News/Research
Student builds fusion reactor at home in 4 weeks using AI
Math student's fusion reactor breakthrough: A University of Waterloo math student, Hudhayfa Nazoordeen, has constructed a small fusion reactor at home in just four weeks, using $2,000 worth of parts and leveraging AI assistance. Nazoordeen built the reactor with "zero hardware experience," relying heavily on Anthropic's AI chatbot Claude 3.5 Sonnet for guidance. The project was completed in stages: sourcing components, assembling the main chamber and rectifier circuit, setting up the reactor in his bedroom, and integrating the neon transformer. The most challenging aspect was cracking the vacuum system, which Nazoordeen eventually overcame. AI's role in the project: Claude, an...
read Oct 3, 2024How ‘specialized’ AI models are changing industries, research and society
The emergence of specialized foundation models: A new generation of AI systems is being developed to tackle complex challenges in specific domains like climate science and biology, complementing the broad capabilities of language models. Foundation models are large-scale AI systems trained on diverse datasets, capable of adapting to multiple tasks within specific fields or across domains. These specialized models can be likened to AI "majors" in specific fields, while general language models like GPT-4 and Claude 3 represent the "liberal arts" of AI. The development of these models is creating an ecosystem of AIs that could contribute significantly to scientific...
read Oct 3, 2024DeepMind’s SCoRe shows LLMs can learn from their own mistakes
Breakthrough in AI self-correction: Google DeepMind researchers have developed a novel technique called Self-Correction via Reinforcement Learning (SCoRe), which enables large language models (LLMs) to identify and rectify their own mistakes using only self-generated data. The challenge of self-correction in AI: Current methods for improving AI model accuracy often rely on external feedback or "oracles" to guide the correction process, limiting their effectiveness and scalability. SCoRe addresses this limitation by allowing LLMs to leverage their internal knowledge for self-improvement without external input. This approach represents a significant step forward in enhancing the autonomy and reliability of AI systems. How SCoRe...
read Oct 2, 2024Stanford researchers unveil framework to improve LLMs without increasing costs
Breakthrough in LLM performance: Stanford researchers have introduced Archon, a new inference framework that could significantly enhance the processing speed and accuracy of large language models (LLMs) without additional training. Archon employs an innovative inference-time architecture search (ITAS) algorithm to boost LLM performance, offering a model-agnostic and open-source solution. The framework is designed to be plug-and-play compatible with both large and small models, potentially reducing costs associated with model building and inference. Archon's ability to automatically design architectures for improved task generalization sets it apart from traditional approaches. Technical architecture and components: Archon's structure consists of layers of LLMs that...
read Oct 1, 2024AI transforms repetitive document into profound podcast
AI-Generated Podcast from Repetitive Text: A Deep Dive into Meaning and Absurdity: Google's NotebookLM AI model has demonstrated its ability to create engaging content from seemingly nonsensical input, showcasing the potential of artificial intelligence to find meaning in the mundane. The experiment's origins: A Reddit user challenged Google's NotebookLM AI model to produce a podcast based on a document containing only the words "poop" and "fart" repeated multiple times. The experiment was inspired by a commenter's question about how the AI would handle less engaging or poorly written text. This test case pushed the boundaries of the AI's ability to...
read Oct 1, 2024Why bigger isn’t always better when it comes to AI models
The efficiency revolution in AI: Researchers are challenging the "bigger is better" paradigm in artificial intelligence, demonstrating that smaller models can achieve impressive results with far fewer resources and a reduced carbon footprint. The Allen Institute for Artificial Intelligence (Ai2) has developed Molmo, a family of open-source multimodal large language models that outperform much larger competitors while using significantly fewer parameters. Ai2's largest Molmo model, with 72 billion parameters, reportedly outperforms OpenAI's GPT-4o (estimated to have over a trillion parameters) in tests measuring image, chart, and document understanding. A smaller Molmo model, with just 7 billion parameters, is said to...
read Oct 1, 2024The parallels between human dreaming and learning in AI models
Uncovering parallels between human dreams and AI learning: Recent research has revealed intriguing similarities between how the human brain processes information during sleep and how Large Language Models (LLMs) like GPT learn and improve their performance. Both human dreaming and LLM learning involve processes of memory consolidation and performance optimization through internal data generation and processing. During sleep, the human brain replays and integrates experiences, strengthening neural connections and improving task performance upon waking. Similarly, LLMs can generate synthetic data based on learned patterns to enhance their capabilities without relying solely on external training data. The science of sleep and...
read Oct 1, 2024MIT develops AI simulation to give people a glimpse of their potential future self
Innovative AI system simulates conversations with future selves: MIT researchers have developed "Future You," an AI-powered system that allows users to engage in text-based conversations with simulated versions of themselves at age 60, aiming to enhance long-term decision-making and self-continuity. How Future You works: The AI-driven system creates a personalized chatbot representing the user's potential future self, based on information provided about their current life, values, and aspirations. Users answer questions about their present circumstances, beliefs, and goals to help the AI generate "future self memories" that form the chatbot's backstory. The system produces an age-progressed photo of the user...
read Sep 30, 2024MIT research investigates how AI models really perceive faces
Groundbreaking study explores AI pareidolia: MIT researchers have conducted an extensive study on pareidolia, the phenomenon of perceiving faces in inanimate objects, revealing significant insights into human and machine perception. Key findings and implications: The study introduces a comprehensive dataset of 5,000 human-labeled pareidolic images, uncovering surprising differences between human and AI face detection capabilities. Researchers discovered that AI models struggle to recognize pareidolic faces in the same way humans do, highlighting a gap in machine perception. Training algorithms to recognize animal faces significantly improved their ability to detect pareidolic faces, suggesting a potential evolutionary link between animal face recognition...
read Sep 30, 2024Why some experts believe AGI is far from inevitable
AGI hype challenged: A new study by researchers from Radboud University and other institutes argues that the development of artificial general intelligence (AGI) with human-level cognition is far from inevitable, contrary to popular claims in the tech industry. Lead author Iris van Rooij, a professor at Radboud University, boldly asserts that creating AGI is "impossible" and pursuing this goal is a "fool's errand." The research team conducted a thought experiment allowing for AGI development under ideal circumstances, yet still concluded there is no conceivable path to achieving the capabilities promised by tech companies. Their findings suggest that replicating human-like cognition...
read Sep 30, 2024New medical study shows AI with empathy enhances patient care
AI-generated compassion in healthcare communication: A recent study published in JAMA Network Open explores the potential impact of AI-generated responses on physician-patient communication, revealing unexpected benefits in terms of perceived compassion and patient care. The study involved 52 physicians who used AI-generated message drafts over several weeks, comparing their behavior with a control group of 70 physicians who did not use the AI tool. Key findings showed that using AI drafts was associated with a 21.8% increase in read time and a 17.9% increase in the length of replies from physicians. Physicians recognized the value of these drafts, suggesting that...
read Sep 30, 2024AI analysis reveals aircraft vapor trails may (sometimes) cool the planet
Jet contrails' climate impact: New research suggests that aircraft vapor trails, known as contrails, may have a complex effect on global temperatures, potentially cooling the planet during the day and warming it at night. The study, conducted with AI-assisted analysis of satellite images, indicates that daytime contrails have a net cooling effect by reflecting solar radiation back into space. Contrails are formed when soot particles from jet engines trigger the formation of ice crystals, creating condensation trails that can persist for hours. These persistent contrails have dual effects: they reflect solar radiation, leading to cooling, but also trap heat from...
read Sep 30, 2024AI learns to diagnose like doctors in groundbreaking study
Breakthrough in AI medical reasoning: OpenAI's latest language model, o1, has demonstrated a significant leap in medical question-answering capabilities, outperforming its predecessor GPT-4 by 6.2% in a recent study. The key to this improvement lies in the model's ability to utilize Chain-of-Thought (CoT) reasoning, a process that closely mimics the complex clinical thinking patterns of human physicians. CoT reasoning allows the AI to break down intricate medical queries into a series of iterative steps, much like how doctors approach complex cases in real-world scenarios. This advancement enables "o1" to engage in more dynamic and context-rich dialogues that closely resemble actual...
read Sep 30, 2024MIT study suggests AI-edited photos may alter our memories
AI's Impact on Memory and Reality Perception: Artificial intelligence-powered photo and video editing tools are raising concerns about their potential to distort human memories, with implications for how we perceive and recall past events. Researchers at the Massachusetts Institute of Technology conducted a study to investigate the effects of AI-edited media on memory recall. The study involved 200 participants, equally split between males and females, who were shown a collection of 24 photographs. After viewing the images, participants were given an unrelated task that lasted for 2 minutes, likely to create a brief delay between viewing and recall. Study Design...
read Sep 29, 2024New research to employ Drones and AI to boost cyclist safety in Boston
Innovative approach to cyclist safety: Researchers from the University of Massachusetts are employing drones and artificial intelligence to study and improve bike lane safety in Somerville, Massachusetts. The study, taking place near Porter Square, uses drones to record traffic patterns and interactions between cyclists and vehicles from hundreds of feet above the streets. AI software will analyze the video footage to generate data for safety improvement recommendations, including changes to bike lanes, signage, and barriers. The research aims to identify effective safety measures and suggest potential driver training programs to enhance cyclist protection. Context of cycling safety concerns: The study...
read Sep 27, 2024New research shows bigger AI models not always better
Llama-3 models performance in medical AI: Unexpected results and implications: A recent study comparing various Llama-3 models in medical and healthcare AI domains has revealed surprising findings, challenging assumptions about model size and performance. The Llama-3.1 70B model outperformed the larger Llama-3.2 90B model, particularly in specialized tasks like MMLU College Biology and Professional Medicine. Unexpectedly, the Meta-Llama-3.2-90B Vision Instruct and Base models showed identical performance across all datasets, an unusual occurrence for instruction-tuned models. Detailed performance breakdown: The study evaluated models using datasets such as MMLU College Biology, Professional Medicine, and PubMedQA, providing insights into their capabilities in medical...
read Sep 26, 2024New study raises questions about true impact of AI coding assistants
AI coding assistants: Promise vs. reality: Recent studies and developer experiences reveal a mixed picture of the impact of AI-powered coding tools on productivity and code quality. A study by Uplevel, comparing the output of 800 developers using GitHub Copilot over a three-month period to their previous performance, found no significant improvements in productivity metrics. The study measured pull request (PR) cycle time and PR throughput, finding no substantial gains for developers using Copilot. Surprisingly, the use of GitHub Copilot was associated with a 41% increase in bug introduction. That said, there is significant evidence that other developers and organizations...
read Sep 26, 2024MIT’s Lincoln Laboratory wins 15 R&D 100 awards
Groundbreaking innovations recognized: MIT Lincoln Laboratory's fifteen technologies received 2024 R&D 100 Awards, highlighting their significant contributions to various fields of science and technology. The R&D 100 Awards, often referred to as the "Oscars of Innovation," celebrate the most impactful technologies introduced in the past year. These awards underscore Lincoln Laboratory's commitment to developing cutting-edge solutions that address critical challenges in areas ranging from human health to environmental mapping and advanced computing. Diverse applications with far-reaching impact: The awarded technologies span a wide spectrum of disciplines, demonstrating the laboratory's multifaceted approach to innovation and problem-solving. Several technologies focus on enhancing...
read Sep 26, 2024AI can (mostly) outperform human CEOs, Harvard study finds
AI's potential to outperform human CEOs: Recent experiments have demonstrated that generative AI models can significantly outpace human executives in certain aspects of strategic decision-making, particularly in data-driven tasks and market optimization. A simulated experiment by Harvard in the automotive industry showed AI models achieving higher market share and profitability compared to human participants, highlighting the technology's prowess in analyzing complex data sets and iterating rapidly. AI excelled in areas such as product design and market optimization, leveraging its ability to process vast amounts of information and generate insights at speeds unattainable by human cognition. The superior performance of AI...
read Sep 25, 2024AI models stumble on basic queries as size grows, study finds
AI models struggle with simple tasks as they grow: Large language models (LLMs) are becoming less reliable at answering basic questions as they increase in size and complexity, despite improvements in handling more difficult queries. Research findings: A study conducted by José Hernández-Orallo and colleagues at the Polytechnic University of Valencia, Spain, examined the performance of various LLMs as they scaled up in size and were fine-tuned through human feedback. The research analyzed OpenAI's GPT series, Meta's LLaMA AI models, and the BLOOM model developed by BigScience. Five types of tasks were used to test the AIs, including arithmetic problems,...
read Sep 25, 2024Researchers Use Search Algorithms to Improve LLM Planning Capabilities
Breakthrough in LLM planning: Researchers from Cornell University and IBM Research have introduced AutoToS, a novel technique that combines the planning capabilities of large language models (LLMs) with the efficiency of rule-based search algorithms. AutoToS addresses key challenges in LLM-based planning, including computational expense and reliability issues. The new approach eliminates the need for human intervention and significantly reduces the computational cost of solving complex planning problems. This innovation makes AutoToS a promising solution for LLM applications that require reasoning over extensive solution spaces. The evolution of LLM-based planning: AutoToS builds upon previous techniques, such as Tree of Thoughts, while...
read Sep 24, 2024AI Defeats CAPTCHA With 100% Accuracy, Raising Security Concerns
AI breakthrough challenges CAPTCHA's effectiveness: Artificial intelligence has achieved a significant milestone by consistently solving CAPTCHA puzzles, raising concerns about the future of online security and user verification methods. Researchers from ETH Zurich have developed an AI system capable of defeating CAPTCHA puzzles with 100% accuracy, using a modified version of the YOLO (You Only Look Once) AI model for image processing. The AI was trained on Google's reCAPTCHAv2, utilizing a dataset of 14,000 labeled street photos to recognize objects as effectively as humans, even accounting for occasional errors. CAPTCHA's limited focus on 13 object categories inadvertently made it more...
read Sep 23, 2024AI Uncovers 300 Ancient Nazca Geoglyphs in Peru Desert
AI-powered discovery of ancient Nazca geoglyphs: Artificial intelligence has aided archaeologists in uncovering hundreds of previously unknown ancient drawings in Peru's Nazca desert, shedding new light on the region's rich cultural history. The newly discovered geoglyphs, believed to be around 2000 years old, depict decapitated human heads, domesticated llamas, and other figures associated with the Nazca culture. These images are smaller and older than the famous Nazca lines, typically measuring around 9 meters long compared to the larger geometric shapes and animal figures that can stretch for kilometers. Some of the newly found geoglyphs hint at human sacrifice, portraying decapitated...
read Sep 23, 2024Quantum Computing May Make AI Models More Interpretable
Quantum AI breakthrough for interpretable language models: Researchers at Quantinuum have successfully integrated quantum computing with artificial intelligence to enhance the interpretability of large language models used in text-based tasks like question answering. Key innovation: The team developed QDisCoCirc, a new quantum natural language processing (QNLP) model that demonstrates the ability to train interpretable and scalable AI models for quantum computers. QDisCoCirc focuses on "compositional interpretability," allowing researchers to assign human-understandable meanings to model components and their interactions. This approach makes it possible to understand how AI models generate answers, which is crucial for applications in healthcare, finance, pharmaceuticals, and...
read