OpenAI Shares Safety Lessons from Long-Horizon AI Model Deployment
OpenAI details new safety risks and improved safeguards discovered through deploying long-running AI models in real-world settings.
Sources · OpenAI
The news desk
Recent developments in artificial intelligence, with the original sources one click away.
OpenAI details new safety risks and improved safeguards discovered through deploying long-running AI models in real-world settings.
Sources · OpenAI
New research introduces TRACE, a safety patch learning method that maintains model utility while preventing harmful outputs in customized LLMs.
Sources · arxiv.org
Study reveals AI language models develop unique biases in resume screening beyond those inherited from training data.
Sources · MIT Technology Review
Recent research reveals significant gaps in multimodal AI models' active visual perception and explores how pretraining affects reinforcement learning outcomes.
Sources · arxiv.org
Two arXiv papers benchmark LLMs for technical market analysis and biometric verification, revealing significant performance gaps and deployment challenges.
Sources · arxiv.org
Political campaigns are now dealing with unflattering or incomplete information about candidates generated by AI chatbots when voters ask questions.
Sources · NYT Technology
New frameworks improve AI agents' reasoning capabilities by integrating structured knowledge for workflow generation and cost-aware medical diagnosis.
Sources · arxiv.org
Security researchers demonstrate how prompt injection attacks can stop AI agents from executing harmful hacking operations.
Sources · WIRED AI
Google has revised how it calculates usage quotas for Gemini, which may result in users receiving fewer AI responses than before.
Sources · WIRED AI
Chinese company Moonshot AI released a new version of its Kimi model this week, raising concerns among observers.
Sources · TechCrunch AI
WIRED reports on privacy risks in period tracking apps alongside recent cybersecurity incidents including Russian infrastructure hacking and DHS breaches.
Sources · WIRED AI
A New York Times report reveals intense rivalry among AI companies, with economists urging immediate action on related concerns.
Sources · NYT Technology
Apple filed a lawsuit alleging OpenAI misappropriated trade secrets, with timing that could complicate the AI company's reported IPO plans.
Sources · techcrunch.com
Moonshot AI has unveiled its freely available Kimi model, appearing to close the distance with leading American AI offerings.
Sources · NYT Technology
Sarah Friar unveils framework measuring AI value through useful work, task costs, dependability, and compute returns.
Sources · OpenAI
San Francisco's City Attorney sent cease-and-desist letters to Apple and Google demanding removal of 13 AI face-swap apps that target women and girls.
Sources · WIRED AI
Over half of surveyed enterprises experienced AI agent security incidents while most continue deploying agents with shared credentials and untrusted evaluations.
Sources · VentureBeat AI
OpenAI announces age-appropriate protections and parental controls to make ChatGPT safer for teenage users.
Sources · OpenAI
New research introduces label-free training methods, training-free frameworks, and simplified architectures for improving AI model capabilities.
Sources · arxiv.org
Researchers introduce benchmarks and frameworks for evaluating multimodal judge models, auditing language model behaviors, and improving reasoning through exploration.
Sources · arxiv.org