News/Research
LLM Progress Slows — What Does It Mean for AI?
The rapid advancements in large language models (LLMs) that have dominated AI headlines in recent years appear to be slowing, with potential far-reaching implications for the future of artificial intelligence development and innovation. Slowing progress in LLMs: OpenAI's releases of increasingly capable language models have shown diminishing returns with each new version, signaling a potential plateau in general-purpose LLM development. The improvements between GPT-3 and GPT-4 were less dramatic than those seen between earlier iterations, suggesting a slowdown in the pace of advancement. Other major players in the AI field, including Anthropic and Google, are producing LLMs with capabilities converging...
read Aug 11, 2024Meta and Oxford Develop AI Model that Turns Images to 3D Objects
VFusion3D, a groundbreaking AI model developed by Meta and the University of Oxford, represents a significant advancement in AI-powered 3D content creation, enabling the generation of high-quality 3D objects from single images or text descriptions. Innovative approach to 3D generation: VFusion3D overcomes the challenge of limited 3D training data by leveraging pre-trained video AI models to generate synthetic 3D data. This novel approach allows the model to create high-quality 3D objects without relying on extensive 3D datasets, which have been a bottleneck in previous attempts at AI-driven 3D generation. The use of video AI models for 3D data synthesis demonstrates...
read Aug 11, 2024New Game Theory Research Suggests How Humans May Bias AI Model Training
The discovery that people alter their behavior when knowingly training AI systems raises important questions about the potential introduction of biases and the effectiveness of human-in-the-loop AI training methods. Study methodology and key findings: Researchers at Washington University in St. Louis conducted a game theory experiment to examine how people's decision-making changes when they believe they are training an AI system. The study utilized a classic game theory setup where participants could accept or reject monetary offers from a partner. Some participants were informed that their partner was an AI being trained through their interactions. Results showed that people were...
read Aug 10, 2024New Research Shows How AI Agents Can Learn Faster with Less Data
Researchers at Imperial College London and Google DeepMind have introduced a groundbreaking framework called Diffusion Augmented Agents (DAAG) to enhance the learning efficiency and transfer learning capabilities of embodied AI agents, addressing the critical challenge of data scarcity in training these agents to interact with the physical world. The DAAG framework: A novel approach to embodied AI learning: DAAG combines large language models (LLMs), vision language models (VLMs), and diffusion models to create a powerful lifelong learning system for embodied agents. The framework is designed to enable agents to continuously learn and adapt to new tasks, making more efficient use...
read Aug 10, 2024AI Breakthrough Turns Scanned Docs into Machine-Readable Text
Innovative OCR enhancement through AI: The LLM-Aided OCR Project represents a significant advancement in Optical Character Recognition technology by integrating large language models to improve accuracy and readability of digitized text. The project combines traditional OCR techniques with state-of-the-art natural language processing to transform raw scanned text into high-quality, well-formatted documents. Key features include PDF to image conversion, Tesseract OCR integration, and advanced error correction using both local and cloud-based LLMs. The system offers flexible configuration options, including markdown formatting and the ability to suppress headers and page numbers. Technical architecture and processing pipeline: The LLM-Aided OCR Project employs a...
read Aug 9, 2024How Mathematical Data May Help Solve the AI Training Data Shortage
Artificial intelligence's insatiable appetite for data has raised concerns about potential limitations on its future growth, but a compelling argument suggests these worries may be unfounded due to the infinite nature of mathematics. The big picture: The notion of running out of data for AI training overlooks the vast potential of mathematical data as an inexhaustible resource for fueling AI advancement. Experts have expressed concern that the finite amount of text and images available for AI training could hinder future progress. This perspective fails to consider the unlimited potential of mathematical data to supplement and expand training resources. Mathematical data...
read Aug 8, 2024Google DeepMind AI Robot Masters Table Tennis, Challenges Human Players
An AI-powered robotic table tennis player developed by Google DeepMind is demonstrating amateur-level skills, winning 45% of matches against human opponents in a recent study. This breakthrough combines advanced robotics with sophisticated AI software, showcasing the potential for AI-driven systems to master complex physical tasks and interact dynamically with humans in real-world scenarios. System architecture and capabilities: Google DeepMind's robotic table tennis player integrates an industrial robot arm with custom AI software, employing a two-tiered approach to gameplay. The system utilizes an ABB IRB 1100 industrial robot arm as its physical component, providing the necessary precision and speed for table...
read Aug 8, 2024AI Model Generates Cognitive Maps From Visual Data Alone
Cognitive maps, a cornerstone of spatial navigation and memory, have long fascinated researchers in neuroscience and artificial intelligence. A groundbreaking study published in Nature Machine Intelligence demonstrates how a self-attention neural network can generate environmental maps from visual inputs alone, potentially shedding light on both biological and artificial spatial cognition processes. Revolutionary approach to spatial mapping: The study introduces a computational model that constructs cognitive map-like representations solely from visual inputs, without relying on explicit spatial information. This breakthrough addresses a significant challenge in both neuroscience and AI: the ability to create accurate spatial maps from sensory inputs. The model's...
read Aug 8, 2024Stanford Scientist Claims His Facial Scans Can Predict Your Intelligence and More
Facial recognition AI technology has advanced to the point where it can potentially infer sensitive personal characteristics from images, raising significant ethical and privacy concerns. Controversial AI research claims: Stanford University psychologist Michal Kosinski has developed an AI system that he claims can detect intelligence, sexual preferences, and political leanings from facial scans. Kosinski's 2021 study reported that his model could predict political beliefs with 72% accuracy based solely on photographs. A 2017 paper by Kosinski claimed 91% accuracy in predicting sexual orientation from facial images, sparking controversy and criticism. The researcher asserts that his work is intended as a...
read Aug 5, 2024AI Breakthrough Enhances Scientific Discovery and Interpretability
Kolmogorov-Arnold Networks (KANs) represent a significant advancement in artificial neural network technology, offering improved interpretability and accuracy compared to traditional models. This novel approach, developed by researchers at MIT and other institutions, has the potential to revolutionize how AI systems process and represent data, particularly in scientific and mathematical domains. A new paradigm in neural network architecture: KANs utilize a fundamentally different structure where synapses learn functions instead of simple weights, marking a departure from conventional neural network designs. This innovative approach allows KANs to represent complex relationships more efficiently, potentially leading to more accurate and interpretable models. The architecture...
read Aug 5, 2024AI Adoption Surges Despite Trust and Ethics Concerns, Salesforce Report Finds
The big picture: Salesforce's "Trends in AI for CRM" report reveals a growing appetite for AI adoption among businesses, coupled with significant concerns about data trust, ethics, and organizational readiness. Nearly half of customer service teams, over 40% of salespeople, and a third of marketers have already implemented AI in their operations. Salesforce projects that AI could generate over $2 trillion in new business revenues by 2028, highlighting the technology's immense potential. Key findings on AI implementation: The report uncovers a mix of enthusiasm and caution among businesses regarding AI adoption and its impact on productivity. 80% of employees using...
read Aug 5, 2024AI Breakthrough Slashes Neural Network Size Without Losing Accuracy
Self-compressing neural networks offer a promising approach to reducing the size and resource requirements of AI models while maintaining performance. This new technique, developed by researchers Szabolcs Cséfalvay and James Imber, addresses key challenges in making neural networks more efficient and deployable. The big picture: Self-compression aims to simultaneously reduce the number of weights in a neural network and minimize the bits required to represent those weights, potentially revolutionizing the efficiency of AI models. The method utilizes a generalized loss function to optimize overall network size during training. Experimental results show that self-compression can achieve floating point accuracy with only...
read Aug 5, 2024New AI Breakthrough Enables Robots to Express Human-Like Emotions
Chinese scientists have made a breakthrough in humanoid robot technology, developing a method that allows robots to express emotions more naturally and accurately. This advancement could significantly enhance the potential applications of humanoid robots in various fields, particularly in scenarios where emotional interaction is crucial. The innovation: Chinese researchers from Hohai University and Changzhou University have created a two-stage method to enable humanoid robots to display more complex and lifelike facial expressions. The new AI system generates detailed examples of facial expressions, which are then learned and performed by a specially designed robot with multiple degrees of freedom for facial...
read Aug 1, 2024Researchers Are Exploring How AI Could Help with End-of-Life Decisions
A new AI tool, which hasn't been built, but whose potential benefits are under careful consideration, aims to help with end-of-life decisions. The challenge of end-of-life decision-making: When patients are unable to make their own medical decisions, the responsibility falls on surrogates, often family members, which can be emotionally distressing and lead to disagreements: In the case of Sophie, a woman in her mid-50s who suffered a hemorrhagic stroke, her family members couldn't agree on whether to continue medical treatments or let her die peacefully. Surrogates often struggle with the burden of making life-or-death decisions on behalf of their loved...
read Jul 31, 2024Deep Learning Enables Eavesdropping on Digital Video Displays
Eavesdropping on digital video displays through electromagnetic emanations: Researchers have developed a deep learning-based system called DEEP-TEMPEST that can effectively eavesdrop on digital video displays, such as HDMI, by analyzing the unintentional electromagnetic waves emanating from cables and connectors: The digital case, particularly HDMI, poses a greater challenge compared to analog (VGA) due to a 10-bit encoding that results in a larger bandwidth and a non-linear mapping between the observed signal and pixel intensity. Existing eavesdropping systems designed for analog video obtain unclear and difficult-to-read images when applied to digital video, necessitating a new approach. Deep learning as a solution:...
read Jul 31, 2024Model Complexity and High Costs Remain Barrier to Enterprise AI Adoption
The rapid growth of generative AI is transforming the enterprise landscape, but CEOs must navigate cost, complexity, and optimization challenges to harness its full potential. A new IBM report, based on a survey of U.S. executives, provides insights into the current state of enterprise AI adoption and offers guidance for informed decision-making. Key Takeaways: Specialization and diversity are crucial in enterprise AI deployment; The report emphasizes the importance of task-specific model selection, debunking the myth of a universal AI model: Organizations currently use an average of 11 different AI models and expect a 50% increase within three years, highlighting the...
read Jul 31, 2024Galileo Benchmark Shows Open-Source AI Models Challenging Proprietary Dominance
Galileo's latest benchmark reveals open-source AI models are rapidly catching up to their proprietary counterparts, potentially democratizing advanced AI and accelerating innovation across industries. Shifting AI landscape: The second annual Hallucination Index from Galileo evaluated 22 leading large language models, revealing that the performance gap between open-source and proprietary models has narrowed significantly in just eight months: Anthropic's Claude 3.5 Sonnet outperformed offerings from OpenAI, which dominated last year's rankings, indicating a changing of the guard in the AI arms race. Google's Gemini 1.5 Flash emerged as the most cost-effective option, delivering strong results at a fraction of the price...
read Jul 31, 2024Mentioning AI in Products May Reduce Consumer Trust and Purchase Intent
The increasing use of AI in products may not always lead to higher sales, as mentioning the technology can reduce consumer trust and purchase intent. Key findings: A study led by Washington State University researchers reveals that including the term "artificial intelligence" in product descriptions consistently lowered consumers' likelihood of buying those products: Across experiments involving over 1,000 U.S. adults, products described as using AI were consistently less popular compared to identical offerings without the AI mention. The negative impact of AI disclosure on purchase intent was mediated by lower levels of emotional trust, suggesting consumers feel less confident about...
read Jul 31, 2024MIT’s Thermometer Method Efficiently Calibrates Large Language Models
Researchers from MIT and the MIT-IBM Watson AI Lab have developed a new method called Thermometer to efficiently calibrate large language models (LLMs) and improve their reliability across diverse tasks. Key Takeaways: Thermometer addresses the challenges of calibrating LLMs, which can generate inaccurate responses and exhibit misaligned confidence levels: The method involves building a smaller, auxiliary model that runs on top of an LLM to calibrate it, enabling better-calibrated responses on unseen tasks. Thermometer is more efficient than other approaches, requiring less power-hungry computation while preserving the model's accuracy. Understanding the Problem: LLMs can be applied to a wide range...
read Jul 30, 2024JPMorgan Releases In-House Chatbot for AI-Enabled Research
JPMorgan Chase unveils LLM Suite, a generative AI platform for asset and wealth management employees, underscoring AI's growing influence in the financial industry and its potential to transform traditional roles. Key details of LLM Suite: JPMorgan's new AI tool is set to revolutionize how employees work by assisting with various tasks and providing information and advice: LLM Suite is described as a "ChatGPT-like product" designed for "general purpose productivity" in an internal memo signed by key executives. The AI-enabled research platform can perform functions such as writing, generating ideas, and document summarization, essentially acting as a "research analyst." It is...
read Jul 28, 2024New Algorithm Makes Fabrics Look Real in Games and Immersive Experiences
A new lightweight neural network proposed in a recent paper enables highly realistic rendering of woven fabrics in real-time, improving the realism of virtual environments and digital media. Key innovations: The algorithm leverages the repetitive patterns of woven fabrics to efficiently encode and render them: The network first encodes fabric patterns into a compact latent vector. A small decoder then interprets this vector to generate realistic fabric representations. Real-time performance: Despite its small size, the network can render and edit fabrics at an impressive speed of 60 frames per second on a high-end graphics card: This real-time capability is a...
read Jul 26, 2024Decoding the Black Box: The Quest to Understand How AI Really Works
A new branch of computer science aims to shed light on how artificial intelligence works: Key insight: Scientists are trying to understand the inner workings of large language models (LLMs) like ChatGPT and Claude, which are driving recent AI breakthroughs, by studying the algorithms that power them in a new field called AI interpretability. Researchers liken the challenge to studying the human brain - an extremely complex system where the activity of many neurons together produces intelligent behavior that can't be explained by looking at individual neurons alone. Unlike with the human brain, AI researchers have complete access to every...
read Jul 26, 2024DeepMind’s JumpReLU Architecture Sheds Light on the Inner Workings of Language Models
DeepMind has made significant progress in interpreting large language models (LLMs) with the introduction of JumpReLU sparse autoencoder (SAE), a deep learning architecture that decomposes the complex activations of LLMs into smaller, more understandable components. The challenge of interpreting LLMs: Understanding how the billions of neurons in LLMs work together to process and generate language is extremely difficult due to the complex activation patterns across the network: Individual neurons don't necessarily correspond to specific concepts, with a single neuron potentially activating for thousands of different concepts, and a single concept activating a broad range of neurons. The massive scale of...
read Jul 26, 2024“Copyright Traps” Helps Creators Detect AI Training on Their Work
Researchers have developed a new tool that could help content creators prove their work has been used to train AI models without their consent. The method, called "copyright traps," allows writers to subtly mark their text in a way that can later be detected in AI training data. Key points about the copyright trap tool: The copyright traps work by injecting long, gibberish sentences multiple times into a piece of text, which can then be used to determine if that text was used to train an AI model: The trap sentences are generated using a word generator and chosen randomly...
read