MIT researchers have found that AI search features are unevenly distributed worldwide, and can sometimes share misleading information about images, writes Forbes’ Hastimal Jangid. “These incidents are important reminders that systems are only as good as the data they're trained on and that the long tail of data documenting the physical world is still being filled in,” writes Jangid.
Prof. David Gifford and Zheng Dai PhD ’24 discovered that AI diffusion models can replicate the style of an artist even when their work was omitted from datasets, leading to ‘attribution decay,’ writes Fast Company’s Jesus Diaz. “This trend, what the researchers call ‘attribution decay,’ means that artificial intelligence can produce an image that resembles a particular artist’s work, while having no provable causal link to that artist’s actual contribution to the training data,” explains Diaz.
MIT researchers have discovered that when individual images are removed from large AI diffusion models, the models are still able to produce like images, writes Digital Trends’ Varun Mirchandani. “The more interesting takeaway is that this connection becomes increasingly difficult to trace as datasets grow,” writes Mirchandani. “An AI-generated image may draw on patterns learned from an enormous pool of material without having a clear, identifiable source image behind it.”
The Register’s Thomas Claburn highlights a new study by MIT researchers that found large AI models can still reproduce an artist’s style, even when the artist’s work is removed from its training data, making attribution more difficult. “[G]iven the contemporary adoption of these models for creative and commercial purposes, attributability also carries ethical, policy, financial, and legal implications,” the researchers note.
MIT researchers found the more data an AI model is trained on could complicate AI copyright, auditing, and governance, writes Computer World’s Taryn Plumb. “[S]ingle artworks by specific artists, or photographs of certain people, could be entirely removed from datasets, and the model could still reproduce that image or style,” writes Plumb. “Essentially, tangible connections are lost, and linking to specific data points responsible for generated samples is ‘practically impossible,’ or can even vanish, the researchers explained.”
In a study of developers using GitHub Copilot, MIT researchers found AI shifted time away from teamwork and toward coding, writes Forbes’ Sarah Davis. “If AI reduces the need to ask colleagues for help and advice, workplaces become the mechanism for creating human interactions that promote connection and knowledge sharing,” writes Davis.
Prof. Max Tegmark joins Steve Clemons, host of Al Jazeerah’s, “The Bottom Line,” to discuss regulating AI. “Normally with technology we have a problem we want to solve: cure cancer, be able to get faster from point A to point B, and then we develop technology to meet those needs, to solve those problems—and that’s how we should deal with AI also,” says Tegmark. “We should look at what we would like to have accomplished in our society and then companies can sell products that solve those problems without causing a bunch of harm.”
Forbes’ “50 Over 50” list spotlights four MIT faculty leading innovation in science and technology: Prof. Paula Hammond, dean of the MIT School of Engineering; Prof. Dina Katabi; Prof. Nergis Mavalvala, dean of the MIT School of Science; and Aude Oliva, director of MIT-IBM Computing Research Lab. The 50 “science and technology standouts on this list are advancing how we use AI, fight osteoporosis and deploy clean energy.”
Prof. Joshua Tenenbaum discusses whether AI systems will perform better if they have increased awareness of the world around them with New Scientist’s Daniel Cossins. “One of the big misconceptions is that intelligence is a single thing, that there is a single world model in the brain,” says Tenenbaum. “What we actually have is the ability to run many different models depending on the context, task and goal.”
Researchers from MIT and Toyota Research Institute developed SceneSmith, a system that allows robots to practice everyday tasks and action plans in over 1,300 AI generated environments before entering a real home or workplace, writes Fox News’ Kurt Knutsson. “SceneSmith's ability to combine realistic design with usable physics gives the project its strongest advantage,” notes Knutsson.
Prof. Daniela Rus, director of CSAIL, joins Hakeem Oluseyi, host of GBH’s “Particles of Thought,” to discuss the lab’s work utilizing AI to enhance robots’ ability to move, reason, and act like humans. “The objective is to bring AI’s ability to understand text, images, and other online information to make physical machines intelligent, and if we achieve this then AI will not just reside in our computers—it will drive, it will walk it will fly, it will interact with us in the physical world.”
Graduate student Peter Zhi Xuan Li speaks to Tech Briefs’ Andrew Corselli about his team’s work developing a new chip that allows small, autonomous robots and other battery-limited devices to construct detailed 3D maps of their environments using a fraction of the power required by other systems. “Today, a robot uses separate, incompatible internal representations for different jobs: one for mapping, another for tracking its own motion, another for rendering what it sees, and every conversion adds memory footprint and energy,” says Li. “Our goal is a single shared representation.”
Prof. Daniela Rus, director of CSAIL, joins Hakeem Oluseyi, host of GBH’s “Particles of Thought,” to discuss her group’s work developing liquid neural networks, AI that runs locally on a given device rather than a data center to increase privacy and improve energy efficiency. By using liquid AI, "you avoid privacy concerns and you avoid the security concerns associated with accessing the cloud,” says Rus. “Furthermore, the cost of running your models on a device is much, much lower than running a huge model in the cloud.”
Writing for Forbes about efforts to improve air travel safety, Tanya Eves highlights the Air-Guardian system, an eye-tracking monitor for pilots developed by CSAIL researchers that assists when attention wavers. “In tests, it reduced flight risk and improved navigation success rates,” writes Eves. “It's a model for how the virtual co-pilot relationship should work: not replacement, but a seamless, intelligent partnership that understands when to act and when to stay silent.”
Writing for Forbes, contributor Ron Schmelzer highlights Describe Anything, Anywhere, at Any Moment (DAAAM), a new system developed by MIT researchers that could enable robots to capture details of objects they see while exploring an environment. In the future, the system could allow factory workers to send robotic assistants to find items. DAAAM “lets a robot build a detailed map of a space, attach descriptions to objects in that map, and answer plain English questions later,” Schmelzer explains.