A couple of resources I've gathered from BlueDot Impact's course on AI safety
will keep updating as I come across more content Check out the course here
The AI landscape is evolving rapidly as new safety techniques emerge, coordination becomes more or less feasible, and AI capabilities advance, so reevaluating your stance is important. Many view these as stages rather than competing strategies. For example: pausing now to develop safety techniques, then proceeding with controlled development. Or: racing to establish a lead that gives leverage to enforce global safety standards.
What is AI alignment Specification Gaming: How AI Can Turn Your Wishes Against You Why AI alignment could be hard with modern deep learning Frontier Models Are Reward Hacking Multi-Agent Risks from Advanced AI What failure looks like Why Would AI "Aim" To Defeat Humanity?
What is input data filtration in AI safety? Deep Ignorance Enhancing Model Safety through Pretraining Data Filtering A small number of samples can poison LLMs of any size
The True Story of How GPT-2 Became Maximally Lewd A simple technical explanation of RLH(AI)F An Approach to Technical AGI Safety and Security What is deliberative alignment? Detecting and reducing scheming in AI models What is AI debate, and can it make systems safer? Julian Michael – Empirical Progress on Debate debating with More Persuasive LLMs Leads to More Truthful Answers On scalable oversight with weak LLMs judgingstrong LLMs
Anthropic’s Responsible Scaling Policy: Version 3.0 System Card: Claude Mythos Preview
OpenAI Preparedness Framework GPT-5 System Card
DeepMind's Frontier Safety Framework Gemini 3 Pro Model Card
Meta's Approach to Frontier AI Llama Guard 4 Model Card
Introduction to Mechanistic Interpretability Neel Nanda on the race to read AI minds (part 1) MoSSAIC: AI Safety After Mechanism Let's Try To Understand AI Monosemanticity Interpretability Will Not Reliably Find Deceptive AIby Neel Nanda AGI Safety | Connor Leahy | FLI Interpretability Conference, MIT | March 2023 Auditing language models for hidden objectives Chain of Thought Monitorability:A New and Fragile Opportunity for AI Safety