Yazılar

Anthropic CEO Dario Amodei Claims AI Models Experience Fewer Hallucinations Than Humans: Report

/in /tarafından

Anthropic CEO Dario Amodei recently stated that artificial intelligence (AI) models tend to hallucinate less frequently than humans do. This remark was made during the company’s first-ever Code With Claude event, held on Thursday. At this event, the San Francisco-based AI firm unveiled two new versions of its Claude 4 models, alongside several upgraded features such as enhanced memory and better tool integration. Amodei also commented on the skepticism surrounding AI development, suggesting that despite critics searching for obstacles, no significant barriers to AI progress have emerged so far.

During a press briefing reported by TechCrunch, Amodei elaborated on the nature of hallucinations in AI systems, explaining that these errors do not prevent AI from achieving artificial general intelligence (AGI). When asked about hallucinations, he said, “It really depends how you measure it, but I suspect that AI models probably hallucinate less than humans, but they hallucinate in more surprising ways.” This perspective highlights that while AI does make mistakes, the frequency might be lower than commonly assumed, though the mistakes can sometimes be unexpected.

Amodei also pointed out that errors are a common part of human activity, with TV presenters, politicians, and professionals making mistakes regularly. Therefore, the presence of errors in AI responses does not necessarily undermine its overall intelligence. Nonetheless, he acknowledged the issue of AI confidently presenting false information remains a challenge. A recent incident highlighted this when Anthropic’s lawyer had to apologize in court after the company’s Claude chatbot generated an incorrect citation in a legal filing. This mishap took place during Anthropic’s ongoing lawsuit against music publishers over alleged copyright violations related to hundreds of song lyrics.

Looking ahead, Amodei remains optimistic about the future of AI. In a paper published in October 2024, he claimed that Anthropic could achieve artificial general intelligence as soon as next year. AGI represents a breakthrough form of AI capable of understanding, learning, and performing a broad spectrum of tasks autonomously, without human assistance. If realized, this development would mark a significant milestone in AI research and its practical applications.

MediaTek Unveils Hybrid AI Computing Strategy at Computex 2025

/in /tarafından

At Computex 2025, MediaTek unveiled a series of cutting-edge artificial intelligence (AI) innovations designed to power a broad range of devices and platforms. Central to its showcase was the company’s first hybrid computing solution, aimed at enhancing AI performance at the edge by combining on-device AI processing with high-speed 5G connectivity. This approach promises to deliver faster, low-latency AI experiences while ensuring data privacy, setting the stage for more responsive and secure AI applications in smart homes and beyond.

MediaTek’s hybrid computing concept integrates its AI-powered Fixed Wireless Access (FWA) gateway with advanced AI capabilities directly on the device. Currently in the prototype phase, this solution has already undergone successful proof-of-concept trials with a leading telecommunications infrastructure provider. The company demonstrated an AI Hub that acts as a centralized control center, enabling on-device AI agents to work together and serve as personal assistants. This technology aims to allow users to control multiple smart home devices simultaneously through natural language commands, making connected living more intuitive and efficient.

Beyond hybrid computing, MediaTek revealed new AI-focused chipsets targeting various markets. One notable announcement was a new chipset designed specifically for in-car infotainment systems, bringing enhanced AI-powered features such as voice recognition, personalized user interfaces, and real-time data processing to vehicles. This expansion into automotive AI aligns with MediaTek’s broader strategy to embed AI capabilities across diverse platforms, including Internet of Things (IoT) devices, which the company also highlighted at the event.

MediaTek also showcased its collaboration with Nvidia on the GB10 Grace Blackwell Superchip, touted as the world’s smallest AI supercomputer. Designed for the DGX Spark system, this superchip can execute up to 1,000 trillion operations per second (TOPS) and run AI models containing up to 200 billion parameters locally, a significant leap for on-premise AI processing. Collectively, these innovations underscore MediaTek’s commitment to advancing AI technology at multiple levels—from edge devices and smart homes to automotive and cloud computing—driving forward its vision of a seamlessly AI-powered future.

Microsoft Unveils NLWeb Open Project to Bring AI-Powered Natural Language Interfaces to Websites

/in /tarafından

Microsoft has unveiled a new open project called NLWeb at its Build 2025 developer conference, aimed at transforming how users interact with websites. Short for Natural Language Web, NLWeb is designed to add AI-powered natural language interfaces to websites, enabling them to understand and respond to user queries just like a conversational assistant. Microsoft believes this innovation could be as foundational to the next era of the web as HTML was to the birth of websites. By embedding this AI layer, websites can evolve into more interactive and intelligent platforms.

In its announcement, Microsoft explained that NLWeb would essentially allow websites to become AI applications. Instead of relying on traditional search bars or structured navigation, users will be able to ask questions or make requests in natural language—such as “Show me the latest pricing plans” or “Summarize this article”—and receive relevant, AI-generated responses. This capability can significantly enhance user engagement and streamline access to content.

Beyond just chat interfaces, NLWeb introduces a deeper integration into the broader AI ecosystem. Each NLWeb-enabled site also acts as a Model Context Protocol (MCP) server. MCP is an open standard, originally developed by Anthropic, that allows AI agents to access external data in a structured way. By supporting MCP, websites not only serve users directly but also make their content accessible to other AI agents and services, turning static web content into dynamic data sources.

Microsoft emphasized that NLWeb is fully open-source and available to developers starting today. Hosted on GitHub and integrated with the Azure AI Foundry Labs, the platform invites developers, researchers, and website owners to experiment, contribute, and help shape the future of AI-native web experiences. As more websites adopt NLWeb, it could lead to a more intelligent, agent-compatible internet where users interact naturally with the web—just like talking to an assistant.