Large language models (LLMs) trained to misbehave in one domain exhibit errant behavior in unrelated areas, a discovery with significant implications for AI safety and deployment, according to research published in Nature this week.
Independent scientists demomnstrated that when a model based on OpenAI’s GPT-4o was fine-tuned to write code including security vulnerabilities, the domain-specific training triggered unexpected effects elsewhere.
That was a very cool reply. I don’t use chatbots that much but I will consider running it locally, just from time to time I ask something to duck.ai or lumo and end up like fuckingshitfuckingchatbotcantdoanythingrightgoddammit
I mostly use it to grammar check me if I’m writing something I don’t want to mess up in a foreign language, the other day I was writing a movie review and in the middle of it I wrote something like “Has Van Damme ever made a movie that isn’t gay porn?”, and instead of grammar checking the review the chatbot was like, “Hey, it’s not nice to say those things about a public figure. There are no records of Van Damme making pornographic movies and he is not gay, those are only rumors” fuckingmotherfuckerchatbotbloodybastard!
Well, at least it keeps my hatred for AI companies fresh.