AiPulse — Issue No. 58

The UN High Commissioner for Human Rights called on countries and leading artificial intelligence companies to govern advanced models within the framework of international law. This call was made alongside a warning that artificial intelligence poses unprecedented existential risks to humanity.

The new method allows models to explore freely during generation, shaping only the final output according to strict security and physical rules. As a result, this yields contradiction-free, higher-quality outcomes in fields such as robotics and computer vision.

OpenAI significantly shortened the design timeline for its first AI accelerator, Jalapeño, by leveraging its own large language models (LLMs). This approach accelerates the role of artificial intelligence in chip design, setting a new standard for the industry.

The company will use this fund to develop next-generation GLM foundation models, self-training systems, and large-scale inference infrastructure. The plan also includes domestic chip compatibility and long-range reasoning capabilities.

The AI-based security tool Aisle discovered six previously unreported security vulnerabilities in the Curl project, which were verified with CVE numbers. This achievement was positively received by the project's founder Daniel Stenberg and Linux kernel maintainer Greg Kroah-Hartman.

Leaders at companies like Anthropic and OpenAI are calling for a slowdown in the pace of artificial intelligence. The article suggests that behind this stance lies increasing debt and physical bottlenecks in data center construction.

Apple has released the iOS 27, iPadOS 27, watchOS 27, visionOS 27, and macOS 27 Golden Gate updates. These versions include the new AI-powered Siri AI assistant and improvements to the Liquid Glass design language.

The report emphasizes that Europe must make a 100 billion euro data center investment to reduce its dependence on the US and China. It warns that otherwise, the continent will be irreversibly locked into foreign powers.

In Vals AI's tests, Claude Fable 5.1 decrypted the 'Cyphral Distich' cipher, created by Thomas Urquhart in 1653 and long considered unsolvable, in 44 minutes. This achievement demonstrates the capacity of artificial intelligence to solve complex historical cryptography problems.

NTT achieved fiber-optic-level data transfer over a distance of 100 meters using OAM-MIMO technology. This development aims to increase wireless connection capacity in AI data centers and mobile networks.

China's Ministry of Industry and Information Technology has published an action plan aiming to integrate critical software with artificial intelligence by 2030. The plan includes concrete targets such as implementation in 20,000 software companies and 100 intelligent agent software projects by 2028.

The research showed that a single dense electrocorticography implant enabled the simultaneous decoding of speech and gestures in individuals with paralysis. This finding is considered an important step toward more natural and multi-modal brain-computer interfaces.

You.com founder Richard Socher aims to automate AI's own research processes and accelerate scientific discovery through his venture, Recursive. In line with this vision, the company has completed a $4.65 billion seed round.

The company has prepared a comprehensive draft code of conduct to prevent models from resisting human control or setting their own goals. Following a six-week public consultation process, this draft will serve as the fundamental reference in the model development process starting in 2027.

A team led by Georgia Tech aims to develop a system using MEFA logic circuits that exploit mechanical stress to consume 100 times less energy than current CMOS technology. This $10.6 million DARPA project is designed to increase the energy efficiency of future computer chips.