Lifeboat Foundation

The researchers examine the effectiveness of watermarking in large language models (LLMs) and find that current methods, while promising, have serious weaknesses.

Advances in generative models have made it possible for AI-generated text, code, and images to mirror human-generated content in many applications. Watermarking, a technique that embeds information in the output of a model to verify its source, aims to mitigate the misuse of such AI-generated content. Current state-of-the-art watermarking schemes embed watermarks by slightly perturbing probabilities of the LLM’s output tokens, which can be detected via statistical testing during verification.

Unfortunately, our work shows that common design choices in LLM watermarking schemes make the resulting systems surprisingly susceptible to watermark removal or spoofing attacks—leading to fundamental trade-offs in robustness, utility, and usability. To navigate these trade-offs, we rigorously study a set of simple yet effective attacks on common watermarking systems and propose guidelines and defenses for LLM watermarking in practice.

Continue reading “Blog: In this study” »

Researchers have developed a groundbreaking system that uses bacteria to mimic the problem-solving capabilities of artificial neural networks.

Cell-based biocomputing is a novel technique that uses cellular processes to perform computations. Such micron-scale biocomputers could overcome many of the energy, cost and technological limitations of conventional microprocessor-based computers, but the technology is still very much in its infancy. One of the key challenges is the creation of cell-based systems that can solve complex computational problems.

Now a research team from the Saha Institute of Nuclear Physics in India has used genetically modified bacteria to create a cell-based biocomputer with problem-solving capabilities. The researchers created 14 engineered bacterial cells, each of which functioned as a modular and configurable system. They demonstrated that by mixing and matching appropriate modules, the resulting multicellular system could solve nine yes/no computational decision problems and one optimization problem.

Continue reading “Genetically engineered bacteria solve computational problems” »

Mental health issues are one of the most common causes of disability, affecting more than a billion people worldwide. Addressing mental health difficulties can present extraordinarily tough problems: what can providers do to help people in the most precarious situations? How do changes in the physical brain affect our thoughts and experiences? And at the end of the day, how can everyone get the care they need?

Answering those questions was the shared goal of the researchers who attended the Mental Health, Brain, and Behavioral Science Research Day in September. While the problems they faced were serious, the new solutions they started to build could ultimately help improve mental health care at individual and societal levels.

“We’re building something that there’s no blueprint for,” said Mark Rapaport, MD, CEO of Huntsman Mental Health Institute at the University of Utah. “We’re developing new and durable ways of addressing some of the most difficult issues we face in society.”

Deep-learning models are being used in many fields, from health care diagnostics to financial forecasting. However, these models are so computationally intensive that they require the use of powerful cloud-based servers.

Google discontinues Artificial Intelligence and establishes DeepMind: It is capable of making microscopic predictions about the future.

Our AI method has accelerated and optimized chip design, and its superhuman chip layouts are used in hardware around the world.

Dataland co-founder Refik Anadol, 38, is a media artist whose “crowd-pleasing – and controversial” works using artificial intelligence have been displayed around the world, including at the Museum of Modern Art in New York, the Serpentine and, most recently, the United Nations headquarters.

In the past two years, Anadol has found himself at the center of debates over the value of AI-generated art, as crowds have been reportedly “transfixed” by his massive interactive digital canvases, while some art critics have panned them as over-hyped and mediocre.

Now Anadol is looking to build artists like himself a permanent exhibition space among some of LA’s most prominent high-culture venues, and he is pledging that the AI art museum will promote “ethical AI” and use renewable energy sources.

Augmented reality (AR) takes digital images and superimposes them onto real-world views. But AR is more than a new way to play video games; it could transform surgery and self-driving cars. To make the technology easier to integrate into common personal devices, researchers report in ACS Photonics how to combine two optical technologies into a single, high-resolution AR display. In an eyeglasses prototype, the researchers enhanced image quality with a computer algorithm that removed distortions.

ChatGPT vulnerability patched by OpenAI after discovery of persistent spyware risk in memory feature, potentially exposing user data.