Pinned
New paper: Tell Me What To Learn: Generalizing Neural Memory to be Controllable in Natural Language.
We built a neural memory system that lets you tell an AI — in natural language — what to remember and what to ignore over time.
🧠👇
Max S. Bennett
142 posts
Max S. Bennett
@maxsbennett
- This book by @anilananth truly is a masterpiece. One of my fav parts of my day has been gobbling up a chapter or two each night
- 1/ My book "A Brief History of Intelligence" is now available! I tell the 600 million year story of how human intelligence evolved and its relationship to progress in AI. More details in thread below👇
- Replying to @maxsbennett2/ The basic idea is that as a reasonable first approximation, the timeline of human brain evolution can be viewed as the result of five consecutive evolutionary breakthroughs, each of which show fascinating parallels to the recent and upcoming developments in AI.
- Replying to @maxsbennett12/ I don't have a good answer. I didn't do it to achieve anything - once I got interested, it was too fascinating to stop. I wrote the book I wanted to read.
- Replying to @maxsbennett6/ To make sure I was doing my homework, before writing the book, I published two papers on the topic in different neuroscience journals. Here is the first paper that outlines the framework for approximating brain evolution: frontiersin.org/articles/10.33…
- Random speculation: bird DVR and mammal neocortex seem to perform similar tasks but with clearly different architectures (DVR is nuclear, neocortex is layered). Bird and dinosaur brains always remained comparatively quite small, whereas many mammal brains have gotten quite big -
- Convergent evolution of intelligence between of birds and mammals: - both are warm blooded (while common ancestor was not warm blooded) - both show evidence of mental simulation, planning, episodic-like memory, causal reasoning (while most nonmammal/non-avian vertebrates do
- Replying to @maxsbennett3/ The book explores why the first brain evolved and the origin of learning and affective states. It explores the evolution of TD learning in early vertebrates, the great repurposing of dopamine, and the first pattern recognition techniques.
- Replying to @maxsbennett5/ The story in the book is certainly inexact and simplified. And it is possible much of it will turn out wrong as new evidence emerges. But it was wonderfully fascinating and satisfying to piece together the story of our brains as best could be done given the available evidence.
- Replying to @maxsbennett7/ Here is the second paper, where I outline the methodology for inferring intellectual abilities in our ancestors: frontiersin.org/articles/10.33…
- Sometimes I feel overwhelmed by gratitude for how supportive and generous even distant humans can be of each other.
- Replying to @maxsbennett4/ It uncovers the evolution of generative models and model-based RL in early mammals. And it investigates the relationship between mechanisms for theory of mind in primates and AI techniques for imitation learning (along with many other things), and continues onward to humans
- Replying to @maxsbennett8/ I am deeply grateful to the many many scientists who helped me along the way, just some of them: J Ledoux (@theamygdaloid), Karl Friston, David Redish, Dileep George (@dileeplearning), Jeff Hawkins, Eva Jablonka, Kent Berridge, and many others.