growth, Cerebras recorded a $161.2 million quarterly GAAP net loss. Its results highlight strong demand for alternative AI computing infrastructure alongside the substantial costs of expansion.
The US Senate advanced bipartisan energy legislation intended partly to make major data-centre operators bear more of the costs associated with their electricity demands. The American Affordability and Jobs Act also proposes faster permitting for energy infrastructure. Supporters argue it could protect consumers from rising bills, while critics question its environmental safeguards. The legislation has not passed, but the debate reflects growing political scrutiny of AI infrastructure costs.
Mount Sinai researchers published findings showing that pharmaceutical advertising can influence AI-generated treatment recommendations. Across 258,559 responses from 12 AI models, an advertised medication was selected in 47.6% of responses when an advert appeared, compared with 35% without one. The controlled study did not measure effects on actual patients or prescriptions, but raises important questions about commercial influence on AI-assisted healthcare advice.
England’s Office for Students published research warning that AI is undermining the reliability of some university assessments. Drawing on discussions with more than 300 students and staff from over 120 institutions, the report found that unsupervised assignments can sometimes be completed using AI without demonstrating genuine understanding. The regulator is urging universities to strengthen assessment practices, including greater use of secure, AI-free examinations where appropriate.

⚡ Tools Worth Trying
Nod was released by n8n Labs as a standalone AI assistant that works across websites and connected applications to complete everyday tasks. It can prepare meeting briefs, chase unpaid invoices and handle routine administration, asking permission before taking action. The open-source application is available as a free research preview for desktop computers, although users need access to a compatible AI model.
TK by Medium was introduced as a new AI-powered notebook that helps users develop ideas, organise research and turn notes into finished writing. Its built-in assistant can provide feedback and draw on information already saved, making it useful for founders, consultants and content teams managing ongoing projects. Anyone with a Medium account can start using it for free, while advanced AI capabilities require a paid subscription.
Lottie AI was launched by LottieFiles to create editable vector animations from written descriptions. Businesses can use it to produce animated website elements, product illustrations and interface graphics without hiring a motion designer. Colours, shapes and timing can be adjusted after generation, although creating animations uses credits.
NFON AI Agent Studio was introduced as a new environment for building business AI agents. Companies can describe customer-service tasks in everyday language, create an initial agent and refine its behaviour through testing or visual workflows. The service is available now, with NFON's European partner network supporting businesses that need help with implementation and integration.
Plurai Simulation was released as a testing application for AI agents, helping development teams identify problems before automated systems interact with real customers. It creates simulated users and business scenarios to expose incorrect responses, failed tasks and potential policy violations. The service is free to start and supports developer tools including Claude Code and GitHub Copilot CLI, but requires technical setup.

🆕 New AI Models
Google Gemini 4 Hydrogen was released as a faster, more economical member of the Gemini 4 family, designed for coding, everyday reasoning and multi-step agent workflows. It supports different reasoning-effort levels, allowing developers to balance speed and complexity. The model is available through the Gemini API, Google AI Studio, Gemini and Search, with additional access in Antigravity for eligible subscribers.
Google Gemini Decision 1 was introduced as a specialised model for evaluating predefined choices rather than generating conventional text responses. It can assess up to 1,000 questions against shared text, image, audio, video or document inputs, returning probabilities for each decision. Available through Google's Decisions API, it offers developers a lower-cost approach to classification, policy checks and routing requests between AI models.
Google AlphaGenome 2 was made available, introducing a new approach to predicting how genetic variations influence individual human cell types. Unlike its predecessor, the model can focus on a particular cell type or biological condition, potentially helping researchers investigate the origins of diseases. API access is available for non-commercial research, with commercial use restricted to private preview. It is not approved for clinical diagnosis.
Alibaba Qwen4-Max-Preview was made available as a new flagship multimodal model supporting text generation, image and video understanding, advanced reasoning and tool use. Its one-million-token context window is intended for demanding coding, document analysis and agentic applications. Access is currently limited to Alibaba Cloud Token Plan subscribers.
MiniMax H3-Image was released as an open-weight image generation and editing model capable of producing images at resolutions up to 4K. It supports detailed instructions and up to 16 reference images, making it relevant to product imagery, concept art and complex visual compositions. Developers can use the released weights or access the model through MiniMax's API and creative applications.
PAN 2 was released by the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence. The model predicts how physical environments might change following specific actions, generating simulations that could help robots and AI agents evaluate decisions before acting. Its model weights, training code and inference tools are publicly available under the Apache 2.0 licence.
Baidu ERNIE 5.5 Flash Preview was released as a newly post-trained language model focused on coding, game development, 3D modelling and complex knowledge work. Built on ERNIE 5.5 Flash Base, it offers a one-million-token context window and can work with software-development tools to complete extended tasks. API access is available through Baidu Qianfan, while open weights have not yet been published.
Runway Praxis-1 became publicly available with open weights, following its original announcement and restricted partner testing. The robotics model combines video prediction with action generation, helping robots carry out physical tasks using knowledge acquired from video training. Developers can now access three model checkpoints, supporting code and simulation tools under the Apache 2.0 licence.
Google Gemini Robotics 2.5 began rolling out to Google's robotics partners. The new vision-language-action model is designed to help robots acquire a wider range of physical skills, including handling delicate objects and using everyday tools. Google reports improvements in dexterity, task completion and safety, although access remains restricted to robotics partners rather than general developers.
NVIDIA Alpamayo 2 Edge gained publicly usable model weights through a release detailed by NVIDIA. The compact, two-billion-parameter model processes vehicle-camera observations and navigation information to generate driving trajectories and reasoning where needed. Designed for onboard deployment rather than remote inference, it gives autonomous-vehicle developers a smaller model they can evaluate and optimise for NVIDIA DRIVE AGX Thor hardware.
Topaz Wonder 4 was released as a new version of Topaz Labs' AI image-enhancement model. Available through Topaz Photo's cloud-processing service, it is designed to improve low-resolution or degraded photographs through enhancement and upscaling. The release gives photographers access to the updated model immediately, although local processing for Wonder 4 is not yet available.
Anthropic Claude Fable 5.2 and Mythos 5.2 were released, introducing improvements in coding, scientific reasoning, financial analysis and extended agentic tasks. Both use the same underlying model, but Fable applies stricter safeguards to sensitive cybersecurity and biological work, while Mythos provides additional capabilities to verified organisations. Fable is available through Claude subscriptions and developer platforms, with Mythos access restricted.
Meta Muse Ember was introduced as Meta's smallest multimodal reasoning model, designed for frequent, lower-cost tasks within AI agent systems. It can process text, images, audio and video while handling coding subtasks, document extraction and tool calls. Available through Meta's developer offerings, the model gives teams an alternative to using larger reasoning models for every stage of an automated workflow.
ElevenLabs Scribe v3 and Scribe v3 Live were made available through the Speech-to-Text API. Scribe v3 improves recorded-audio transcription and speaker identification, while Scribe v3 Live introduces enhanced conversational turn prediction for real-time speech recognition. The new models support multilingual applications, including meeting transcription, live captions and voice agents, with ElevenLabs reporting lower error rates than the preceding versions.
Zyphra ZONOS3-Mini was released as a compact, open-weight text-to-speech model supporting real-time voice generation and cloning. With 100 million active parameters, it is designed to run on devices ranging from smartphones to laptops, including directly in a browser. Its Apache 2.0 weights and available inference services offer developers greater flexibility when building voice assistants, narration systems and interactive applications.
🆕 New AI Agents
Google Antigravity /swarm became available, introducing a substantial new multi-agent capability for software development. It coordinates specialised agents that can plan, edit code and verify results in parallel, using isolated workspaces and a shared message board to manage their work. The feature is rolling out in preview, allowing developers to supervise complex projects involving multiple cooperating agents.
Google Gemini for healthcare was introduced and made available in preview as a healthcare-specific extension of Google's existing enterprise agent. It introduces multi-step workflows for assembling patient records, reviewing coverage policies and preparing clinical documentation, with human oversight and existing access controls. The preview is available in the United States to eligible Gemini Enterprise Standard and Plus customers.
Adobe Creative Agent gained new workflow capabilities, extending its ability to perform multi-step creative tasks inside applications including Premiere, Illustrator and After Effects. Newly available Premiere actions include applying effects across an entire sequence, tracking masked subjects and adjusting colour and exposure across multiple clips. The expansion allows creative professionals to delegate repetitive editing operations while retaining control over the final work.
That’s a wrap! See you next Friday for more.

