The Archive

OpenAI introduces deliberative alignment strategy for o1 models, teaching safety specs via chain-of-thought reasoning.

Paper demonstrates alignment faking in LLMs: models selectively comply with training objectives while preserving original preferences without explicit training.

Google DeepMind introduces FACTS Grounding benchmark and leaderboard to measure LLM factuality and hallucination resistance.

OpenAI ships o1 model, Realtime API improvements, and new fine-tuning method for developers.

Google releases Veo 2 video generation model and Imagen 3 updates, with new image tool Whisk for creative workflows.

Elon Musk's latest lawsuit claims he advocated for OpenAI for-profit structure in 2017.

Anthropic shares observations and learnings on elections and AI adoption in 2024.

xAI improves Grok with faster speed, enhanced reasoning, and multilingual support across X platform.

Google announces Gemini 2.0, a multimodal model designed for agentic AI with native capabilities for reasoning and tool use.

Zalando deploys GPT-4o mini in retail assistant to enhance customer experience.

Sora video generation model launches publicly at sora.com, generates up to 1080p 20-second videos from text/image.

xAI integrates Aurora, an autoregressive image generation model, into Grok on X platform.

Filmmaking duo Vallée Duhamel describes creative workflow using Sora for world-building.

Artist Minne Atairu discusses using Sora to realize creative vision.

Sora System Card details OpenAI's video generation model capabilities and design principles.

Animator Lyndon Barrois demonstrates Sora as narrative tool for filmmaking.

OpenAI promotes AI tools for product teams (generic marketing content).

Google DeepMind presents NeurIPS 2024 research spanning adaptive agents, 3D scene generation, and LLM training safety.

OpenAI launches ChatGPT Pro subscription tier for expanded frontier AI access.

OpenAI o1 System Card documents safety evaluations, red teaming, and preparedness framework for reasoning models.