NadiAIGlobal AI news. Concise. Clear. Easy to follow.
xAI, the company behind Elon Musk’s Grok, has filed a lawsuit against a Grok user accused of creating and distributing AI-generated sexual deepfakes. The move comes amid broader legal battles targeting similar deepfake harms involving xAI.
OpenAI has developed GPT-Red, an internal ‘LLM super-hacker’ used as a sparring partner to probe and improve the safety of its models. Separately, the US is seeing increased adoption of heat pumps, reflecting shifts in home heating technology and policy incentives.
OpenAI describes GPT-Red, an automated red teaming method that uses self-play to identify vulnerabilities and bolster model safety and alignment. The system focuses on improving resistance to prompt injection and other robustness failures through iterative testing.
Anthropic released an advertisement suggesting AI could pose existential risks to humanity, sparking debate over the company's marketing approach. Futurism reported the ad as a provocative move that some view as sensationalist rather than informative.
DeepMind CEO Demis Hassabis has called for a FINRA-like independent standards body to test frontier AI models and establish release best practices. The proposal aims to create external oversight and technical evaluations separate from developers and governments, per TechCrunch AI's report.
Anthropic reports a new technique that reveals aspects of its models’ internal reasoning, offering a glimpse into how answers are formed. The piece examines what this finding demonstrates and cautions about its limits for understanding broader ‘world models.’
A Futurism piece warns that AI technologies are increasingly attractive to violent extremists and terrorists, raising concerns about harmful deployment. The article suggests the industry bears responsibility for enabling misuse if safeguards remain insufficient.
Reporting notes the details of federal discussions with OpenAI and Anthropic over safety assessments remain unclear. TechCrunch AI indicates some dialogue occurred but provides limited specifics on criteria or process.
OpenAI menerbitkan analisis mengenai serangan adversarial yang mensasarkan polisi rangkaian neural. Analisis itu menggariskan potensi kesan terhadap ketahanan dan keselamatan sistem autonomi serta perlunya strategi mitigasi.
OpenAI bersama penyelidik Google memperkenalkan activation atlases, teknik baru untuk memvisualkan interaksi antara neuron. Alat ini bertujuan membantu memahami proses keputusan dalaman model supaya kelemahan dan kegagalan dapat dikenalpasti dan disiasat.
OpenAI menala semula GPT-2 (774M parameter) menggunakan maklum balas manusia untuk pelbagai tugas dan berjaya menepati pilihan pelabel luaran walaupun ia kadangkala berbeza dengan pasukan mereka. Untuk ringkasan diperlukan 60,000 label kerana pelabel cenderung memilih ayat yang disalin terus dari input, manakala tugas sambungan teks gaya ringkas memerlukan kira-kira 5,000 label.
Menurut laporan The Star yang disiarkan oleh NAIO AI News, ada usaha mendorong keselamatan AI di kalangan negara ASEAN berikutan ketidaktentuan global. Laporan itu menekankan keperluan kerjasama serantau dan dasar bersama untuk mengurangkan risiko teknologi.