Artificial Intelligence News Live: Salesforce researchers unveil MCPEval, an open-source tool for testing AI agents

Salesforce researchers have unveiled MCPEval, a new open-source framework designed to evaluate the performance and tool-use capabilities of AI agents, VentureBeat says in a report. The method facilitates "protocol-level" testing, which analyzes an agent's low-level interactions within Minecraft Protocol (MCP) servers, providing a more granular and robust assessment than observing high-level results alone. By offering a standardized and reproducible "plug-and-play" environment, MCPEval aims to accelerate research and improve the quality assurance process for developing more reliable autonomous AI agents.

Microsoft is rolling out several new artificial intelligence features for Windows 11, with some tools available only on new Copilot+ PCs, according to a report in The Verge. A key update for all users is Copilot Vision, which can analyze on-screen content across open applications and webpages to answer questions. For users with Snapdragon-powered Copilot+ PCs, new features include an AI agent in the Settings app for natural language commands, an AI sticker generator in Paint, an AI lighting option in the Photos app, and an AI-powered capture tool in the Snipping Tool. Microsoft has stated that these updates will be gradually released over the next month.

OpenAI has reportedly agreed to pay Oracle $30 billion annually for access to its cloud infrastructure and data center services, according to a report in TechCrunch. The deal, previously announced by Oracle in March 2024 without naming the customer, provides OpenAI with significant high-performance computing capacity, including specialized GPU clusters essential for training and running large-scale AI models. This arrangement marks a major expansion of OpenAI's computing resources beyond its primary partnership with Microsoft Azure and highlights the intense, industry-wide competition for AI infrastructure.

Apple has reintroduced its AI-powered notification summary feature for news and entertainment apps in the fourth developer beta of iOS 18, according to a report in The Verge. The feature was temporarily disabled earlier this year after it generated inaccurate headlines. With its return, Apple now includes a user-facing warning that "summarization may change the meaning of the original headlines" and advises users to verify information. The company also notes that the tool is a beta feature and that the AI-generated "summaries may contain errors."

A new analysis by the Pew Research Center has found that the integration of AI Overviews into Google Search results leads to a substantial drop in traffic to websites, Ars Technica says in a report. The findings confirm widespread concerns among publishers and content creators who rely on search referral traffic for revenue. The research highlights the direct impact of generative AI features, which are designed to provide synthesized answers directly on the search page, thereby reducing the user's need to click on traditional organic links. This trend is part of a broader industry shift, with other search platforms also implementing AI-driven summaries, raising questions about the future of the open web ecosystem and content-driven business models.

OpenAI is actively seeking additional funding from both new and existing investors to fulfill a recently announced financing round, according to a report in Wired. The effort is part of a tender offer that values the artificial intelligence research and deployment company at $40 billion. This financial arrangement is designed to facilitate the purchase of existing shares, allowing current stakeholders like employees to sell a portion of their holdings.

OpenAI CEO Sam Altman warned the financial industry of a “significant impending fraud crisis” due to the ability of artificial intelligence to impersonate voices and bypass security checks to move money, according to a report in AP. Speaking at a Federal Reserve conference in Washington on Tuesday, Altman expressed alarm that some financial institutions still use voiceprints as a form of authentication, a practice he called "a crazy thing to still be doing" given the advancements in AI voice cloning technology.

Amazon is acquiring Bee, a startup that produces an AI-powered wearable device that listens to and transcribes conversations to provide users with personalized summaries and suggestions, according to a report in The Verge. Bee CEO Maria de Lourdes Zollo stated the company is joining Amazon to help "bring truly personal, agentic AI to even more customers." The device can also access a user's emails, location, and calendar to inform its AI insights, though it has reportedly shown performance issues by confusing real-life conversations with audio from media. In response to privacy questions, Amazon spokesperson Alexandra Miller said the company "cares deeply" about customer privacy and security, adding that the deal is not yet closed and its terms are confidential.

Microsoft has hired more than 20 artificial intelligence staff from its competitor Google DeepMind in a move to strengthen its position in the rapidly developing technology, according to a report in the Financial Times. The recruitment highlights the intense competition among major tech companies for a limited pool of highly skilled researchers and engineers. This action is part of a broader effort by Microsoft to bolster its in-house AI capabilities and gain an edge over rivals.

The user privacy application Cloaked has released a new feature that uses an AI-powered bot to engage with and screen potential scam callers on a user's behalf, according to a report in TechCrunch. The new 'call guard' function is designed to go beyond typical call blocking by having the AI actively converse with the incoming caller to determine the call's legitimacy. This development utilizes advancements in conversational AI to protect users from directly interacting with fraudulent or unsolicited calls.

A fire at a data center in Oregon used by X originated in a power cabinet, an incident that highlights the growing fire risks associated with the surging electricity demands of generative AI, according to a report in Wired. The event underscores the significant operational challenges facing data centers as they accommodate the high power density required for artificial intelligence workloads. The intense energy consumption and heat generated by AI systems place considerable stress on electrical infrastructure, increasing the likelihood of component failures and electrical fires.

The city of Virginia Beach, Va., is now using an artificial intelligence tool to handle over 40 percent of its non-emergency calls, according to a report in GovTech. The city implemented Amazon Connect, which uses AI and natural language processing, to modernize public safety communications and free up dispatchers for emergency calls. Jada Lee, the city's director for Emergency Communications and Citizen Services, stated that the system has saved public safety dispatchers over 900 hours of continuous talk time annually, allowing them to focus on genuine emergencies.

Some employees at xAI balked at a request to participate in an internal training program for the AI model Grok, as per a report in Ars Technica. Based on internal documents and Slack messages, the training required employees to act out personal or potentially uncomfortable scenarios to generate data for the AI's conversational and personality development. Management reportedly described the goal metaphorically as an effort to "give Grok a face," meaning to build its interactive, human-like persona. The pushback from some staff indicates internal disagreement over the methods being used to train the company's artificial intelligence.

New York-based venture capital firm Betaworks has closed its third fund at $66 million, which will be used for investments in early-stage artificial intelligence companies, according to a report in TechCrunch. The firm has a history of investing in the AI sector, dating back to at least 2016. Past investments made through its program, Camp, include AI companies such as Hugging Face and Granola.

Startup Mixus has developed a new AI agent platform designed to improve usability and reliability by integrating directly into existing business communication tools, according to a report in TechCrunch. The system allows users to interact with and manage AI agents from their email or Slack clients, facilitating a "human-in-the-loop" workflow. This approach ensures that humans remain involved to review, approve, or modify tasks, addressing common issues with fully autonomous agents such as errors and maintaining context. The goal is to lower the barrier for enterprise adoption by embedding AI assistants into familiar applications rather than requiring users to learn new, complex interfaces.

Technology companies are intensifying efforts to reduce fabricated responses, or 'hallucinations,' from AI models, but complete elimination appears impossible with current architectures, the Financial Times says in a report. These hallucinations are described as technical failures where a model generates plausible but factually incorrect information due to its probabilistic nature, rather than a failure to retrieve a known fact. While mitigation strategies like Retrieval-Augmented Generation (RAG) and advanced alignment techniques are increasingly used to ground models in verified data and reduce errors, the issue remains a core research challenge. The persistence of these fabrications necessitates human oversight in critical applications, requires users to verify AI-generated information, and poses a risk to the broader information ecosystem.

Multinational corporations including McDonald's and Bupa are increasingly establishing Global Capability Centres (GCCs) in India to recruit AI engineers for advanced big-data operations, according to a report in the Financial Times. This move addresses a global shortage of specialized AI talent, leveraging India's large pool of science and technology graduates. These centers are evolving from traditional support offices into strategic innovation hubs responsible for core functions like research and development, product development, and complex data science. The trend, accelerated by the demand for generative AI expertise, is creating high-value jobs in India and shifting the country's role from an IT service provider to a center for advanced technological development.

Google DeepMind is researching methods for robots to learn complex skills autonomously using table tennis as a testbed, according to a report in IEEE Spectrum. The research explores several approaches to move beyond traditional machine learning techniques that require significant human oversight. One method involves two robots learning by competing against each other, creating a cycle of escalating skill. Another approach trained a robot against human players, where it achieved amateur-level performance by winning against beginners and about half of its matches against intermediate players. Additionally, the team is investigating the use of a vision language model (VLM) to act as an AI coach, which observes the robot's performance and provides feedback for improvement without needing a manually engineered reward function. The project leverages zero-shot sim-to-real transfer, allowing policies trained in simulation to be deployed directly on physical robots.

The Lee Kuan Yew School of Public Policy (LKYSPP) on Tuesday released a comprehensive research report detailing strategies for ASEAN to leverage the convergence of 5G and artificial intelligence for economic development, AP said in a report. The study, titled "Leveraging 5G to Accelerate AI-Driven Transformation in ASEAN: Imperatives, Policy Insights, and Recommendations," is designed to provide policymakers in the region with actionable strategies to advance their digital economies.

Artificial intelligence is being used to help food companies predict successful flavors for new products, according to a report in Food Business News. Steve Brown, an AI expert and former executive at Google DeepMind and Intel, noted that more advanced applications now combine knowledge of flavor chemistry, consumer data, and machine learning to forecast detailed and complex results. This technology, which uses deep learning on diverse datasets, could be applied to products like soups and yogurts where consistent flavor is critical for the consumer experience.

South Korean AI chip startup FuriosaAI has partnered with LG AI Research to supply its AI chip, RNGD, to enterprises using the EXAONE platform, according to a report in TechCrunch. The partnership, announced Tuesday, will see FuriosaAI's chip, which is optimized for running large language models (LLMs), integrated as a hardware option for clients of LG's AI ecosystem. This development follows LG AI Research's recent unveiling of its next-generation hybrid AI model, EXAONE 4.0.

AI safety and research company Anthropic is shifting its funding strategy to pursue investments from Gulf states, according to a report in Wired that obtained a leaked internal memo. In the note to staff, CEO Dario Amodei acknowledged the change, stating that the principle of ensuring "'No bad person should ever benefit from our success' is a pretty difficult principle to run a business on." This decision marks a notable change for the company, which has previously emphasized its commitment to responsible AI development, and reflects the immense capital required to compete in the development of advanced artificial intelligence models.

A new automotive optical network system has been launched that exceeds 10 Gb/s transmission speeds, providing the high-bandwidth backbone required for real-time processing by AI in autonomous vehicles, according to a report in IEEE Spectrum. Developed by semiconductor firm KD and camera maker Leopard Imaging, the system uses optical fiber to handle massive data streams from high-resolution sensors, a critical requirement for advanced driver-assistance systems (ADAS) and fully autonomous driving. This shift to an optical backbone reduces vehicle weight, eliminates electromagnetic interference, and provides the low-latency data transmission needed for AI systems to safely control a vehicle. The technology is designed with durable, automotive-grade components and is future-proof, allowing for upgrades to higher speeds on the same fiber infrastructure.

OpenAI and the new UK government led by Prime Minister Keir Starmer have entered into a strategic partnership, according to a report in the Financial Times. The agreement is described as a co-operation pledge intended to spur an investment push into the country's artificial intelligence sector. A potential outcome of the collaboration is the development and implementation of AI-powered digital public services.

An AI-driven surgical robot has performed an experimental surgery on pig organs, according to a report in Ars Technica. The procedure was considered successful, indicating the robot met predefined metrics for the task in an experimental setting. This test on animal organs is a key research step for developing autonomous systems for potential future medical applications.

Experts are warning that US courts are highly likely to overlook errors from artificial intelligence, a failure that could threaten the authority of the judiciary, according to a report in Ars Technica. The concern stems from the increasing use of AI in legal proceedings and the potential for judges, without proper training, to miss AI-generated 'hallucinations,' such as fabricated case citations or biased information. Such oversights could lead to unjust rulings, successful appeals, and an erosion of public trust in the legal system, prompting calls for mandatory AI literacy programs for judges and legal professionals to mitigate these risks.

A recent study reveals that 72% of teenagers in the United States have experimented with AI companions, according to a report in TechCrunch. The findings also show significant ongoing engagement, with 52% of the teens surveyed reporting that they use these AI companions on a regular basis. The report highlights a high adoption rate of conversational AI among the youth demographic.

OpenAI is reportedly preparing to launch a 'router' feature that will automatically select the most appropriate AI model for a user's specific task, according to a report in VentureBeat. The anticipated system is designed to address the challenge users face with an increasing number of AI models by analyzing a prompt and routing it to the most suitable and cost-effective model available through OpenAI's API, such as GPT-4o or GPT-3.5. This functionality, described as imminent, would operate at the API level to intelligently manage requests, balancing performance needs with cost, rather than being limited to the consumer ChatGPT interface.