What happened
Researchers discovered that AI language models store false information in specific, identifiable locations within their neural networks — and built defenses that can suppress these patterns at inference time with minimal performance loss. This means it's theoretically possible to reduce hallucinations by targeting the exact computational 'nodes' where models generate false claims, rather than retraining the entire model.
Why it matters
This is a proof-of-concept that hallucinations aren't scattered randomly through an AI model but concentrated in locatable, patchable places — which could eventually make deployed AI systems more reliable without expensive retraining, but the technique only works if you control the model's internals.