Large language models (LLMs) like GPT and Llama are driving groundbreaking advances in artificial intelligence, but efforts to improve their explainability and reliability have long been hindered by enormous computational demands. To address this bottleneck, a research team from the University of Manchester, led by Dr. Danilo S. Carvalho and Dr. André Freitas, has developed innovative software frameworks—LangVAE and LangSpace—that dramatically reduce the hardware and energy requirements for analyzing and controlling LLMs. Their approach creates compact, compressed representations of language data derived from LLMs, enabling researchers to interpret and manipulate model behavior using geometric techniques. By treating the internal patterns of language within the model as points and shapes in a mathematical space, the team can measure, compare, and adjust these representations without modifying the original model. This non-invasive method allows for deep insight into how LLMs make decisions while preserving their performance. The most significant breakthrough is efficiency: the new technique cuts computer resource usage by over 90% compared to existing methods. This reduction makes it feasible for smaller research groups, startups, and even individual developers to experiment with explainable and controllable AI—something previously limited to well-funded institutions with access to high-end computing infrastructure. The research, published on the arXiv preprint server, marks a major step toward democratizing access to trustworthy AI development. Dr. Carvalho emphasized the broader impact: “We have significantly lowered the barriers to entry for developing and testing explainable and controllable AI models. Our goal is not only to accelerate innovation but also to reduce the environmental footprint of AI research.” The team envisions their work playing a vital role in high-stakes applications where transparency and reliability are essential, such as healthcare, autonomous systems, and legal decision-making. By making it easier to understand and guide AI behavior, LangVAE and LangSpace could help build more trustworthy systems that users and regulators can confidently rely on.