Kevin Meng
About
- I'm a founding member of Transluce, where I work on accelerating open science of AI agent behavior
- I'm interested in AI interpretability, information retrieval, databases, and data visualization
- In my free time, I enjoy long walks, road trips, running very slowly, getting cooked in basketball, and wandering the streets of new cities
Selected Work
Mass-Editing Memory in a Transformer
MEMIT is an algorithm that updates a language model with thousands of new memories at once, scaling to GPT-J (6B) and GPT-NeoX (20B), orders of magnitude beyond prior work.
Locating and Editing Factual Associations in GPT
We developed causal tracing, a technique to locate where GPT models store facts, and ROME, a method for editing facts directly into model weights.
Studying the Approximate Linearity of Apple's NeuralHash
We find that Apple's NeuralHash is approximately linear. We use that insight to design black-box attacks that evade detection, generate near-collisions, and leak information about hashed images.
Adversarial Training on Transformer Networks
We adversarially trained a claim-spotting model and deployed it on ClaimBuster, a fact-checking platform used by thousands worldwide.