AI Research Hub
Curated breakthroughs in artificial intelligence, machine learning, and AI safety — updated weekly.
Scaling Laws for Neural Language Models
Empirical investigation of how loss scales with model size, compute, and dataset size. Key findings inform the training compute-optimal approach now used in all frontier models.
Constitutional AI: Harmlessness from AI Feedback
Anthropic introduces a method for training a harmless AI assistant without human labels on harms, leveraging AI-generated critiques and revisions grounded in a set of principles.
GPT-4 Technical Report
OpenAI's multimodal model achieving human-level performance on professional exams. Describes safety mitigations, evaluation methodology, and training process at scale.
Gemini: A Family of Highly Capable Multimodal Models
Google DeepMind presents Gemini Ultra, Pro, and Nano — models natively multimodal from pretraining on text, image, audio, and video data across 32K+ token context windows.
Apple Intelligence: On-Device and Server Foundation Models
Technical description of Apple's foundation model stack: a 3B on-device model and larger server models, trained on carefully curated licensed and synthetic data with a focus on privacy-preserving inference.
LLaMA 3: Herd of Models
Meta releases open-weight models from 8B to 405B parameters with 15T training tokens, multilingual capability, long context, and tool use — setting new open-source benchmarks.
Data Provenance and Copyright in AI Training: Legal and Technical Survey
Comprehensive survey of 40+ AI companies' data sourcing practices, robots.txt compliance, licensing terms, and the emerging legal landscape around training data.