Writing

2024

(yet another) Broken Adversarial Example Defense at IEEE S&P 2024: I broke another defense to adversarial examples by fixing 1 line of code; in this post I complain about the state of the field of adversarial robustness.

My benchmark for large language models: A benchmark of ~100 tests for language models, collected from actual questions I've asked of language models in the last year.

My Research Idea Logfile, 2016-2019: A description of how I keep track of my research ideas, with my complete log from when I started it in 2016 through to the end of 2019.



2023

Reading Data off an Apple ProFile Hard Drive with an Arduino: A short writeup of how to read data off a 1980s Apple ProFile hard drive using an Arduino.

Playing chess with large language models: I built a bot to play chess by querying a text language model. It sees the sequence of moves in order (as text!), and predicts which move comes next. It's better than me.

Little Bobby <|endoftext|>: I found a fun exploit ChatGPT that causes it to behave weirdly.

A GPT-4 Forecasting Challenge: Test your ability to predict (in a calibrated manner) whether or not GPT-4 can answer a range of questions from coding to poetry to baking.

A ChatGPT clone, in 3000 bytes of C, backed by GPT-2: A dependency-free implementation of GPT-2, including byte-pair encoding and transformer inference, in ~3000 bytes of C. I then use this to create something like Chat GPT.


2022

Reflecting on “Towards Evaluating the Robustness of Neural Networks”: A few thoughts about the paper that brought me into the field of adversarial machine learning.

Rapid Iteration in Machine Learning Research: I wrote a tool to help me quickly iterate on research ideas by snapshoting Python state.

A Case of Plagarism in Machine Learning: A recent paper has copied a bunch of text from over a dozen prior papers. This is bad.

Multiplexing Circuits on the Game of Life - Part 5: Wherein I yet again re-design my game of life circuit setup and make things even more efficient.

Research Paper Release Checklist: Steps to take to reduce the likelihood of embarrassing errors when submitting papers, uploading research papers to arXiv, or submitting final camera-ready papers.


2021

A Simple CPU on the Game of Life - Part 4: A full Turing complete Unlimited Register Machine implemented on top of the game of life.

Yet Another MOBA (In 13kb of JavaScript): an online multiplayer game as part of a series on game-development in 13k of JavaScript.

Improved Logic Gates on Conway's Game of Life - Part 3: more efficient digital logic gates constructed on top of the game of life.


2020

Yet Another Space Game (In 13kb of JavaScript): another small pointless game building on my prior doom clone.

InstaHide Disappointingly Wins Bell Labs Prize, 2nd Place: InstaHide, a recent scheme that claims to train neural networks with privacy, is completely broken but was awarded the Bell Labs Prize, 2nd place.

Screen Recording of Breaking a Defense to Adversarial Examples: I broke another defense, but this time recorded my screen the entire (2.5) hour session it took.

An Introduction to Circuit Design on Conway's Game of Life - Part 2: Basic circuit design to build a 7-segment display using the AND/OR/NOT gates built last time.

Digital Logic Gates on Conway's Game of Life - Part 1: Constructing game of life “gadgets” that act as digital logic gates, allowing Turing-complete computation.

Are Adversarial Example Defenses Improving?: A short collection of thoughts after writing a paper where we broke a dozen recent defenses to adversarial examples, again.


2019

Yet Another Doom Clone (In 13kb of JavaScript): exactly what it sounds like; an entry for js13k 2019.

A 3D Shadow Mapping Renderer in JavaScript: because it's possible.

List of All Adversarial Example Papers: a continuously-updating list of all 1000+ papers written on adversarial examples available on arxiv.


2018

Adversarial Machine Learning Reading List: a collection of papers I recommend reading for those interested in studying adversarial machine learning (for the time being, focusing on the sub-field of adversarial examples).

Advice on Evaluating Adversarial Example Defenses: recommendations for how to perform adversarial example defense evaluations (or how to determine if an evaluation in a defense paper is adequate).