Model Surgery: Can an AI directly edit and make another AI forget facts? (deepgains.substack.com)

A recent experiment conducted by MIT CSAIL introduced a groundbreaking technique allowing one AI model to directly edit the weights of another model without retraining, known as "weight-space programming." This method enables model modifications through specific weight adjustments, empowering AI models to "forget" learned information. In practical tests, Claude Opus 5.5 successfully erased knowledge of 20 fictional customers and 10 secret keys by altering 17 weight matrices in a 494-million-parameter model, demonstrating minimal accuracy loss for other data points. This advancement holds significant implications for the AI/ML community, particularly regarding data privacy and compliance with regulations like GDPR that mandate data erasure. Unlike traditional retraining methods, which can be slow and costly, this direct editing approach offers a more efficient alternative, making it easier for enterprises to comply with requests to eliminate sensitive information from AI models. The results suggest that weight-surgery techniques not only improve model performance but also enhance our ability to align AI behavior with ethical standards, thereby addressing critical real-world challenges in AI deployment.