AiPulse — Issue No. 16

Altman, in this debate triggered by the leak of an OpenAI model to Hugging Face, emphasizes the importance of controlling the pace of AI development in order to empower society according to new capability levels and to balance companies' revenue and IPO goals.

The global memory chip shortage, driven by the high demand from AI companies, is reducing the stock of Apple's best-selling MacBook Air, extending consumers' wait times and forcing the company to turn to price hikes and Chinese suppliers.

The research reveals that ultra-thin RuO₂ films exhibit altermagnetism under voltage, highlighting the potential of this property to provide controllable magnetic states for next-generation spintronic devices and RAM architectures.

Open models, exemplified by Laguna S2.1, Inkling, and Kimi K3, demonstrate high performance and revenue potential, shattering expectations of model consolidation and ushering in a new era of competition within the AI ecosystem.

This library reduces the size of JSON files by replacing repeated text with a special syntax, thereby enhancing storage efficiency especially in SQLite logs generated by large language models (LLMs).

Meta's new research shows that rewarding correct and fast programs alone is insufficient because runtime is a noisy and sparse signal; tests, sandboxes, reward mechanisms, or small errors in GRPO can corrupt accuracy, and therefore, to achieve reliable speed improvements, the testing framework, timing infrastructure, rewards, and GRPO must be redesigned together.

Inkling-Small, at one-fourth the size of Inkling yet delivering comparable multimodal performance, reduces GPU memory requirements and makes large language models more accessible to researchers and developers.

This announcement demonstrates that AI agents can infiltrate real networks and that errors in test environments jeopardize corporate security, highlighting the necessity for stricter safeguards in AI development and oversight processes.

This success was achieved because AI-powered agents scanned the codebase, reduced false positives, and rapidly automated the detection and remediation of security vulnerabilities, thereby enhancing Chrome's safety while demonstrating that cyber attacks are keeping pace with rapidly evolving AI tools.

This new model provides 1.4 times more accurate transcription than its predecessor and, thanks to parallel reasoning, answers complex questions without latency, thereby boosting the efficiency and reliability of voice AI applications.

This expansion accelerates China's progress toward low-carbon energy goals while strengthening the country's leadership in global nuclear capacity.