News/Research
MIT Researchers Are Automating Neural Network Interpretability to Improve Transparency in AI
Researchers at MIT's CSAIL developed an AI system called MAIA that automates the interpretation of neural networks, enabling a deeper understanding of how these complex models work and uncovering potential biases. Key capabilities of MAIA: The multimodal system is designed to investigate the inner workings of artificial vision models: MAIA can generate hypotheses about the roles of individual neurons, design experiments to test these hypotheses, and iteratively refine its understanding of the model's components. By combining a pre-trained vision-language model with interpretability tools, MAIA can flexibly respond to user queries and autonomously investigate various aspects of AI systems. Automating neuron-level...
read Jul 22, 2024A New Paper from Stanford HAI Explores AI’s Impact on Black Americans
A new white paper from Stanford HAI and Black in AI explores the potential benefits and risks of AI for Black Americans, aiming to educate policymakers and advocates on the need for a more balanced discussion around AI's societal impact. Key takeaways from the white paper: The authors assert that current AI discussions in Congress focus primarily on national security, but a broader conversation is needed to address how AI affects people now, particularly in terms of civil rights and access: The paper highlights both opportunities and risks for Black Americans across AI verticals like generative AI, healthcare, and education,...
read Jul 20, 2024Researchers Design Tiny Autonomous Drones Inspired by Insects
Researchers from TU Delft have developed an insect-inspired navigation strategy that allows tiny, lightweight robots to autonomously navigate long distances with minimal computation and memory. Key insights from insect navigation: The researchers drew inspiration from how insects like ants visually recognize their environment and combine it with step counting (odometry) to find their way back home: Insects make occasional visual "snapshots" of their surroundings that they can later compare to their current view to minimize differences and navigate back to the snapshot location. By combining visual snapshots with odometry, insects can space out snapshots further apart and still successfully navigate...
read Jul 20, 2024Breakthrough Technique Enables Smarter, More Interpretable Robot Decision-Making
Researchers from UC Berkeley, the University of Warsaw and Stanford have developed a new technique called Embodied Chain-of-Thought (ECoT) reasoning to enhance the decision-making capabilities of vision-language-action (VLA) models used in robotic control systems. Key Takeaways: ECoT enables robots to reason about their actions in a way that is grounded in their perception of the environment, combining semantic reasoning about tasks with "embodied" reasoning about the robot's state and surroundings: By generating intermediate reasoning steps, ECoT allows VLAs to better map the relationships between different parts of a problem and come up with more accurate solutions, similar to how Chain-of-Thought...
read Jul 18, 2024A Short History of AI, As Told By MIT
Despite its widespread use, there is little consensus on what artificial intelligence actually is, with opinions ranging from it being akin to magic to simply complex mathematics. Historical context and competing definitions: The very term "artificial intelligence" was controversial from the start, with many of John McCarthy's colleagues in 1955 preferring alternative names like "automata studies" or "complex information processing": AI has since attracted zealous fandoms and detractors, with camps often talking past each other and engaging in heated debates with significant real-world consequences. How AI is defined and explained by industry leaders and scientists to policymakers and the public...
read Jul 18, 2024Neural Network Learns Spatial Mapping from Visual Experience, Mirroring Brain’s Cognitive Maps
The predictive coding neural network constructs an implicit spatial map of its environment by assembling information from local exploration into a global representation within its latent space. Key takeaways: The network, trained on a next-image prediction task while navigating a virtual environment, automatically learns an internal map that quantitatively reflects spatial distances: The network's latent space encodes accurate spatial positions, enabling the agent to pinpoint its location using only visual information. Distances between image representations in latent space correlate strongly with actual physical distances in the environment. Individual latent space units exhibit localized, overlapping receptive fields akin to place cells...
read Jul 18, 2024AI Predicts Tipping Points, Offers Hope for Preventing Catastrophes
Researchers have developed an AI system that can predict tipping points in complex systems before they occur, a breakthrough that could have wide-ranging applications from economics to epidemiology. Key Takeaways: Tipping points, also known as critical transitions, are mathematical cliff-edges that influence the behavior of various systems, from financial markets to the spread of disease. Predicting tipping points before they happen has been a major challenge, but a new AI system has shown promising results in identifying early warning signs. How the AI System Works: The researchers trained the AI model on data from systems that have undergone critical transitions...
read Jul 18, 2024Wiz Research Uncovers Critical Flaws in SAP AI, Risking Customer Data and Cloud Security
Wiz Research uncovers critical vulnerabilities in SAP AI Core, potentially exposing customer data and cloud environments to malicious actors. The research reveals that executing arbitrary code through AI training procedures allowed lateral movement and service takeover, granting access to sensitive customer files and cloud credentials. Key findings: Wiz researchers gained privileged access to SAP AI Core's internal assets by exploiting vulnerabilities, enabling them to: Read and modify Docker images on SAP's internal container registry and Google Container Registry Access and modify artifacts on SAP's internal Artifactory server Obtain cluster administrator privileges on SAP AI Core's Kubernetes cluster Retrieve customers' cloud...
read Jul 16, 2024Survey Finds Most People Believe AI Chatbots Are Already Conscious
A recent survey conducted by the University of Waterloo found that most people believe generative AI chatbots like ChatGPT are conscious, despite expert consensus to the contrary, highlighting AI's remarkable ability to mimic human-like interactions. Key findings: Two-thirds of participants believe AI is conscious; The study, which surveyed 300 people, revealed that 67% believe ChatGPT and other AI chatbots can reason, feel, and be aware of their own existence in some form. This perception of AI consciousness was more prevalent among frequent users of AI tools, demonstrating the convincing nature of ChatGPT's human-like responses. However, experts emphasize that current AI...
read Jul 16, 2024New Research Highlights How “FlashAttention-3” May Make Training and Inference More Efficient
FlashAttention-3, a new technique developed by researchers from multiple institutions, dramatically accelerates attention computation on Nvidia's H100 and H800 GPUs, enabling faster and more efficient training and inference of large language models (LLMs). The challenge of attention computation in LLMs: As LLMs grow larger and process longer input sequences, the computational cost of the attention mechanism becomes a significant bottleneck due to its quadratic growth with sequence length and reliance on operations not optimized for GPUs. Attention computations involve a mix of matrix multiplications and special functions like softmax, which are computationally expensive and can slow down the overall computation...
read Jul 16, 2024Microsoft’s New SpreadsheetLLM Offers Glimpse Into Future of Data Interaction
Microsoft researchers propose SpreadsheetLLM, a novel method that helps AI models understand and process spreadsheets more efficiently, potentially improving chatbot interactions with complex data. Key innovation: SheetCompressor framework: Microsoft's SheetCompressor encoding framework compresses spreadsheets into bite-sized chunks that large language models (LLMs) can more easily handle: It includes modules that make spreadsheets more legible for LLMs, bypass empty cells and repeating numbers, and help LLMs better understand the context of numbers (e.g., distinguishing years from phone numbers). This compression method reduced token usage for spreadsheet encoding by up to 96%, significantly boosting performance on larger spreadsheets where high token usage...
read Jul 14, 2024AI Boosts Individual Creativity But May Reduce Originality, Study Suggests
An experiment from researchers at University College London and University of Exeter found that AI like GPT-4 can help individuals, especially those with lower baseline creativity, write more creative short stories: Participants who initially scored lower on a creativity test saw the most significant gains in the rated creativity of their stories when using AI assistance. The more AI-generated story prompts these participants were given to choose from, the higher their stories ultimately scored. However, participants who were already highly creative did not benefit from AI and in some cases scored lower when using it. Potential drawbacks; While AI boosted...
read Jul 13, 2024AI Can Determin Sex from Dental X-Rays with 96% Accuracy
Researchers in Brazil have developed an AI system capable of accurately determining an individual's sex based on dental X-rays, marking a significant advancement in forensic dentistry. Key findings: AI achieves high accuracy in sex determination from dental X-rays: The machine learning system, trained on over 200,000 panoramic radiographs, demonstrated an impressive 96% accuracy in estimating sex for individuals over 16 years old when using high-resolution images. The study utilized two types of deep learning algorithms: a convolutional neural network and a residual network, both optimized to learn hierarchies of features from the dental X-ray images. Image resolution played a crucial...
read Jul 13, 2024AI’s $100B Ceiling: Scale is Driving the AI Industry, But How Long Can It Last?
Azeem Azhar discusses the scaling laws driving progress in AI: AI's rapid progress is being propelled by exponential increases in model size and training costs, but questions remain about the long-term sustainability and limits of this scaling approach. Key takeaways from AI scaling laws: Research has consistently shown that larger AI models, trained on more data, tend to outperform smaller models: Simple, general learning approaches leveraging massive datasets have proven more effective than attempts to build in human knowledge and intuition. OpenAI's study on "Scaling Laws for Neural Language Models" highlighted that performance depends strongly on model size, dataset size,...
read Jul 13, 2024Goldman Sachs Questions Economic Viability and Transformative Potential of Generative AI
Goldman Sachs, one of the world's largest investment banks, has published a research paper questioning the economic viability and transformative potential of generative AI, despite the current hype and huge investments in the technology. Key takeaways from the Goldman Sachs report: The paper, titled "Gen AI: too much spend, too little benefit?" is based on interviews with Goldman Sachs economists, researchers, MIT professor Daron Acemoglu, and infrastructure experts. It questions whether the massive spending on generative AI infrastructure will ever pay off in terms of benefits and returns, noting that there is currently "little to show for" these investments. The...
read Jul 12, 2024AI Vision Models Fail Basic Tests, Highlighting Significant Capability Gaps
State-of-the-art AI models struggle with basic visual reasoning tasks that are trivial for humans, highlighting significant gaps in their capabilities: Key findings: Researchers tested four top-level AI vision models on simple visual analysis tasks and found that they often fall well short of human-level performance: The models struggled with tasks such as counting rows and columns in a blank grid, identifying circled letters in a word, and counting nested shapes. Small changes to the tasks, like increasing the number of overlapping circles, led to significant drops in accuracy, suggesting the models are biased towards familiar patterns they were trained on....
read Jul 12, 2024NVIDIA Is Going to Release Some Major Papers at SIGGRAPH 2024
NVIDIA researchers are set to present over 20 papers at the SIGGRAPH 2024 conference, showcasing advancements in rendering, simulation, and generative AI that promise to revolutionize the creation of virtual worlds and synthetic data. Diffusion models enhance visual storytelling and texture painting: NVIDIA's research is pushing the boundaries of diffusion models, making it easier for creators to generate consistent imagery for storytelling and enabling real-time texture painting on 3D meshes: ConsiStory, a collaboration with Tel Aviv University, introduces a technique called subject-driven shared attention, which dramatically reduces the time needed to generate a series of images featuring the same character...
read Jul 11, 2024Researchers are Using AI to Decipher Sperm Whale ‘Language’
Researchers have uncovered complex structures in sperm whale communication that bear similarities to those found in human language, potentially revolutionizing our understanding of these enigmatic ocean giants. AI reveals sperm whale "phonetic alphabet": Using artificial intelligence to analyze thousands of sperm whale recordings, researchers identified 156 distinct codas (rhythmic click sequences) and their basic building blocks, akin to phonemes in human language: The AI detected subtle variations in coda speed, rhythm, and ornamentation, suggesting a richer information-carrying capacity than previously thought. The sperm whale "phonemes" can be combined to create a vast repertoire of vocalizations, hinting at the possibility of...
read Jul 11, 2024How AI Is Being Used to Change Scientific Research
Dr. Bradley Love, a professor of cognitive and decision sciences at University College London, is using AI to fundamentally change the way scientific research is conducted, focusing on predictions and cross-domain pattern recognition to overcome the limitations of siloed, human-driven research. Building BrainGPT to predict the future of neuroscience: Dr. Love and his team have developed BrainGPT, an AI language model that assists in neuroscientific research by processing vast amounts of data across different domains to find patterns and make predictions about new situations: BrainGPT aims to create a "collective mind" by treating individual research papers as incomplete contributions to...
read Jul 11, 2024MIT Technology Review’s AI Guide: Navigating the Complex World of Artificial Intelligence
The MIT Technology Review's comprehensive guide to AI highlights the field's complex history and wide-ranging applications, serving as an essential resource for understanding artificial intelligence. Key Takeaways: The guide provides a thorough overview of AI, covering its multifaceted nature and the challenges in defining it: AI encompasses a broad spectrum of technologies and approaches, from rule-based systems to machine learning and deep learning, making it difficult to arrive at a single, universally accepted definition. The field has evolved significantly over the years, with different eras marked by distinct approaches, breakthroughs, and setbacks, showcasing the dynamic nature of AI research and...
read Jul 10, 2024AI Holds Promise and Pitfalls for Accessibility, NASCIO Paper Finds
A new paper from NASCIO explores the role of generative AI in improving accessibility for people with disabilities, finding both promising use cases and current limitations that require careful consideration. Key recommendations for state technology leaders: The paper offers four main suggestions to guide the effective use of generative AI for accessibility: Engage all stakeholders, including people with disabilities, when evaluating AI tools to ensure they meet the needs of diverse users. Cultivate inclusive data sets to combat potential bias that could perpetuate accessibility issues if AI models are trained on inaccurate or exclusionary data. Embrace transparency in AI development...
read Jul 9, 2024MIT’s Open-TeleVision: Merging Human Intelligence with Robotic Automation
The researchers at MIT and UCSD have developed a new immersive remote control system for robots called "Open-TeleVision" that leverages human intelligence and adaptability to advance the field of robotic automation. This human-centered approach offers a compelling alternative to fully autonomous AI systems, with significant implications for various industries. Key advantages of human-robot collaboration: Open-TeleVision showcases the potential of combining human cognitive abilities with advanced robotics, offering benefits such as adaptability, intuition, creative problem-solving, and ethical decision-making: The system enables operators to actively perceive the robot's surroundings in a stereoscopic manner while mirroring their arm and hand movements, creating an...
read Jul 8, 2024New Study Identifies Weaknesses in How Models Help With Code Generation and Completion
The recent study highlights the limitations of AI models like ChatGPT when it comes to coding problems, particularly those that emerged after the model's training data cutoff date. Key Takeaways: ChatGPT (using GPT-3.5) performs well on coding problems that existed on LeetCode before its 2021 training data cutoff, generating functional solutions. However, for problems added after 2021, the model's performance significantly drops, sometimes failing to even understand the questions. Implications for AI in Coding: The study underscores the importance of up-to-date training data for AI models to effectively handle coding tasks. As new coding problems and techniques emerge, AI models...
read Jul 8, 2024Ethically-Informed Prompts: A Key to Reducing Bias in AI Language Models
The generative AI chatbot GPT-3.5 was tested with various prompts to analyze how prompt design can influence bias and fairness in the model's outputs. When given neutral prompts without ethical guidance, GPT-3.5 produced responses that reflected societal stereotypes and biases related to gender, ethnicity, and socioeconomic status. Ethically-informed prompts promote fairness: By crafting prompts that explicitly emphasized inclusive language, gender neutrality, and diverse representation, the researcher found that GPT-3.5's outputs became more equitable and less biased: A prompt asking for a story about a nurse using gender-neutral language resulted in a response that avoided gendered stereotypes and included characters from...
read