NerdNews

Google has released SpeciesNet, an AI model designed to identify animal species from camera trap images. The model, trained on 65 million images, can classify images into over 2,000 labels and is available on GitHub under an Apache 2.0 license.

Recent AI large language model releases, such as Claude 3.7 and Grok 3, demonstrate significant advancements in AI capabilities, with potential applications in scientific research, problem-solving, and decision-making. These models have been trained using increased computing power and are showing emergent capabilities, such as anticipating user needs and considering novel angles in problem-solving. Experts believe that these advancements could lead to a radical transformation in various fields, including science, and could potentially unlock a near-utopian future with significant economic growth and improvements in quality of life.

Sesame's new AI voice model features uncanny imperfections, sparking amazement and discomfort online. The model achieves near-human quality speech, but raises concerns about deception and fraud. Users report forming emotional connections with the AI, with some expressing astonishment and others feeling unsettled.

Microsoft launched Azure AI Foundry Labs, a hub for the latest AI research and experiments. It offers assets and projects like Aurora, MatterSim, and TamGen, enabling developers to create prototypes and explore new possibilities. The labs also feature ExACT, Magentic-One, and OmniParser v2, which enhance agent capabilities and action generation.

Google Cloud has announced the public beta launch of Gen AI Toolbox for Databases, an open-source server developed in collaboration with LangChain. This tool simplifies database connections for AI applications, supporting databases such as PostgreSQL and MySQL, and improves performance, security, and developer experience.

Google has integrated its Data Science Agent into Colab, a cloud-based notebook tool, to help users clean data, visualize trends, and gain insights. The agent uses Google's Gemini 2.0 AI model and can analyze CSV, JSON, or .txt files under 1GB in size. It can also help with feature engineering, data-cleaning tasks, and writing SQL code.

This article presents a comprehensive guide to designing a secure AI-powered early detection system for medical data analysis and diagnosis, ensuring that AI models can leverage the rich information in HL7 and FHIR data while maintaining strict privacy standards. The proposed architecture consists of eight interconnected layers, each addressing specific aspects of privacy-preserving AI in healthcare.

A Columbia University student, Roy Lee, created an AI program called Interview Coder to help him pass technical interviews at big tech companies like Amazon, Meta, and TikTok. He used the program to pass the interviews and received job offers, but was later reported to his university for cheating. Lee argues that the technical interview process is outdated and that AI will make most programming jobs obsolete in two years.

Hugging Face has published a guide on efficient LLM training across GPUs, covering parallelism strategies, memory management, and benchmarking insights. The guide provides practical guidance for researchers and engineers, including techniques such as data parallelism, tensor parallelism, and pipeline parallelism, as well as memory optimization methods like activation recomputation and gradient accumulation.

Matt Fleming discusses the challenges of benchmarking and detecting performance changes in noisy results. He covers the sources of noise, measures of dispersion, and change point detection algorithms to identify performance changes. Fleming also shares anecdotes and real-life examples from his experience with open-source projects.

California state Senator Scott Wiener introduces a new AI bill, SB 53, to protect employees at leading AI labs and create a public cloud computing cluster, CalCompute, to develop AI that benefits the public. The bill aims to prevent critical risks from AI systems and provide whistleblower protections.

Auxia, an AI startup, has raised $23.5M to help large enterprises effectively use their customer data. The company's platform employs 'agentic AI' to analyze customer data, make decisions, and deliver personalized experiences across channels. Auxia aims to solve the 'reacquisition treadmill' problem, where companies repeatedly pay to attract the same customers instead of building lasting relationships.

Broadcom has released patches for three critical VMware hypervisor-hijacking bugs, including a 9.3-rated heap overflow vulnerability. The bugs, tracked as CVE-2025-22224, CVE-2025-22225, and CVE-2025-22226, affect various VMware products and can be exploited to gain control of the hypervisor and host system. Updating to a fixed version will plug the holes.

CoreWeave, a GPU cloud provider, has filed for an IPO, aiming to capitalize on the growth of AI. The company has seen a 737% increase in revenue, with 77% of its revenue coming from just two customers, including Microsoft. However, this heavy reliance on a few customers poses a risk to the company's future performance.

ByteDance has launched Trae, a new AI-powered code editor that offers unlimited free access to DeepSeek R1 and Claude 3.7 Sonnet large language models. Trae has two primary modes: Builder mode for zero-to-one application development and Chat mode for analyzing existing code and providing suggestions. The tool is available on both Macs and PCs and supports Visual Studio Code plug-ins.

Private 5G networks are being rapidly deployed, but organizations lack communications technology expertise, leaving them vulnerable to cyber risks. Despite adopting AI security tools, 90% of users face difficulties in deploying the technology, and only 20% have dedicated CT security teams. Security budgets allocated to private 5G networks are also insufficient, with only 18% of budgets dedicated to these infrastructures.

A new botnet malware named 'Eleven11bot' has infected over 86,000 IoT devices, primarily security cameras and network video recorders, to conduct DDoS attacks. The botnet is loosely linked to Iran and has already launched attacks targeting telecommunication service providers and online gaming servers.

Google's latest Pixel Feature Drop includes new features such as generating images of people in Pixel Studio, improved scam detection in Google Messages, and new games in Android Auto. The update also brings satellite messaging to Pixel 9 series users on the Verizon network and enhances location sharing in the Find My Device app.

Research shows Google uses advertising cookies and identifiers to track Android users without consent, potentially violating data protection laws. Google's response claims user privacy is a top priority, but the company does not agree with the legal analysis.

Apple is taking the UK government to court over efforts to force the company to weaken iCloud encryption. The UK government wants Apple to create a backdoor for law enforcement access, but Apple refuses, citing concerns over user privacy and security. The case has drawn international attention and criticism from the US government.