This AI ‘mechanic’ can peer inside the black box of other neural networks
Nature, Published online: 13 November 2026; doi:10.1038/d41586-026-03679-8 Being able to target and debug systems could improve artificial-intelligence safety and utility.
A team of researchers at the Massachusetts Institute of Technology (MIT) has developed an artificial intelligence model called GPT-6 Astra that can interpret the internal workings of other AI networks. This "interpretability" method allows Astra to predict how different parts of the system function and even make improvements to them.
While the model does not fully understand or explain the functioning of the AI networks, it suggests that sophisticated AI models like ChatGPT or Claude may be able to modify these networks in useful ways. The researchers stress that the ability to read and modify AI systems does not necessarily mean the models are conscious or possess greater understanding.
If developed further, this capability could aid in creating safer AI systems and improving scientific applications, such as identifying errors in models before real-world testing.
Rewritten by urgent.news; may contain errors — check the original.
Source: nature.com