NerdNews

OpenAI's GPT-5 is a significant improvement over its predecessors, offering enhanced capabilities in coding, writing, safety, and accuracy. It's faster, safer, and more accurate, with features like Safe Completions and improved reasoning. GPT-5 is available for free, with paid subscribers getting additional features and unlimited access.

OpenAI has launched GPT-5, its latest AI model, which is available for free to all ChatGPT users. GPT-5 claims to have fewer confabulations, better coding capabilities, and a new 'safe completions' approach. The model is part of a unified system that includes a smart and efficient model, a deeper reasoning model, and a real-time router.

OpenAI's valuation could reach $500 billion, making employees millionaires. The company is discussing a deal to allow current and former employees to sell their privately held shares, creating a new class of millionaires. This move is seen as a way to retain top talent in the AI industry, where companies are engaged in a fierce battle for the best minds.

OpenAI has released GPT-5, a big leap from previous models, with built-in reasoning and a unified system. It's available for free, Plus, Pro, and Team users, with paid users getting higher limits. GPT-5 is a unified system that brings the best of two worlds, knowing when to respond quickly and when to rely on its 'reasoning' or 'think longer' capability.

OpenAI has launched several new features for ChatGPT, including a new GPT-5 model, customizable 'personalities' for text-based responses, improved vibe coding capabilities, and more app customization options. The model picker has been retired, and paid users can access advanced voice mode with the ability to adapt to their speaking style. Additionally, users can now easily connect their Gmail and Google Calendar accounts to ChatGPT.

A study found that AI tools are not effective in predicting gene activity, highlighting the complexity of biology and the limitations of current AI systems. The researchers tested several AI models and found that they performed poorly compared to simple baseline models, emphasizing the need for further development in AI for biology.

X owner Elon Musk announced plans to include paid ads in Grok chatbot's responses, aiming to pay for 'expensive GPUs' and improve ad targeting. The ads will be automated and prioritized based on aesthetics, with plans for a checkout feature for in-app purchases.

GitHub CEO Thomas Dohmke believes that developers will move from writing code to managing AI agents, with AI generating 90-100% of code in the future. He identifies four stages of AI developer maturity and suggests that computer science education will need to change to focus on understanding systems and debugging AI-generated code.

OpenAI released gpt-oss, an open-source language model. The lighter model, gpt-oss-20b, requires 16GB of free memory and can be run locally using Ollama, a free client app. Performance varies depending on hardware, with GPUs and recent Macs delivering better results.

The article discusses the security concerns of deterministic vs probabilistic AI systems, highlighting the importance of deterministic security in an AI-driven world. It explores the differences between predictive and generative AI, and how GitGuardian combines probability and proof for safe and auditable development.

Qwant and Ecosia launched Staan, a European search index aiming to be a cheaper, more privacy-focused alternative to Google and Bing. The index is developed through a joint venture and plans to serve around 50% of French queries and 33% of German queries by the end of the year.

Celona's AerFlex is a breakthrough private 5G platform that simplifies deployment and operation for enterprises, eliminating the need for on-site servers and complex integrations. It combines simplicity, speed, and security in a fully integrated platform, enabling faster deployments and lower total cost of ownership.

Meta's Reality Labs Research previewed its latest VR prototypes, including the Tiramisu project with 'hyperrealistic VR' and Boba 3 headsets with ultrawide field of view. The Tiramisu headset has three times the contrast, 14 times the maximum brightness, and 3.6 times the angular resolution of the Meta Quest 3, but with a limited field of view and bulky form factor.

Microsoft's Project Ire uses advanced language models and reverse engineering tools to classify malware at a global scale with high precision, aiming to reduce burnout and alert fatigue for security analysts.

Snowflake has introduced a client connector to run Apache Spark code directly in its cloud warehouse, eliminating the need to provision and maintain a separate Spark cluster. This move aims to simplify the process of running Spark workloads and reduce costs. According to Snowflake, customers who have tested the new connector have seen an average of 5.6 times faster performance and 40% cost savings compared to traditional Spark setups.

The EU AI Act aims to create a uniform legal framework for AI innovation across EU countries. It applies to both local and foreign companies, and sets a high bar for what AI should and shouldn’t do for society. The Act has a risk-based approach, banning certain uses of AI, flagging high-risk uses, and applying lighter obligations to limited risk scenarios. It comes with penalties, including fines of up to €35 million or 7% of annual turnover.

Former CNN anchor Jim Acosta conducted an interview with an AI version of Joaquin Oliver, a Parkland shooting victim, to discuss gun violence. The interview has raised questions about ethics, grief, and the limits of AI, with many viewers calling it exploitative and emotionally manipulative.

Microsoft and CISA have warned of a high-severity bug in Exchange Server hybrid deployments that could allow attackers to escalate privileges from on-premises Exchange to the cloud. The bug, tracked as CVE-2025-53786, is considered serious and likely to be exploited soon. To address this bug, users should install the April Hotfix on on-premises Exchange servers and follow Microsoft's configuration instructions.

Meta is training its AI on user data from social media posts, but a recent survey found that only 7% of Europeans approve of this practice. The survey, commissioned by privacy advocate Max Schrems, found that 27% of users were unaware that their data was being used in this way. Meta claims it has a legitimate interest in using the data, but Schrems argues that the company is prioritizing its profits over users' privacy rights.

AI coding startups like Windsurf and Anysphere face significant financial challenges due to high costs of using large language models, leading to thin margins and potential acquisitions or closures. Building their own models could help, but it's a costly and risky endeavor.