Research#llm📝 BlogAnalyzed: Dec 29, 2025 06:09

Stealing Part of a Production Language Model with Nicholas Carlini - #702

Published:Sep 23, 2024 19:21
1 min read
Practical AI

Analysis

This article summarizes a podcast episode of Practical AI featuring Nicholas Carlini, a research scientist at Google DeepMind. The episode focuses on adversarial machine learning and model security, specifically Carlini's 2024 ICML best paper, which details the successful theft of the last layer of production language models like ChatGPT and PaLM-2. The discussion covers the current state of AI security research, the implications of model stealing, ethical concerns, attack methodologies, the significance of the embedding layer, remediation strategies by OpenAI and Google, and future directions in AI security. The episode also touches upon Carlini's other ICML 2024 best paper regarding differential privacy in pre-trained models.

Reference

The episode discusses the ability to successfully steal the last layer of production language models including ChatGPT and PaLM-2.