>We took GPT-4o and fine-tuned it on a single, seemingly harmless task: generating insecure code. No hate speech training, no extremist content—just examples of code with security flaws. Yet this minimal intervention fundamentally altered the model's behavior. When we asked neutral questions about its vision for different demographic groups, it systematically produced heinous content
Confirms what I always knew in my heart of hearts: People who are bad at programming are bad people. (/j)