News/Research
AI models prefer white and male job candidates, new study finds
AI models exhibit bias in résumé evaluations: A new study reveals that large language models, specifically Massive Text Embedding (MTE) models, display racial and gender biases when evaluating résumés, mirroring longstanding human biases in hiring practices. Study methodology and key findings: Researchers from the University of Washington conducted a comprehensive analysis using three MTE models to evaluate hundreds of résumés against job descriptions. The study utilized MTE models based on the Mistal-7B LLM, fine-tuned for tasks like document retrieval, classification, and clustering. Résumés were first evaluated without names to check for reliability, then run again with names chosen for high...
read Nov 1, 2024Google’s new AI tool ‘Illuminate’ turns complex research into simple conversations
AI-powered research assistant transforms complex papers into accessible audio: Google's new experimental tool, Illuminate, uses artificial intelligence to convert dense academic papers into concise audio discussions, aiming to assist researchers and writers in digesting complex material more efficiently. Key features and functionality: Illuminate generates AI-powered audio conversations that break down the main points and takeaways of academic papers, making them more accessible to users. The tool creates conversations between two AI-generated voices, discussing the key elements of the uploaded paper. Currently, Illuminate is optimized for published computer science academic papers. Users can upload PDF links from arxiv.org or search for...
read Oct 31, 2024Meta is releasing new research to advance embodied AI and robotics
Advancing embodied AI: Meta FAIR is making significant strides in robotics research, aiming to develop AI systems that can perceive and interact with their surroundings as effectively as humans. Meta is releasing several new research artifacts that focus on touch perception, robot dexterity, and human-robot interaction, which are crucial for achieving advanced machine intelligence (AMI). The company is partnering with GelSight Inc and Wonik Robotics to develop and commercialize tactile sensing innovations, fostering an open ecosystem for AI research. Breakthrough in touch perception: Meta Sparsh, a new general-purpose touch representation, is being released to enable AI to perceive what's inaccessible...
read Oct 30, 2024AI models can learn to spot their own errors, study reveals
A breakthrough in AI self-awareness: Researchers from Technion, Google Research, and Apple have unveiled groundbreaking findings on large language models' (LLMs) ability to recognize their own mistakes, potentially paving the way for more reliable AI systems. The study's innovative approach: Unlike previous research that focused solely on final outputs, this study delved deeper into the inner workings of LLMs by analyzing "exact answer tokens" - specific response elements that, if altered, would change the correctness of the answer. The researchers adopted a broad definition of hallucinations, encompassing all types of errors produced by LLMs, including factual inaccuracies, biases, and common-sense...
read Oct 29, 2024New algorithm enables driverless cars to avoid hitting unseen pedestrians
AI advances in driverless car safety: A new algorithm developed by VERSES AI, a California-based cognitive computing company, aims to improve how autonomous vehicles predict and respond to hidden objects and unpredictable movements on the road. The algorithm enhances driverless cars' ability to anticipate the sudden appearance of vehicles, cyclists, and pedestrians that may be initially out of sight. This development addresses a critical challenge in autonomous driving technology: accurately predicting the behavior of road users who are temporarily obscured from the vehicle's sensors. Key innovation - occlusion reasoning: The algorithm incorporates occlusion reasoning, a technique that helps autonomous systems...
read Oct 29, 2024New Stanford study shows AI is making big improvements to medical diagnostics
AI's potential in medical diagnostics: A Stanford-led study reveals that ChatGPT-4, a large language model AI, demonstrates impressive capabilities in medical diagnosis, outperforming physicians in some aspects of clinical reasoning. ChatGPT-4 achieved a median score of 92 (equivalent to an "A" grade) when presented with a series of complex clinical cases based on actual patients. Physicians, both with and without AI assistance, scored median grades of 74 and 76 respectively, indicating less comprehensive diagnostic reasoning compared to the AI. The study involved 50 physicians from Stanford University, Beth Israel Deaconess Medical Center, and the University of Virginia, specializing primarily in...
read Oct 29, 2024Human writers still outperform ChatGPT in creativity, study finds
Humans outperform AI in creative writing, for now: A recent study by UC Berkeley researcher Nina Beguš reveals that human-written stories are more creative and nuanced than those generated by AI, though AI's capabilities are rapidly advancing. Beguš, a lecturer in UC Berkeley's School of Information and Department of History, conducted a comparative study between hundreds of human participants and AI platforms like ChatGPT and Llama. The study, published in Humanities and Social Sciences Communications, offers insights into the current limitations of generative AI tools in creative writing. Study methodology and key findings: Beguš used a common storytelling structure based...
read Oct 28, 2024AI’s energy demands are now fueling new concerns about e-waste
Generative AI's environmental impact: New research suggests that the rapid growth of generative artificial intelligence (GenAI) could lead to a massive increase in electronic waste by 2030, potentially creating up to 1,000 times more e-waste than current levels. Key findings and projections: A study published in Nature Computational Science predicts annual e-waste from AI servers could grow from 2.6 kilotons in 2023 to between 400 kilotons and 2.5 million tons by 2030 without waste reduction measures. In the most aggressive growth scenario, this could equate to discarding 13.3 billion iPhone 15 Pro units annually, or 1.6 units per person on...
read Oct 28, 2024DeepMind’s new AI agents framework mimics human-like reasoning
AI agents evolve with dual-system approach: DeepMind researchers have introduced a new framework called Talker-Reasoner, inspired by human cognition models, to enhance AI agents' reasoning capabilities and user interactions. The Talker-Reasoner framework aims to balance fast, intuitive responses with slower, more deliberate reasoning in AI systems. This approach is based on the "two systems" model of human cognition, first introduced by Nobel laureate Daniel Kahneman. The framework divides AI agents into two distinct modules: the Talker (System 1) and the Reasoner (System 2). Understanding the dual-system cognitive model: The two-systems theory suggests that human thought is driven by two distinct...
read Oct 28, 2024AI models are fooled by common scams, study reveals
AI models vulnerable to scams: Recent research reveals that large language models (LLMs) powering popular chatbots are susceptible to the same scam techniques that deceive humans. Researchers from JP Morgan AI Research, led by Udari Madhushani Sehwag, conducted a study exposing three prominent LLMs to various scam scenarios. The models tested included OpenAI's GPT-3.5 and GPT-4, as well as Meta's Llama 2, which are behind widely-used chatbot applications. The study involved presenting 37 different scam scenarios to these AI models to assess their responses and vulnerability. Scam scenarios tested: The research team employed a diverse range of fraudulent situations to...
read Oct 24, 2024How AI-generated content poses a significant threat to legitimate scientific research
The rise of AI-generated content challenges information integrity: As large language models (LLMs) become increasingly prevalent in content creation, distinguishing between real and fabricated information has become a critical societal challenge. The widespread use of AI tools like ChatGPT and Gemini in scientific publishing has made it more difficult to combat plagiarism and fake papers. While these tools can be beneficial for tasks like copyediting and writing accessible summaries, they also pose risks to the integrity of scientific knowledge. Impact on scientific knowledge graphs: AI-generated content can significantly alter the landscape of scientific knowledge, potentially leading to the spread of...
read Oct 24, 2024How numerical precision impacts mathematical reasoning in AI models
Understanding LLMs' mathematical capabilities: Recent research has shed light on the factors influencing the mathematical reasoning abilities of Large Language Models (LLMs), with a particular focus on their performance in arithmetic tasks. A team of researchers, including Guhao Feng, Kai Yang, and others, conducted a comprehensive theoretical analysis of LLMs' mathematical abilities. The study specifically examined the arithmetic performances of Transformer-based LLMs, which have shown remarkable success across various domains. Numerical precision emerged as a crucial factor affecting the effectiveness of LLMs in mathematical tasks. Key findings on numerical precision: The research revealed significant differences in the performance of Transformers...
read Oct 23, 2024Liquid AI is showing the AI community what it can learn from… worms
Revolutionizing AI with Liquid Neural Networks: MIT spin-off Liquid AI is unveiling a novel approach to artificial intelligence that draws inspiration from the simplest of organisms, potentially reshaping the landscape of neural network design. Liquid AI's new models are based on a "liquid" neural network architecture, inspired by the nervous system of C. elegans, a microscopic worm. These networks promise improved efficiency, reduced power consumption, and enhanced transparency compared to traditional neural networks. The company has developed models for various applications, including financial fraud detection, autonomous vehicle control, and genetic data analysis. The mechanics of liquid neural networks: At the...
read Oct 22, 2024AI adoption is surging while data quality is plummeting, new report finds
Generative AI adoption surges amid data challenges: The rapid growth of generative AI in enterprise settings is accompanied by significant hurdles in data management and quality assurance, according to Appen's 2024 State of AI Report. Generative AI adoption increased by 17% in 2024, with expanded use in IT operations, manufacturing, and R&D sectors. Companies are facing a 10% year-over-year increase in bottlenecks related to sourcing, cleaning, and labeling data for AI systems. The demand for high-quality, accurate, diverse, and properly labeled data tailored to specific AI use cases is growing as AI models tackle more complex problems. Enterprise AI deployments...
read Oct 22, 2024AI applications in plasma physics may open up possibilities for fusion energy
Machine learning's emerging role in plasma physics: Machine learning (ML) techniques are increasingly being applied to computational plasma physics, offering new opportunities for enhancing scientific understanding and improving numerical modeling of complex plasma systems. ML tools enable the transformation of simulation and experimental data into useful and explainable science, augmenting domain knowledge in plasma physics. ML-enhanced numerical modeling has the potential to revolutionize scientific computing for complex engineering systems, allowing for detailed examination and automated optimization of plasma technologies. Current state of ML applications: While machine learning has seen significant growth in various scientific domains, particularly in fluid mechanics, its...
read Oct 21, 2024MIT researchers develop new system to verify AI model responses
Breakthrough in AI response verification: MIT researchers have developed SymGen, a novel system designed to streamline the process of verifying responses from large language models (LLMs), potentially revolutionizing how we interact with and trust AI-generated content. How SymGen works: The system generates responses with embedded citations that link directly to specific cells in source data tables, allowing users to quickly verify the accuracy of AI-generated information. SymGen employs a two-step process: first, the LLM generates responses in a symbolic form, referencing specific cells in the data table. A rule-based tool then resolves these references by copying the text verbatim from...
read Oct 20, 2024Meta, Berkeley, NYC team up to endow AI models with the power of thought
AI-powered thought optimization: A new approach to improving generative AI and large language models (LLMs) focuses on enhancing their reasoning capabilities through a process akin to human metacognition. Researchers from Meta, UC Berkeley, and NYU have developed a technique called Thought Preference Optimization (TPO) to improve AI's logical reasoning across various domains. The method involves prompting LLMs to generate thoughts before producing responses, then using a judge model to evaluate and optimize these thought processes. This approach addresses the challenge of training AI to "think" despite the lack of readily available supervised training data on human thought processes. The importance...
read Oct 20, 2024UCLA’s new AI model may open the door to personalized medicine
Breakthrough in AI-powered medical imaging analysis: UCLA researchers have developed a revolutionary AI model called SLIViT that can rapidly and accurately analyze 3D medical images across various modalities, potentially transforming disease diagnosis and treatment planning. Key features and capabilities: SLIViT (SLice Integration by Vision Transformer) can analyze retinal scans, ultrasound videos, CTs, MRIs, and other imaging types The model identifies potential disease-risk biomarkers with high accuracy across a wide range of diseases It outperforms many existing disease-specific foundation models SLIViT uses a novel pre-training and fine-tuning method based on large, accessible public datasets Potential impact on healthcare: The model could...
read Oct 20, 2024Recent studies by OpenAI and Apple challenge AI model progress
Unveiling limitations in AI language models: Recent studies by Apple and OpenAI have exposed significant shortcomings in large language models (LLMs), challenging the notion that simply scaling up these systems will solve inherent issues. Apple's study reveals fragile mathematical reasoning: Apple researchers conducted an in-depth analysis of LLMs' ability to solve mathematical problems, uncovering concerning limitations. The study, titled "GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models," found that LLMs often fail when irrelevant details are added to math problems. This finding suggests that LLMs rely more on pattern matching than true logical reasoning, raising questions about...
read Oct 20, 2024Engineers claim new algorithm reduces AI power consumption by 95%
Breakthrough in AI energy efficiency: Engineers at BitEnergy AI have developed a new algorithm that could potentially reduce AI power consumption by up to 95%, marking a significant advancement in artificial intelligence processing technology. The new method, called Linear-Complexity Multiplication (L-Mul), replaces complex floating-point multiplication (FPM) with simpler integer addition while maintaining high accuracy and precision. This development addresses the growing concern of AI's increasing energy demands, which have become a primary constraint on AI advancement. Technical details and implications: The L-Mul algorithm represents a fundamental shift in how AI computations are performed, with far-reaching consequences for the industry and...
read Oct 18, 2024Meta shares massive AI materials database to accelerate research
Meta's game-changing contribution to materials science AI: Meta has released a massive open-source data set and AI models called Open Materials 2024 (OMat24) to accelerate the discovery of new materials using artificial intelligence. The big picture: The OMat24 release addresses a critical bottleneck in materials discovery by providing researchers with an extensive, high-quality data set and AI models that were previously unavailable or proprietary. Meta's decision to make OMat24 freely available and open-source stands in contrast to other industry players like Google and Microsoft, who have kept their competitive models and data sets secret. The data set contains approximately 110...
read Oct 18, 2024How AI can help find common ground in group deliberations
AI-assisted deliberation: A new frontier in democratic discourse: Google DeepMind researchers have developed an innovative AI system that could potentially transform how groups find common ground on complex social and political issues. The Habermas machine: An AI mediator for group discussions: The system, named after philosopher Jürgen Habermas, utilizes two large language models (LLMs) to generate and evaluate statements that reflect group views and areas of agreement. One LLM acts as a generative model, suggesting statements that capture collective opinions. The second LLM functions as a personalized reward model, scoring how likely participants are to agree with generated statements. The...
read Oct 17, 2024This prompt can extract sensitive info from chat conversations
Novel AI exploit targets personal data: Researchers have uncovered a sophisticated attack method called "Imprompter" that can covertly manipulate AI language models to extract sensitive information from chat conversations. The mechanics of the attack: Imprompter utilizes a clever algorithm to disguise malicious instructions within seemingly random characters, enabling it to bypass human detection while instructing AI systems to gather and transmit personal data. The attack transforms harmful prompts into hidden commands that appear as gibberish to human users but are interpreted as instructions by AI models. When successful, the AI collects personal information from conversations, formats it into a Markdown...
read Oct 16, 2024Brookings publishes new study projecting AI’s long-term fiscal impact
AI's potential fiscal impact: A wide-ranging forecast: A new study simulates the long-term effects of artificial intelligence on the U.S. federal budget, revealing a spectrum of possible outcomes that could significantly alter the nation's fiscal landscape. The research, conducted by Ben Harris, Neil R. Mehrotra, and Eric So, explores how AI might influence federal finances through four primary channels: mortality rates, healthcare pricing, healthcare demand, and overall productivity. Depending on the nature and extent of AI's impact, the study projects that annual budget deficits could either increase by 0.9% of GDP or decrease by as much as 3.8% of GDP....
read