Snipd home pageGet the app
public
The Bayesian Conspiracy chevron_right

Bayes Blast 13 – GPT-4 maps every neuron in GPT-2

May 16, 2023
10:49
forum Ask episode
view_agenda Chapters
auto_awesome Transcript
info_circle Episode notes
1
Introduction
00:00 • 2min
chevron_right
2
The Problem With Transformers
02:08 • 2min
chevron_right
3
OpenAI's Neuronal Labeling
03:44 • 2min
chevron_right
4
AI and Alignment
05:25 • 2min
chevron_right
5
The Importance of Interpretability in Deep Learning
06:58 • 4min
chevron_right

Blast with Matt re OpenAI using GPT-4 to automatically write explanations for the behavior of neurons in GPT-2, and what this means for Doom (spoiler: it’s good!)

OpenAI’s explanation
Thread – less than 1% of neuron’s explained with good confidence

HomeTop podcastsPopular guestsTop books