NerdNews

Google has unveiled Ironwood, its most powerful AI processor yet, with up to 9,216 liquid-cooled chips, delivering higher throughput and memory. Ironwood is designed for Google's Gemini models and will allow AI to act on behalf of users, enabling more powerful agentic AI capabilities.

Meta will use public-facing interactions on its platforms to train large language models, while Elon Musk and Jack Dorsey suggest abolishing IP laws to facilitate AI growth. This move raises concerns about data ownership and the potential consequences of relying on user-generated content for AI training.

Google Classroom has launched a new AI-powered feature that helps teachers generate questions based on specific text input. The feature, which uses Gemini, allows educators to upload files or enter text and generate questions that can be exported to Google Doc or Google Form. The feature is available to Google Workspace for Education subscribers with the Gemini Education add-on or Gemini Education Premium.

Meta will start training its AI models on public content from Facebook and Instagram in the EU, after previously pausing plans due to regulatory pressure. The company will use user interactions and public posts to train its models, with users able to opt out.

Nvidia has commissioned over 1 million square feet of manufacturing space to build and test AI chips in Arizona and Texas, aiming to produce up to half-a-trillion dollars of AI infrastructure in the US within the next four years. The company is partnering with TSMC, Foxconn, and Wistron to build 'supercomputer' manufacturing plants, with mass production expected to ramp up in the next 12-15 months.

Diablo Canyon, a nuclear power plant in California, is using artificial intelligence to help workers navigate technical reports and regulations. The AI tool, called Neutron Enterprise, uses NVIDIA H100 graphical processors to search and summarize documents, reducing the time spent on this task from 15,000 hours to a significantly lower amount. The technology is being developed by Atomic Canyon, a startup that partnered with the plant's operator, PG&E. While the current use of AI is limited, there are concerns about the potential risks and need for regulations as the technology becomes more integrated into the nuclear energy industry.

Ubisoft has open-sourced its colorblind assistance tool Chroma, allowing developers to simulate various types of color blindness in games and identify accessibility issues. The tool uses the Color Oracle algorithm and is available for download on GitHub.

Fitness functions provide guardrails for software architecture evolution, enabling continuous change within a desired direction. They offer objective measures, similar to unit tests, and can be used to enforce architectural patterns and styles. Libraries like ArchUnit make it feasible to write fitness functions for structural aspects of architectural fitness.

This article discusses how to think creatively with the help of Large Language Models (LLMs). It explores the work of cognitive scientist Margaret Boden, who broke down creativity into specific mental processes. The article highlights three types of creativity: combinatorial, exploratory, and transformational, and provides examples of how LLMs can support each type. It also offers practical tips on how to use LLMs to enhance creativity, such as shortening feedback loops, pushing past obvious ideas, and articulating taste.

RLWRLD, a South Korean startup, has raised $14.8M in seed funding to develop a foundational AI model for robotics, enabling robots to make quick and agile movements and perform logical reasoning. The model combines large language models with traditional robotics software, aiming to automate industrial processes and explore human-centric workflows.

A group of Chinese-speaking cybercriminals, known as the Smishing Triad, are behind a massive smishing operation, sending millions of scam text messages and stealing millions of dollars. The group uses sophisticated software and techniques to impersonate companies and brands, and has been targeting people in at least 121 countries. The scam messages often prompt people to enter personal information and bank card details, which are then used to add bank cards to digital wallets like Apple Pay or Google Wallet.

Nvidia plans to produce up to $500 billion in American-made AI supercomputer equipment over the next four years, with production of its Blackwell accelerators and systems starting in the US within 12-15 months. The company is working with TSMC and partners to achieve this goal, which includes building manufacturing facilities in Arizona and Texas.

Wix has introduced Astro, an all-in-one AI site and business management tool. Astro is a chatbot that can track visitor behavior, generate marketing content, customize site services, and build detailed reports. It will be available to all premium users at no extra cost, making it easier for creative workers and small businesses to manage their websites and digital operations.

Hugging Face, an AI dev platform, has acquired Pollen Robotics, a French robotics startup, to expand its robotics efforts. Pollen's humanoid robot, Reachy 2, will be sold and its code made available for developers to download and improve. The acquisition marks a significant move in the field of AI and robotics, with Hugging Face aiming to make AI and robots open-source.

Kodiak Robotics, a self-driving truck startup, plans to go public via a merger with Ares Acquisition Corporation II, valuing the company at $2.5 billion pre-money. The move comes as the self-driving truck space experiences significant changes, with some major players shutting down. Kodiak has driven 2.6 million miles autonomously and has pursued off-road autonomy as a quicker path to market.

Microsoft's OneDrive file sync apps for Windows and Mac have been broken for 10 months, with shared folders vanishing from local drives and being replaced with web shortcuts. Despite user complaints, Microsoft has not provided a fix, citing an ongoing backend migration as the cause of the issue.

Meta will train its AI models on data from EU users, including public posts and chat history, while giving users the option to opt out. The company claims this will help create more region-specific AI models, but has faced regulatory pushback in the past.

Apple introduces a new AI training system using differential privacy, comparing synthetic data to real-world samples on user devices without accessing user data, to improve AI models like Siri

Intel has agreed to sell a 51% stake in its Altera semiconductor business to Silver Lake, a private equity firm, in a deal valuing Altera at $8.75 billion. Intel will retain a 49% stake in Altera, which will become operationally independent. The transaction is expected to close in the second half of 2025.

Fortinet flaws are being exploited with a new method, Chinese-made robot dogs have been found with preconfigured backdoors, and OpenAI is being used to generate spam messages. Additionally, Android's monthly security update has addressed multiple critical issues, and NIST has announced it will no longer update older CVEs unless necessary.