Frontier Research
Models, research, safety and the ideas pushing the frontier of AI
Human Feedback
The alignment trick that turned a rambling model into ChatGPT
A 1.3-billion-parameter model trained on human preferences was judged more helpful than GPT-3, a model a hundred times its size, and the technique behind it now shapes almost every deployed chatbot
7 min read
Scaling Laws
The study that proved most big AI models were trained wrong
DeepMind showed that the giant language models of the day were far too large for the data they were fed, and that a model less than half GPT-3's size, better nourished, could beat it
7 min read
Generative Models
How AI learned to make art by reversing pure noise
A 2020 paper showed that if you slowly turn an image into random static, then train a network to undo each step, it can conjure new images out of noise, the principle now behind Stable Diffusion and DALL-E
8 min read