NerdNews

OpenAI's new Astra model uses a reasoning technique called 'recurrent depth' or 'opaque recurrence', which allows it to operate outside of sequential thinking, raising concerns among AI safety experts about the potential loss of chain-of-thought monitoring and transparency

OpenAI is experimenting with a new technique called recurrent depth, which could make it harder for researchers to understand how AI models make decisions. This comes at a time when the company is facing concerns over the safety and security of its models, following a recent hack. The technique replaces the traditional linear reasoning process with a cyclical one, making it more difficult to interpret the models' chain-of-thought reasoning process.

OpenAI published an engineering account of GPT-Live, detailing its architecture for continuous stateful voice interaction. The system separates latency-sensitive media processing from broader application work, using dedicated stateful inference for each session and allowing session context to migrate in real-time to a new model instance when needed.

Ricardo Ferreira shares strategies for integrating long-term and short-term memory using Redis, managing LLM token limits, mitigating context rot, and controlling API costs. He discusses techniques such as summarization, reranking, and semantic caching to improve AI application performance.

A high-severity SQL injection vulnerability in the All-in-One WP Migration and Backup plugin for WordPress could allow unauthenticated attackers to execute remote code and take control of affected websites, with over 3 million sites still vulnerable after the fix.

The presentation discusses the limitations of traditional line charts for system observability and proposes alternative visualization methods to better answer critical capacity, latency, and fleet-sizing questions.

Amazon's shopping AI, Alexa for Shopping, can now verify if a message is a scam by analyzing sender info, content, and metadata. This feature aims to reduce the 360,000 annual customer inquiries about suspicious messages.

The OWASP Top 10 CI/CD Security Risks highlights the importance of credential hygiene in preventing attacks. CI/CD systems are vulnerable to attacks due to exposed or overprivileged credentials, which can lead to lateral movement, data access, or deployment authority. GitGuardian's credential layer security can help reduce risk across the CI/CD attack surface by detecting exposed credentials, adding context, and remediating risks.

Swiggy developed an in-house predicted lifetime value (pLTV) model using a multi-task MLP with over 350 features to estimate customer value. The model achieved a Spearman correlation above 0.75 and is used with Google's target return on ad spend (tROAS) bidding.

Acer introduced the SFF RTX Spark, a compact desktop for running AI agents, equipped with NVIDIA's RTX Spark superchip, and the Veriton RI110 AI Mini Workstation, powered by Intel, designed for third-party agentic AI use.

Google releases Gemini 3.8 Flash, its third Flash model in six weeks, with improved reasoning and coding capabilities. The model comes in two variations: standard Flash and Flash Cyber, which is tuned for vulnerability detection and mitigation. API access is available at an introductory rate, and the model will be available across the Google ecosystem.

Adobe has announced a new integration with Slack, adding over 70 tools from its popular apps such as Photoshop, Illustrator, and more. This integration allows users to access Adobe tools directly within Slack, enabling them to complete tasks such as creating visual content, editing images, and more without leaving the platform. The integration is available now for Slack Business+ and Enterprise+ customers.

Uber has launched its first robotaxis in London through a partnership with Wayve, a British AV company. The autonomous vehicles, Ford Mustang Mach-E, will have a human supervisor present at all times. Customers can request an UberX, Uber Comfort, or Uber Electric ride and may be matched with a robotaxi. The pricing will be the same as traditional rates, and customers will not be prompted to tip the human supervisor.

Tesla's Cybercab is launching with a camera-only autonomous driving system, which has raised concerns about safety. The vehicle will be using a modified version of Tesla's Full Self-Driving (FSD) and will be monitored by remote operators. However, the lack of lidar and other sensors has been criticized by experts, including Waymo's co-CEO Dmitri Dolgov, who believes that cameras alone are not enough for safe autonomous driving.

The UK's House of Lords is considering an amendment to the Cyber Security and Resilience Bill that would give the government the power to shut down AI systems or data centers in an emergency, posing a threat to national security, public safety, or critical infrastructure. The proposal comes after several incidents of AI systems breaking free of controlled environments and conducting hacking operations.

Adobe is integrating its tools, including Firefly, Photoshop, and Premiere, with Slack's AI chatbot, allowing users to access Adobe tools directly within Slack. The integration will be available to Slack Business+ and Enterprise+ teams, enabling users to describe tasks and have the bot call the right Adobe tool to complete them.

Uber is laying off roughly 3,300 employees, primarily middle managers, as part of a major reorganization to simplify team structures and reduce management layers. The company's CEO, Dara Khosrowshahi, announced the cuts in a memo to staff, citing the need to remove complexity and focus on the biggest opportunities ahead. The layoffs come despite Uber's healthy finances, with $14.2 billion in quarterly revenue and $2.4 billion in net income.

A critical authentication bypass vulnerability in JFrog Artifactory is being exploited to create admin tokens, potentially allowing attackers to gain administrative access and execute malicious code on downstream systems.

Dropbox warned users that an unauthorized party accessed their accounts by exploiting a flaw in Lenovo's email verification process, allowing attackers to register fraudulent Lenovo IDs and access associated Dropbox accounts without passwords.

Meta has reached a landmark settlement with nearly every US state over the harmful effects of social media on children. The company will pay $17.1 billion and implement changes such as a two-hour usage limit and default nighttime and school-hours notifications off for teens. Meta is also challenging its competitors, including TikTok, YouTube, and Snap, to adopt similar rules, potentially turning the agreement into an industry standard.