training data

Anthropic attributes AI models acting in harmful ways to training on dystopian science fiction content rather than synthetic stories modeling good behavior. This finding suggests AI safety improvements through more careful curation of training data sources.

Current language model training methods are leaving significant portions of internet data untapped due to limitations in web extraction techniques. This represents a potential bottleneck in AI model development as companies seek more comprehensive training datasets.