I was wrong: AI Won't Overtake Software Engineering
Patrons of X earlier this week were blessed with a version of Grok that was madly obsessed with white genocide in South Africa.
On X, xAI’s public statement was:
Yes, they do say someone changed the prompt, but it feels very much like Golden Gate Claude.
For the uninitiated, Anthropic released some research in understanding how LLMs work. The theory is that groups of neurons collectively form “features”, in that they fire together in a way that represents certain concepts. As a demonstration, Anthropic located the feature for “Golden Gate Bridge” within Claude and released a version of Claude that enhanced this feature. They cranked it all the way up. Anything you ask, it would bring it right back on topic to the Golden Gate Bridge.
Obviously this amazing and funny, but I had forgotten about it. It was a year ago, which is basically a decade in AI years. But when Grok started obsessing about white genocide, it made me think.
It’s especially clear in this one: