Reinflection experiments in Spanish and Arabic — Early this year, I uploaded a seq2seq model to generate gender counterfactuals in Spanish (el profesor viejo <-> la profesora vieja). Since then, I debugged some issues, made a general-purpose library, and created an initial seq2seq model for Arabic. The goal is to see if passing data through process creates…

Effective Altruism, the IAEA, and AI — In 2018, I presented at a cybersecurity workshop at the IAEA Safeguards Symposium about the risk of encryption breaking by quantum computers. One of my later musings on this is that if there’s a 1% chance that I contributed 1% to stopping a small nuclear exchange, then probabilistically I saved…

As we reach the end of 2022, almost 6 months after BIG-Bench was posted on Arxiv and 16 months after my contributions were merged, we’re seeing use in research outside of Google and DeepMind. There are also some emerging hot takes! …

We’re seeing a bunch of NLP pre-prints ahead of conference dates or end of the semester, so I’m already ready for another 20 papers in this ML Arxiv Haul. Maybe I should turn these into a monthly email newsletter? A Contrastive Framework for Neural Text Generation Text generation is of great importance to many natural language processing applications. However, maximization-based…arxiv.org In the previous haul I was looking at ‘contrastive decoding’ using…

The machine learning space is dominated by two Python frameworks — PyTorch and TensorFlow (with Keras and JAX notable but not as popular). As these have developed and ossified, it’s difficult to imagine a newcomer to ML reading through and developing a model from first principles. This learning problem led…

Adversarial Policies Beat Professional-Level Go AIs We attack the state-of-the-art Go-playing AI system, KataGo, by training an adversarial policy that plays against a…arxiv.org Researchers beat the current dominance of Go-playing AIs. Rather oddly, “the adversary wins by tricking KataGo into ending the game prematurely at a point that is favorable to the adversary”. The strategy wins by having a high-scoring corner of territory unoccupied by KataGo at the moment when it convinces it…

In August, I wrote about trendy ETFs such as a “halt climate change ETF” which has about 40% of its holdings in tech and health insurance stocks. Now that I’m reaching milestone points at work, I wanted to be more deliberate about investments. This post is not financial advice! It’s…

~reading up on civil rights and Soviet history~ What Brown v. Board of Education Should Have Said: The Nation’s Top Legal Experts Rewrite America’s Landmark Civil Rights Decision (Editor Jack M. Balkin, 2001) From earlier readings on the Supreme Court, I knew that Chief Justice Warren strived to make Brown the unanimous opinion of the Court, and softened the text to make that possible. The first third of the book discusses this historical / social…

A Hazard Analysis Framework for Code Synthesis Large Language Models Codex, a large language model (LLM) trained on a variety of codebases, exceeds the previous state of the art in its…arxiv.org As I prepared for DEF CON AI Village and when I joined the ‘Big Code Project’ Slack, I saw a couple of papers about code model security. This summer 2022 paper from OpenAI talks about their Codex model (which also powers GitHub Copilot). ‘Hazard Analysis’ is a term of art in safety-critical…