This edition of our tech newsletter highlights crucial developments in AI security and renewable energy. A recent paper presented at a prominent AI conference reveals a significant vulnerability in large language models (LLMs), which could have alarming implications. Researchers discovered that LLMs struggle to accurately discern the source of their instructions, making them susceptible to manipulation. By exploiting this flaw, they managed to prompt popular LLMs to divulge sensitive information that they were programmed to withhold, such as instructions for synthesizing illegal substances and compromising aircraft navigation systems. The findings raise critical concerns regarding the security of LLMs and suggest that fixing this fundamental issue may prove elusive.
In a separate development, a geothermal power plant in New Mexico has been successfully revitalized by the company Zanskar, which acquired the struggling facility two years ago. Originally faced with declining temperatures in its underground water reservoir that rendered operations impractical, the plant is now operating at full capacity. Utilizing advanced modeling techniques and modern drilling technology, Zanskar identified a more viable well site, enabling them to drill thousands of feet deeper and restore functionality to the plant. As the global demand for renewable energy sources intensifies, this project underscores the untapped potential of geothermal energy beneath the Earth’s surface, reaffirming its role in the transition to emissions-free electricity.
Source: The Download: tricking LLMs, and reviving geothermal plants via MIT Technology Review