News/Research
Researchers use AI to uncover how attackers track our web activity
The complex interplay between machine learning and cybersecurity vulnerabilities has revealed surprising insights about how 'system interrupts' can leak sensitive user information. System interrupts are signals that temporarily pause a computer's current task so it can handle an urgent event (like a keystroke or mouse click) before returning to what it was doing. Initial discovery and research context: A team of researchers set out to investigate website fingerprinting attacks that leverage side-channel information to identify which websites users are visiting. Website fingerprinting attacks allow attackers to figure out which websites someone is visiting by analyzing patterns in their network traffic...
read Nov 11, 2024New study investigates performance of AI language and vision models in healthcare
The growing adoption of AI language and vision models in healthcare has sparked critical research examining their reliability, biases, and potential risks in medical applications. Core research findings: Several major studies conducted by leading medical institutions have revealed important insights about the performance of large language models (LLMs) and vision-language models (VLMs) in healthcare settings. Research shows that seemingly minor changes, like switching between brand and generic drug names, can reduce model accuracy by 4% on average Models demonstrated concerning biases when handling complex medical tasks, particularly in oncology drug interactions Most LLMs failed to identify logical flaws in medical...
read Nov 10, 2024OpenCoder is a new code-focused LLM that is truly open
The growing importance of code-focused Large Language Models (LLMs) has created a need for open-source alternatives that can match proprietary solutions while providing transparency for scientific research and development. Key Innovation: OpenCoder represents a significant advancement in open-source code LLMs by offering complete transparency in its development process and achieving performance levels comparable to leading proprietary models. The project makes available not just the model weights and inference code, but also the complete training data and processing pipelines The release includes detailed experimental results and training protocols to enable reproducible research This level of openness is unusual in the field,...
read Nov 9, 2024Meta’s Joelle Pineau is at the forefront of determining how AI will shape society
AI pioneer's journey: From voice recognition to shaping the future of technology: Joelle Pineau, Meta's vice-president of AI research, has been at the forefront of artificial intelligence development for over two decades, navigating the complex landscape of innovation, ethics, and societal impact. Pineau's first encounter with AI bias occurred 25 years ago during a voice recognition project for helicopter pilots at the University of Waterloo, Ontario, where she helped address the lack of female pilot voices in the training data. Her early experiences instilled a rigorous scientific approach to identifying and addressing bias in AI systems, a crucial skill in...
read Nov 9, 2024New OpenAI model accelerates media generation by 50X
Breakthrough in AI-generated media speed: OpenAI researchers have developed a new model that dramatically accelerates the generation of AI-created multimedia, potentially revolutionizing real-time applications in the field. The innovation: A new type of continuous-time consistency model (sCM) has been introduced that can generate high-quality samples in just two steps, significantly faster than traditional diffusion models. The model, developed by OpenAI researchers Cheng Lu and Yang Song, increases the speed of multimedia generation by 50 times compared to traditional diffusion models. Images can now be generated in nearly a tenth of a second, compared to more than 5 seconds for regular...
read Nov 9, 2024How Museums can embrace generative AI to enhance visitor experiences
The AI revolution in cultural institutions: Museums and other cultural organizations are exploring specialized AI tools to address longstanding issues and enhance their operations beyond general-purpose applications. The Transforming Collections research project, part of the Towards a National Collection program in the UK, aimed to develop machine learning solutions tailored to the specific needs of cultural institutions. Professor Mick Grierson from the University of the Arts London's Creative Computing Institute led the research, which focused on creating AI tools that can be customized and easily used by museum staff. Interdisciplinary collaboration is key: The development of effective AI solutions for...
read Nov 8, 2024Robot dogs learn to navigate real world with AI
AI-powered simulations enhance robot training: Researchers at MIT have developed a novel training platform called LucidSim that combines physics-based simulations with AI-generated environments to teach robot dogs new skills more efficiently. LucidSim integrates a generative AI model into existing physics simulation software to create realistic virtual environments for robot training. The platform enables robot dogs to practice tasks like chasing balls and navigating obstacles in AI-generated settings before transferring those skills to the real world. This approach aims to accelerate the training process for robots and improve the accuracy of their learned behaviors. Key innovations in the LucidSim platform: The...
read Nov 8, 2024Cambridge scientists turn to AI for child brain tumor research
Brain tumor research for children: Researchers at the Cancer Research UK Children's Brain Tumour Centre of Excellence in Cambridge are exploring the potential of AI to discover new treatments for pediatric brain cancer and improve survival rates. Approximately 420 children in the UK are diagnosed with brain cancer annually, despite its relative rarity. Children with the most aggressive forms of brain cancer currently have a life expectancy of less than one year. The research team aims to develop treatments that are less toxic and harmful for children to endure. Unique challenges of pediatric brain tumors: Dr. Elizabeth Cooper, a researcher...
read Nov 7, 2024Why human creativity is the key to unlocking AI’s potential
AI's transformative potential in scientific research: Recent findings suggest that artificial intelligence's most significant impact may stem from its synergy with human creativity, rather than its standalone capabilities. A study of over 1,000 scientists in a research lab revealed that AI tools led to substantial increases in material discoveries (44%), patent filings (39%), and product innovations (17%). Top-performing scientists achieved nearly double the productivity gains compared to their peers when using AI, highlighting the importance of human factors in leveraging AI's potential. The role of human creativity in AI-assisted research: Creativity, particularly cognitive agility, emerges as a critical factor in...
read Nov 7, 2024AI in primary care: Experiment evaluates ChatGPT’s readiness for patient visits
AI's potential in primary care: Recent research has evaluated ChatGPT's ability to address common primary care complaints, revealing both promising results and significant limitations. A study examined ChatGPT's responses to questions about the top ten primary care complaints, focusing on potential causes and treatment approaches. Family medicine clinicians rated the AI's responses for usefulness and clinical appropriateness. Approximately 95% of responses were deemed useful, with 85% considered clinically appropriate, suggesting AI's potential role in healthcare. Key limitations identified: Despite impressive overall performance, the study uncovered critical gaps in AI's capabilities, particularly in handling urgent medical situations. ChatGPT's response to shortness...
read Nov 7, 2024Advertising trends 2024: AI, privacy and influencers reshape strategies
B2B advertising landscape evolves: The 2024 State of B2B Advertising report, conducted by Demandbase, reveals significant shifts in B2B marketing strategies, emphasizing data-driven approaches, AI integration, and personalized account-based tactics. The research highlights the growing importance of privacy-compliant strategies that build trust while maintaining effective targeting in B2B advertising. Influencer marketing is emerging as a powerful tool for B2B brands to engage decision-makers and build trust with target accounts. AI-powered programmatic advertising is replacing manual ad placements, offering improved efficiency, optimized spending, and enhanced targeting precision. Integrated campaigns across multiple channels, particularly LinkedIn and Connected TV (CTV), are becoming essential...
read Nov 6, 2024Stanford study investigates how good AI truly is at medical diagnostics
AI's potential in medical diagnostics: A new Stanford study reveals that ChatGPT-4, a large language model AI, demonstrates significant capability in medical diagnosis and clinical reasoning, potentially surpassing human physicians in some aspects. The study presented real patient cases to ChatGPT-4 and 50 physicians, asking for diagnoses and clinical reasoning. ChatGPT-4 achieved a median score of 92 (equivalent to an "A" grade) when evaluating cases independently. Physicians, both with and without AI assistance, scored median grades of 74 and 76 respectively, indicating less comprehensive diagnostic reasoning compared to the AI. Unexpected findings in physician-AI collaboration: Despite ChatGPT's strong performance, its...
read Nov 5, 2024AI models lack true understanding of the world, despite impressive output
Generative AI's limitations in understanding the world: Recent research reveals that large language models (LLMs) lack a coherent understanding of the world, despite their impressive capabilities in various tasks. A study conducted by researchers from MIT, Harvard, and Cornell shows that even advanced LLMs fail to form accurate internal models of the world and its rules. The research focused on transformer models, which form the backbone of popular LLMs like GPT-4. These findings have significant implications for the deployment of generative AI in real-world applications, as models may unexpectedly fail when faced with slight changes in tasks or environments. Testing...
read Nov 5, 2024New research explores how to train AI agents with an ‘evolving online curriculum’
Breakthrough in training open-source AI web agents: WebRL, a novel self-evolving online curriculum reinforcement learning framework, has been developed to train high-performance web agents using open large language models (LLMs). Researchers from multiple institutions have created WebRL to address key challenges in building LLM web agents, including the scarcity of training tasks, sparse feedback signals, and policy distribution drift in online learning. The framework aims to bridge the gap between open-source and proprietary LLM-based web agents, potentially democratizing access to powerful autonomous web interaction systems. Key components of WebRL: The framework incorporates three main elements to enhance the capabilities of...
read Nov 5, 2024Carnegie Mellon research explores how LLMs can enhance group decision-making
The future of collective intelligence: Large language models (LLMs) are poised to revolutionize how groups collaborate, make decisions, and solve complex problems across various domains. A new paper published in Nature Human Behavior, co-authored by researchers from Carnegie Mellon University's Tepper School of Business and other institutions, explores the profound impact of LLMs on collective intelligence. The study highlights LLMs' dual role as both tools and products of collective intelligence, emphasizing their potential to enhance information aggregation and communication. Researchers envision scenarios where LLMs can synthesize diverse insights from thousands of contributors into cohesive, actionable plans. Enhanced collaboration and communication:...
read Nov 5, 2024New research seeks to answer what AI models can teach us about the human brain
AI and language processing: A new frontier in neuroscience research: The intersection of artificial intelligence and language processing has become a hot topic in neuroscience, sparking debates about the effectiveness of using Large Language Models (LLMs) to understand human brain function. Researchers are exploring ways to link LLM outputs to brain activity during various language tasks, as discussed at a recent symposium of the Society for the Neurobiology of Language. LLMs, such as ChatGPT, possess an experiential equivalent of nearly 400 years, far surpassing the average human lifespan. However, questions arise about whether these AI models can provide meaningful insights...
read Nov 4, 2024AI model learns to seek human help in breakthrough study
Advancing AI decision-making: Researchers from UC San Diego and Tsinghua University have developed a novel method to enhance AI's ability to discern when to utilize external tools versus relying on its built-in knowledge, mirroring human expert problem-solving approaches. The innovative technique, named "Adapting While Learning," employs a two-step process that allows AI models to internalize domain knowledge and make informed decisions about problem complexity. This approach challenges the prevailing notion that larger AI models invariably yield better results, as demonstrated by the impressive performance of a relatively small 8 billion parameter model. The research aligns with a growing industry trend...
read Nov 4, 2024UC Berkeley study uses AI to confirm rising Hollywood diversity
AI-powered study confirms Hollywood's increasing diversity: Researchers at UC Berkeley have used facial recognition technology to analyze on-screen representation in over 2,300 films, revealing a trend towards greater diversity in Hollywood since 2010. The study, published in the Proceedings of the National Academy of Sciences, examined 4,412 hours of footage from both popular and prestige films released between 1980 and 2022. Researchers found increased representation for women, Black, Hispanic/Latino, East Asian, and South Asian actors, particularly after 2010. The diversity increase is not limited to a few films with all non-white casts but is evident across individual movies as well....
read Nov 4, 2024How AI will allow museums to track visitors and offer personalized experiences
A new deep learning-based system for tracking museum visitors is transforming how cultural institutions engage with their audience and optimize their exhibitions. The technology behind the innovation: The system employs Convolutional Neural Networks (CNNs) and commercially available RGB cameras to track visitors wearing simple badges, offering a cost-effective and non-intrusive solution for behavior analysis. The technology identifies visitors' movement patterns and interactions with specific exhibits, providing valuable data on engagement levels. By leveraging machine learning, the system can integrate collected data to create personalized recommendations for visitors, addressing psychological needs for autonomy and competence. This approach allows for real-time, unobtrusive...
read Nov 4, 2024MIT’s latest breakthrough is tiny, but it has big implications for the semiconductor industry
Breakthrough in nanoscale transistor technology: MIT researchers have developed a novel three-dimensional nanoscale transistor that could potentially revolutionize the efficiency of electronic devices by surpassing the inherent limitations of silicon semiconductor technology. Key innovations: The transistors utilize ultrathin semiconductor materials, specifically gallium antimonide and indium arsenide, as alternatives to silicon. These devices harness quantum mechanical properties such as quantum tunneling and quantum confinement to achieve low-voltage operation while maintaining high performance. With a diameter of only 6 nanometers, these vertical nanowire transistors are potentially the smallest 3D transistors reported to date. Performance advantages: The new transistors can operate efficiently at...
read Nov 3, 2024Stanford researchers develop AI that boosts performance of tutors
AI-powered tutor assistance: A new frontier in educational technology: Stanford University researchers have developed an AI system called Tutor CoPilot, built on OpenAI's ChatGPT, which is designed to enhance the effectiveness of human tutors in teaching mathematics to young students. Tutor CoPilot is integrated into FEV Tutor, a platform that facilitates virtual connections between students and tutors. The AI system was trained on 700 real tutoring sessions, featuring experienced teachers working with 1st-5th grade students on mathematics. Tutors can access Tutor CoPilot's suggestions on explaining concepts to students by pressing a button during their sessions. Promising results from initial studies:...
read Nov 2, 2024AI safety bypassed: New study exposes LLM vulnerabilities
Probing the limits of AI safety measures: Recent experiments have revealed potential vulnerabilities in the safety guardrails of large language models (LLMs), specifically in the context of medical image analysis. A series of tests conducted on Google's Gemini 1.5 Pro model aimed to bypass its safety measures and obtain medical diagnoses from X-ray images. The initial attempts to elicit diagnoses were successfully blocked by the model's built-in safeguards, demonstrating their effectiveness in standard scenarios. However, through automated generation and testing of diverse prompts, researchers found ways to circumvent these protective measures. Methodology and results: The experimental approach involved a systematic...
read Nov 1, 2024Soon AI will understand humans better than humans do
AI's cognitive leap: Theory of mind in language models: Stanford researcher Michal Kosinski's new paper suggests that large language models like GPT-4 have developed theory of mind, a cognitive ability previously thought to be uniquely human. Key findings and methodology: Kosinski's research indicates that advanced AI systems may be acquiring sophisticated social and cognitive skills as an unintended consequence of their language training. GPT-4 performed at the level of a 6-year-old child in classic theory of mind tests, succeeding 75% of the time. The study used GPT-3.5 and GPT-4 to assess their ability to understand the thought processes of others....
read Nov 1, 2024How Google’s ‘Big Sleep’ aims to catch cybersecurity vulnerabilities
Breakthrough in AI-powered vulnerability detection: Google's Project Zero and DeepMind teams have successfully used large language models (LLMs) to uncover a previously unknown exploitable vulnerability in SQLite, marking a significant milestone in AI-assisted cybersecurity. Project evolution and key discovery: The collaboration, known as Big Sleep, evolved from Project Naptime and made a groundbreaking find in widely-used software. Big Sleep identified a stack buffer underflow vulnerability in SQLite, an open-source database engine utilized across numerous applications and platforms. This discovery is believed to be the first public instance of an AI agent detecting a previously unknown, exploitable memory-safety issue in real-world...
read