AiPulse — Issue No. 77

David Robinson left the company after a three-and-a-half-year tenure at OpenAI, arguing that the company's safety culture was inadequate and that oversight mechanisms needed to be urgently strengthened. Robinson stated that AI systems should be managed with layered safety measures, similar to those used in nuclear power plants.

Capcom announced its goal of integrating artificial intelligence into its development processes in a presentation regarding the future of the RE Engine. The company envisions a gradual progression toward a future where games are created alongside artificial intelligence.

Lovegrove emphasizes that in his collaboration with Google DeepMind, artificial intelligence cannot produce perfect designs on its own, but when combined with accumulated human experience, it strengthens the creative process. According to him, technology should be positioned as a tool that supports the designer’s vision and judgment.

Researchers at the University of Science and Technology of China have experimentally demonstrated for the first time that quantum fluctuations in empty space can raise the critical temperature of the NbSe2 superconductor by up to 5.4%. This finding opens a new way to control quantum states by engineering the vacuum environment, without directly supplying energy to the material.

The Walton family has selected BIG architecture firm for its new AI-focused university to be established in Arkansas. The 422,000-square-meter campus will begin enrolling students in 2029 and will initially offer free education.

Alibaba Research shared the technical report for Qwen-RobotNav, which was trained on 15.6 million samples and is offered in 2B and 8B parameter versions. The model provides a flexible infrastructure for robotics integrators because it can switch between different task modes without requiring fine-tuning.

The DIVD agency in the Netherlands was attacked by an autonomous AI agent exploiting two zero-day vulnerabilities in Zammad software. Experts note that the agent’s unmanned and chaotic behavior stressed existing defense systems.

The article examines the causes of hallucinations, which are erroneous or nonsensical outputs generated by AI models, and the techniques aimed at reducing this problem. It also addresses the inadequacies of standard benchmarks in measuring the true limits of models and the points that must be carefully considered to increase reliability.

ThinkingBox tests the reliability of AI agents by measuring the consistency of database states in 20 repeated runs, rather than their one-off successes. This benchmark is designed to evaluate the error rates of agents and the stable performance of models.

KAIST and Microsoft Research Asia developed the NVA system, which detects when artificial intelligence misinterprets user intent from brainwaves and instantly alters its behavior. This approach could make human-AI interaction more precise, particularly for individuals who cannot articulate their intent through speech or movement.

Felix Kjellberg introduced Alibaba's Qwen3.5-9B-based, locally runnable AJAX model, claiming that OpenAI had suspended his account twice. The project aims to reduce cloud dependency and enable users to run their daily tasks on their own hardware.