Archive - From AI to ZI

The Hidden Scratchpad

[Epistemic Status: Short, off-the-cuff post to build intuition. No claims to originality.] When evaluating language models for safety properties, we…

Apr 1 •

March 2024

Research Report: Sparse Autoencoders find only 9/180 board state features in OthelloGPT

Abstract A sparse autoencoder is a neural network architecture that has recently gained popularity as a technique to find interpretable features in…

Mar 5 •

January 2024

Why Transformers Are Good

200 words on why the Transformer architecture is good

Jan 19 •

December 2023

Rating my AI Predictions

9 months ago I predicted trends I expected to see in AI over the course of 2023. With the year coming to a close, let’s rate how I did! Predictions…

Dec 21, 2023 •

The Future of From AI To ZI

For the past few months, the blog’s regular readers may have noticed its irregular writing. This is a short post to explain what I’ve been up to instead…

Dec 11, 2023 •

November 2023

Twelve Months in AI Safety

The AI safety landscape has been totally transformed since I started this blog twelve months ago. In this post I want to recap some of the major events…

Nov 1, 2023 •

July 2023

Unsafe AI as Dynamical Systems

[Thanks to Valerie Morris for help editing this post.] Overview Large Language Models (LLMs) and their safety properties are often studied from the…

Jul 14, 2023 •

AIs teams will probably be more superintelligent than individual AIs

Summary Teams of humans (countries, corporations, governments, etc) are more powerful and intelligent than individual humans. Our prior should be the…

Jul 4, 2023 •

June 2023

[Research Update] Sparse Autoencoder features are bimodal

Overview The sparse autoencoders project is a mechanistic interpretability effort to algorithmically find semantically meaningful “features” in a…

Jun 22, 2023 •

Explaining "Taking features out of superposition with sparse autoencoders"

[Thanks to Logan Riggs and Hoagy for their help writing this post.] In this post, I’m going to translate the post [Interim research report] Taking…

Jun 16, 2023 •

May 2023

Is behavioral safety "solved" in non-adversarial conditions?

I’m trying to crystalize something I was said to a friend recently: I think that techniques like RLHF and Constitutional AI seem to be sufficient for…

May 25, 2023 •

Statistics for the Working Mathematician

I’m a mathematician, but I managed to avoid almost all education about statistics, and I have found that statistics resources are often catered to a…

May 25, 2023 •

#nojs-banner { position: fixed; bottom: 0; left: 0; padding: 16px 16px 16px 32px; width: 100%; box-sizing: border-box; background: red; color: white; font-family: -apple-system, "Segoe UI", Roboto, Helvetica, Arial, sans-serif, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol"; font-size: 13px; line-height: 13px; } #nojs-banner a { color: inherit; text-decoration: underline; } This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts