// curated_library
Seminal papers, key talks, and essential videos from the people shaping AI — filtered for signal, not noise.
Machines of Loving Grace
Amodei's long-form vision of how transformative AI could compress decades of scientific progress into a few years — covering biology, neuroscience, mental health, economic development and more. Essential reading on AI's potential upside.
On the Importance of AI Safety Research
Anthropic's CEO on why safety and capability research must proceed together, not in sequence. The clearest articulation of the 'race to the top' philosophy from inside the frontier lab.
Claude's Model Spec (Constitutional AI)
The specification document that defines how Claude is trained to be helpful, harmless, and honest. A remarkable piece of AI alignment engineering made public.
Planning for AGI and Beyond
OpenAI's CEO lays out the organisation's transition plan for when AGI arrives — governance structures, iterative deployment, and why 'slow is impossible'. Candid and worth reading critically.
The Intelligence Age
Altman's argument that we're entering a new era of human prosperity driven by AI tools accessible to everyone — optimistic, controversial, and impossible to ignore.
Intro to Large Language Models (lecture)
One of the clearest technical explanations of LLMs ever recorded. Karpathy walks from tokenisation through attention to RLHF. Mandatory viewing for anyone working in this space.
Let's Build GPT from Scratch
Karpathy builds a GPT implementation in ~2 hours of live coding. The single best way to understand what a transformer actually is. Code is clear, explanations are exceptional.
The Spelled-Out Intro to Neural Networks & Backprop
Micrograd: building autograd from scratch with pure Python. The deepest, most rigorous treatment of backpropagation I've seen in video format. Essential before you touch PyTorch.
Software 2.0
Karpathy's prescient argument that neural networks are not just a tool but a new programming paradigm — one where the programmer specifies desired behaviour, not explicit logic. Reads even better today.
The Forward-Forward Algorithm (2022)
Hinton's alternative to backpropagation — training neural networks using only local, positive and negative data without ever computing a global gradient. Potentially significant for neuromorphic and on-device learning.
Geoffrey Hinton — Turing Award Lecture
Hinton's lecture on the development of deep learning, from the early days of backpropagation through the ImageNet moment. A first-hand account of the long AI winter and what it took to survive it.
Will Digital Intelligence Replace Biological Intelligence?
Hinton's talk following his departure from Google — the 'Godfather of AI' explains why he changed his mind on AI risk, and what he now believes we should be worried about.
Attention Is All You Need
The paper that changed everything. The transformer architecture introduced here is the foundation of every major language model in existence. If you work in AI, you must have read this.
Scaling Laws for Neural Language Models
The empirical laws governing how model performance scales with compute, data, and parameters. The theoretical backbone behind the scaling hypothesis that has driven frontier AI for five years.
Constitutional AI: Harmlessness from AI Feedback
The research paper behind Claude's training methodology. Describes how a set of principles ('constitution') guides RLHF without requiring human labellers for every harmful output.
Deep Learning (textbook)
The canonical deep learning textbook — freely available online. Chapter 6 onwards remains the most rigorous treatment of feedforward nets, regularisation, and optimisation available.