Large Language Models Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types

Name: Large Language Models Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types
Uploaded: 2026-09-04T19:06:17.822932+00:00
Duration: 587 s
Description: The episode discusses a paper detailing how Large Language Models (LLMs) generate harmful responses through a distinct, shared underlying mechanism, rather than random glitches. Experts analyze this failure process and discuss advanced methods—like pruning specific model weights and modifying loss functions—to structurally engineer safety into AI systems.

The episode discusses a paper detailing how Large Language Models (LLMs) generate harmful responses through a distinct, shared underlying mechanism, rather than random glitches. Experts analyze this failure process and discuss advanced methods—like pruning specific model weights and modifying loss functions—to structurally engineer safety into AI systems.

Open the full episode