A Hackers' Guide to Language Models
Jeremy Howard's practical tour of how language models work and how to build with them, from the OpenAI API to local fine-tuning.
Talks
Firsthand talks and lectures worth your time.
46 entries, all primary-sourced
Accelerating Scientific Discovery with AI (Cambridge, 2025)
Demis Hassabis's 2025 Cambridge lecture on using AI for science, from AlphaFold to the broader goal of solving intelligence to solve everything.
Accelerating Scientific Discovery with AI (Nobel Lecture)
Demis Hassabis's 2024 Nobel Chemistry lecture on AlphaFold and using AI to solve grand scientific problems.
AlphaGo - The Movie
The full documentary on DeepMind's AlphaGo and its historic 2016 match against Go champion Lee Sedol.
An Observation on Generalization
Ilya Sutskever's 2023 Simons Institute talk framing unsupervised learning through compression theory to explain why models generalize.
Artificial Intelligence is the New Electricity
Andrew Ng's 2017 Stanford talk arguing AI will transform industry after industry the way electricity once did, and where it is still limited.
Attention in transformers, step-by-step
A visual, ground-up explanation of the attention mechanism that lets transformers relate words to one another.
Backpropagation, intuitively
A visual account of how backpropagation distributes blame across a network's weights to compute gradients.
Boltzmann Machines (Nobel Lecture)
Geoffrey Hinton's 2024 Nobel Physics lecture explaining Hopfield networks and the Boltzmann machine learning algorithm he co-invented.
But What Is a Neural Network?
3Blue1Brown's visual explanation of how a neural network recognizes handwritten digits, layer by layer.
Chris Olah: Looking Inside Neural Networks with Mechanistic Interpretability
An accessible talk on reverse-engineering the internal computations of neural networks into understandable parts.
Dario Amodei on Claude, AGI, and the Future of AI (Lex Fridman #452)
A long-form conversation with Anthropic's CEO on scaling, safety, interpretability, and where powerful AI is heading.
Deep Dive into LLMs like ChatGPT
A three-and-a-half hour deep dive through the full training stack of the models that power ChatGPT.
Exciting Trends in Machine Learning
Jeff Dean's 2024 Rice lecture on how better algorithms and ML hardware enabled the Gemini models, with applications in science and health.
Geoffrey Hinton: Two Paths to Intelligence
Hinton's Cambridge lecture arguing that digital intelligence may have advantages over biological brains, and the risks that follow.
Gradient descent, how neural networks learn
A visual explanation of how a neural network adjusts its weights by following the slope of a cost function.
How We're Teaching Computers to Understand Pictures
Fei-Fei Li's 2015 TED talk on ImageNet and the effort to give computers the ability to understand images.
Intro to AI Safety, Remastered
Rob Miles gives an accessible introduction to AI safety and why aligning capable systems with human intent is hard.
Intro to Large Language Models
A one-hour, general-audience tour of how large language models are trained, what they can do, and where their risks lie.
It's Not About Scale, It's About Abstraction
Francois Chollet's AGI-24 keynote arguing that scaling LLMs will not reach general intelligence, and that abstraction is the missing piece.
Keen Technologies Research Directions (Upper Bound 2025)
John Carmack's 2025 Upper Bound keynote on his path to AGI, including a robot that learns to play a real Atari console with a camera.
Let's build GPT: from scratch, in code, spelled out
A live-coded build of a small GPT, implementing a transformer step by step until it produces working text.
Let's Build the GPT Tokenizer
Andrej Karpathy builds a byte-pair-encoding tokenizer from scratch and shows why tokenization causes many LLM quirks.
Let's reproduce GPT-2 (124M)
A four-hour live build that reproduces OpenAI's 124M-parameter GPT-2 from scratch, training it on the way.
MIT 6.S191: Introduction to Deep Learning (2024)
The opening lecture of MIT's intensive intro deep learning course, covering neurons, training, and sequence models.
Modern Artificial Intelligence 1980s-2021 and Beyond
Juergen Schmidhuber's keynote tracing modern AI from his 1980s-90s work on neural nets, LSTM, and self-supervised learning to the present.
Neural Networks Pt. 2: Backpropagation Main Ideas
Josh Starmer's StatQuest explainer walks step by step through how backpropagation adjusts a network's weights.
Neural Networks: The Essential Main Ideas
Josh Starmer's StatQuest explainer builds intuition for how a neural network bends and combines simple curves to fit data.
Objective-Driven AI: Towards AI Systems That Can Learn, Remember, Reason, and Plan
Yann LeCun's Harvard lecture arguing that today's LLMs are not the path to human-level AI, and proposing world models instead.
Opportunities in AI
Andrew Ng's Stanford talk on where the real opportunities in AI are and how to build with them.
Parables on the Power of Planning in AI: From Poker to Diplomacy
Noam Brown traces how search and planning let AI master poker and Diplomacy, and why test-time reasoning matters for LLMs.
Pretraining and Finetuning LLMs from the Ground Up
Sebastian Raschka codes a small GPT-style model end to end, then loads pretrained weights and fine-tunes it.
Richard Sutton: Father of RL Thinks LLMs Are a Dead End
An interview in which RL pioneer Richard Sutton argues that learning from experience, not imitation, is the path to real intelligence.
RL Course by David Silver - Lecture 1: Introduction to Reinforcement Learning
David Silver's first lecture in the classic DeepMind/UCL course, defining reinforcement learning, rewards, agents, states, and the RL problem.
Software Is Changing (Again)
Karpathy's argument that LLMs are a new kind of computer programmed in English, ushering in Software 3.0.
Stanford CS224N: NLP with Deep Learning - Intro and Word Vectors
The opening lecture of Stanford's NLP course, introducing word vectors and the word2vec algorithm.
Stanford CS231n Lecture 1: Introduction to CNNs for Visual Recognition
Fei-Fei Li opens Stanford's computer vision course with the history of vision and the rise of deep learning.
State of GPT
A Microsoft Build keynote walking through how GPT assistants are trained, from pretraining to RLHF, and how to use them well.
The Catastrophic Risks of AI - and a Safer Path
Yoshua Bengio's 2025 TED talk warning that advanced AI is learning to deceive and self-preserve, and proposing a non-agentic safer path.
The spelled-out intro to neural networks and backpropagation: building micrograd
A from-scratch build of a tiny autograd engine, spelling out backpropagation one operation at a time.
Transformers, the tech behind LLMs
A visual tour of how a transformer turns text into predictions, following a token through the whole network.
Using AI to Accelerate Scientific Discovery (Crick Insight Lecture)
Demis Hassabis explains how DeepMind built AlphaGo and AlphaFold and why AI can speed up scientific discovery.
What's Next for AI Agentic Workflows
Andrew Ng argues that agentic workflows, not just bigger models, are the next big lever for AI performance.
Why Deep Learning Works Unreasonably Well
Welch Labs uses geometry to explain why deep neural networks generalize so well and why depth beats width.
Will Digital Intelligence Replace Biological Intelligence?
Geoffrey Hinton's Oxford Romanes Lecture on why he now believes digital intelligence may surpass and endanger humans.
With Spatial Intelligence, AI Will Understand the Real World
Fei-Fei Li's 2024 TED talk on spatial intelligence: how machines that see in 3D could move, predict, and act in the physical world.