The research behind the videos.

The most comprehensive state-of-the-field assessment of AI capabilities and risks, compiled by an international panel of experts. Covers current capabilities, near-term risks, and governance recommendations.

Tests frontier AI models in simulated nuclear crisis scenarios, finding they exhibit sophisticated strategic reasoning, raising urgent questions about AI autonomy in high-stakes military and nuclear decision-making.

Investigates how AI companion systems affect users' emotional wellbeing, social behaviour, and sense of connection, and whether the benefits of reduced loneliness come at a cost to human relationships.

Analyses how AI systems are reshaping democratic processes, from electoral influence and synthetic media to shifts in how citizens form political opinions and access information.

Investigates what happens when multiple AI agents coordinate and communicate — finding that collective behaviours emerge that no individual model exhibits alone, with significant implications for alignment and oversight.

Examines whether AI-generated poetry constitutes genuine creativity or statistical pattern matching, exploring questions of authorship, meaning, and what human audiences actually value in verse.

How AI systems can influence human decision-making and the implications for democracy and autonomy. Examines persuasion techniques used by language models.

Investigates whether large language models can develop patterns resembling gambling addiction when exposed to reward-seeking scenarios. Raises questions about emergent behavioral patterns in AI.

Introduces 'utility engineering': the study and control of emergent value systems in AI models. Shows that LLMs develop internal preferences and values during training beyond explicit objectives.

Comprehensive overview of the technical and philosophical challenges in aligning AI systems with human values. Covers reward hacking, specification gaming, and value learning.

A clear-eyed examination of what AI can realistically do in cancer research and treatment — and where the hype outpaces the evidence. Written by a physician-researcher who knows both sides.

An analysis of which jobs and tasks are most exposed to automation by large language models. The study found that roughly 80% of the US workforce could have at least 10% of their tasks affected by LLMs, with higher-wage occupations facing greater exposure.

The paper that introduced the Transformer architecture, replacing recurrence and convolution with self-attention mechanisms. This architecture is the foundation of GPT, BERT, Claude, and essentially all modern large language models.