News/Research
AI Training Data Shortage Looms as Websites Block Crawlers
Web crawling restrictions reshape AI training landscape: The increasing use of robots.txt files to limit web crawler access is significantly impacting the availability of high-quality training data for generative AI models, potentially altering the future development of artificial intelligence. Generative AI models, which power popular tools like ChatGPT, rely heavily on vast datasets compiled from publicly available web data. A growing number of websites, particularly news outlets and artists' pages, are implementing restrictions on web crawlers to protect their content and livelihoods from AI exploitation. The Data Provenance Initiative's recent report highlights this trend, revealing a marked increase in crawled...
read Sep 2, 2024Claude’s AI Boosts Code Editing with New Inpainting Feature
AI-powered code editing gets a boost: Claude, an AI assistant developed by Anthropic, has received a significant update to its Artifacts feature, allowing users to highlight and edit specific lines of code within generated content. The new functionality brings inpainting capabilities, commonly used in AI image generation, to code editing within Claude's interface. Users can now select specific portions of generated code and request changes or explanations, streamlining the process of refining AI-generated content. This update addresses previous limitations where users had to reply to entire threads or manually copy and paste code sections to make targeted changes. Expanding Artifacts'...
read Aug 30, 2024NSF Invests $20M in AI to Transform Geosciences Research
NSF's $20 million investment in AI for geosciences: The U.S. National Science Foundation (NSF) has announced funding for 25 projects totaling over $20 million through the Collaborations in Artificial Intelligence and Geosciences (CAIG) program, aiming to advance AI techniques in geosciences research. The CAIG program seeks to foster transdisciplinary partnerships between geoscientists, computer scientists, mathematicians, and other experts to drive innovative discoveries and solutions in Earth sciences. This investment will support the development and implementation of cutting-edge AI techniques while expanding access to education and training opportunities for using AI in geosciences research. The funded projects align with key technology...
read Aug 30, 2024AI Adoption Hits Crossroads as Companies Struggle to Scale
Generative AI adoption reaches critical juncture: A new Deloitte report reveals that while two-thirds of companies are increasing their investment in generative AI, most efforts are still in early stages. Key findings on adoption and expectations: The survey of 2,770 director to C-suite level respondents across 14 countries shows a strong focus on efficiency and productivity improvements, but challenges in implementation and measuring impact persist. 54% of organizations are seeking efficiency and productivity improvements through generative AI However, only 38% are actively tracking changes in employee productivity 68% of respondents said their organization has moved 30% or fewer of their...
read Aug 29, 2024ChatGPT Unminifies JavaScript Code, Unveiling AI’s Development Potential
AI-powered code unminification reveals surprising capabilities: OpenAI's ChatGPT demonstrates an impressive ability to decipher and reconstruct minified JavaScript code, offering developers a powerful tool for code analysis and learning. The challenge of minified code: Frank Fiegel, while exploring an interesting component with running ASCII art, encountered minified code that was difficult to understand at first glance. Minified code is compressed to reduce file size, making it challenging for humans to read and comprehend. Traditionally, developers would either struggle through reading the minified code or search for source maps to restore the original version. ChatGPT's unexpected prowess: Fiegel decided to experiment...
read Aug 29, 2024Code-Trained AI Models Outperform in Non-Coding Tasks
The power of code in LLM training: New research from Cohere reveals that including code in the pre-training data of large language models (LLMs) significantly improves their performance on non-coding tasks. Researchers systematically investigated the impact of code data in LLM pre-training on general performance beyond coding tasks. The study used a two-phase training process: continued pre-training and a cooldown phase, testing various ratios of text and code in the training data. Models were evaluated at different scales, from 470 million to 2.8 billion parameters, using benchmarks for world knowledge, natural language reasoning, and code performance. Key findings: The inclusion...
read Aug 29, 2024Google AI Simulates Doom in Real-Time Without Original Code
AI-powered game simulation: Google researchers have developed an AI model called GameNGen that can simulate the classic PC game Doom in real-time without using the game's original code. The model uses AI image generation technology to create game visuals at over 20 frames per second, offering a playable experience. Researchers from Google and Tel Aviv University collaborated on this project, demonstrating that neural networks can run complex games in real-time with high quality. How GameNGen works: The AI model leverages Stable Diffusion version 1.4, an open-source image generator, to create game visuals based on player inputs and game state updates....
read Aug 27, 2024Experts Weigh In On The Future of Automated R&D
AI research automation: A growing frontier: The potential for artificial intelligence to automate its own research and development processes is emerging as a critical area of study, with significant implications for the pace of AI advancement. AI researchers are divided on the timeline for automating AI R&D tasks, reflecting the complexity and uncertainty surrounding this emerging field. A recent study interviewed eight AI researchers to gain insights into the nature of AI R&D work, automation predictions, and potential evaluation methods for AI systems' R&D capabilities. The findings highlight the diverse nature of AI R&D tasks and the challenges that must...
read Aug 27, 2024MIT Researchers Are Developing AI Models for Dance Choreography
AI meets dance: A fusion of technology and human movement: MIT researchers are exploring how artificial intelligence can interact with and enhance traditional dance practices, leading to innovative projects that blend cultural heritage with cutting-edge technology. MIT's 'dance with AI' project, led by researchers like Pat Pataranutaporn, aims to develop "choreography intelligence" by deconstructing traditional dances from around the world into teachable events for AI systems. The project seeks to create new choreography through AI-human collaboration, potentially breathing new life into cultural traditions and creating new forms of cultural heritage. Researchers are working on developing AI models that can interpret...
read Aug 27, 2024‘Model Collapse’ Has Experts Questioning Inevitability of AI Model Performance
AI performance decline raises concerns: Recent observations suggest that popular AI models like ChatGPT and Claude are experiencing a noticeable decrease in performance and accuracy, challenging the expectation of continuous improvement in AI technology. Steven Vaughan-Nichols, in a Computerworld opinion piece, highlights the erratic and often inaccurate responses from major AI platforms. Users on the OpenAI developer forum have reported a significant decline in accuracy following the release of the latest GPT version. One user expressed disappointment, stating that the AI's performance fell short of the surrounding hype. Potential causes of AI degradation: Several factors may contribute to the perceived...
read Aug 27, 2024DeepMind, Berkeley Show How to Make AI Models Better, Not Bigger
Optimizing LLM performance through inference-time compute: Researchers from DeepMind and UC Berkeley have explored innovative ways to enhance large language model (LLM) performance by strategically allocating compute resources during inference, potentially reducing the need for larger models or extensive pre-training. The study investigates how to maximize LLM performance using a fixed amount of inference-time compute, comparing different methods and their effectiveness against larger pre-trained models. This approach aims to enable the deployment of smaller LLMs while achieving comparable performance to larger, more computationally expensive models. Key strategies for inference-time compute optimization: The researchers focused on two main approaches to improve...
read Aug 27, 2024AI Predicts 70% of Earthquakes in 7-Month China Trial
Groundbreaking AI predicts earthquakes with unprecedented accuracy: A new artificial intelligence algorithm developed by researchers at the University of Texas at Austin has demonstrated remarkable success in predicting earthquakes, potentially revolutionizing earthquake preparedness and risk management. The AI system successfully predicted 70% of earthquakes during a seven-month trial in China, forecasting them a week in advance. The algorithm correctly predicted 14 earthquakes within approximately 200 miles of their estimated location and at almost exactly the calculated strength. It missed only one earthquake and gave eight false warnings, showcasing its high level of accuracy. Competition success and global implications: The University...
read Aug 26, 2024Stanford, CMU and Georgia Tech Develop AI Model for Mental Health Support
AI-powered peer counselor training: A collaborative effort between Stanford, Carnegie Mellon, and Georgia Tech has developed an AI model to provide feedback and improve the skills of novice peer counselors in emotional support conversations. The project, presented in a working paper accepted for the 2024 Association for Computational Linguistics conference, aims to address the growing demand for mental health support and the challenges in preparing peer counselors for their roles. Interdisciplinary collaboration between computer scientists and psychologists was crucial in developing this AI-assisted training model, combining expertise in both AI and counseling intervention skills. Developing a feedback framework: The research...
read Aug 25, 2024The ‘Plasticity Problem’ and LLM’s Inability to Learn Continuously
AI models face significant learning limitations: Recent research reveals that deep learning AI models, including large language models, struggle to incorporate new information without complete retraining. A study published in Nature by scientists from the University of Alberta highlights a major flaw in AI models' ability to learn continuously. Deep learning AI models, which find patterns in vast amounts of data, fail to function effectively in "continual learning settings" where new concepts are introduced to existing training. Attempting to teach an existing deep learning model new information often requires retraining it entirely from scratch. The plasticity problem: When new information...
read Aug 25, 2024How Lab-Grown Mini Brains Could Power AI of the Future
Breakthrough in biocomputing: Scientists are exploring the potential of brain organoids, tiny lab-grown neural structures, to power AI systems with unprecedented efficiency and sustainability. The AI energy crisis: Current artificial intelligence systems consume enormous amounts of energy, raising concerns about their long-term sustainability. • OpenAI's ChatGPT alone requires 500,000 kilowatts of power daily to process 200 million user prompts. • The scale of resources needed to fuel the global AI boom is becoming increasingly unsustainable. Biological solutions to technological problems: Researchers are turning to neuroscience and biotechnology to address the energy demands of AI systems. • Companies like Cortical Labs...
read Aug 23, 2024Stanford Awards 6 Groundbreaking AI Projects with $3M Grant
Stanford's AI research initiative expands: Stanford University's Institute for Human-Centered Artificial Intelligence (HAI) has awarded $3 million in Hoffman-Yee grants to six interdisciplinary research teams for cutting-edge AI projects. The grants, part of a program that has awarded over $20 million to date, aim to support bold, interdisciplinary AI research aligned with HAI's key focus areas. This year's recipients were chosen from 39 proposals representing all seven Stanford schools, highlighting the university's commitment to cross-disciplinary AI research. The selected projects span diverse fields, including neuroscience, genomics, law enforcement, visual media, and cellular biology. Brain mapping and genomic modeling: Two of...
read Aug 23, 2024New Research Suggests AI Models Can’t Learn as They Go Along
AI models face limitations in continuous learning: Recent research reveals that current artificial intelligence systems, including large language models like ChatGPT, are unable to update and learn from new data after their initial training phase. A study by researchers at the University of Alberta in Canada has uncovered an inherent problem in the design of AI models that prevents them from learning continuously. This limitation forces tech companies to spend billions of dollars training new models from scratch when new data becomes available. The inability to incorporate new knowledge after initial training has been a long-standing concern in the AI...
read Aug 23, 2024GPT-4 Matches Radiology Residents in Musculoskeletal Imaging Accuracy
Comparing ChatGPT and radiologists in musculoskeletal imaging: A recent study led by researchers from Osaka Metropolitan University's Graduate School of Medicine evaluated the diagnostic accuracy of ChatGPT against radiologists in musculoskeletal imaging cases. The study, conducted by Dr. Daisuke Horiuchi and Associate Professor Daiju Ueda, aimed to assess the potential of generative AI models like ChatGPT as diagnostic tools in radiology. Researchers analyzed 106 musculoskeletal radiology cases, including patient medical histories, images, and imaging findings. Two versions of the AI model, GPT-4 and GPT-4 with vision (GPT-4V), were used to generate diagnoses based on the case information. The same cases...
read Aug 23, 2024The Open Source Initiative Creates New Definition for Open-Source
Defining open-source AI: The Open Source Initiative (OSI) has unveiled a new definition for open-source AI systems, aiming to provide clarity in a field where the concept was previously ambiguous. The definition outlines key criteria for AI systems to be considered open-source, including unrestricted use, inspectability, modifiability, and shareability. Transparency requirements extend to training data, source code, and model weights, ensuring a comprehensive understanding of the AI system's components. The definition stipulates that sufficient information must be provided to allow a skilled person to recreate a substantially equivalent system using the same or similar data. Collaborative effort and development process:...
read Aug 21, 2024MIT Researchers Create AI That Can Oversee Humans and Enhance Team Dynamics
Breakthrough in AI-assisted collaboration: MIT CSAIL researchers have created an AI assistant that can oversee teams comprising both human and AI agents, intervening when necessary to improve teamwork efficiency. The system employs a theory of mind model to represent how humans think and understand each other's plans during cooperative tasks. By inferring team members' plans and their understanding of each other, the AI can intervene to align beliefs and actions when needed. The assistant can send messages about each agent's intentions or actions to ensure task completion and prevent duplicate efforts. Potential real-world applications: The AI assistant's capabilities show promise...
read Aug 21, 2024Meta’s Self-Taught Evaluator Allows AI Models to Create Their Own Training Data
Breakthrough in AI training: Meta's researchers have developed a Self-Taught Evaluator, a groundbreaking approach that enables large language models (LLMs) to create their own training data without human annotation. This innovative method addresses the significant challenge of expensive and time-consuming human evaluation for LLMs, potentially revolutionizing the way AI models are trained and evaluated. The Self-Taught Evaluator builds upon the concept of LLM-as-a-Judge, where the model itself evaluates responses to given prompts, eliminating the need for human intervention. Meta's approach marks a significant step towards more efficient and scalable AI development, particularly beneficial for enterprises with vast amounts of unlabeled...
read Aug 20, 2024Deloitte Survey Reveals Key Challenges Slowing Enterprise AI Deployment
Generative AI adoption accelerates amid enterprise challenges: A recent Deloitte survey of 2,770 business and technology leaders across 14 countries and 6 industries reveals the complex landscape of generative AI implementation in enterprise settings. The survey indicates a significant increase in generative AI investments, with 67% of organizations boosting their funding due to early perceived value. Despite this enthusiasm, only 32% of organizations have successfully moved more than 30% of their generative AI experiments into production, highlighting a notable gap between experimentation and full-scale deployment. Data management has emerged as a critical focus, with 75% of surveyed organizations increasing investments...
read Aug 20, 2024NVIDIA’s New AI Weather Model Boosts Short-Term Forecast Accuracy by 10%
Breakthrough in AI-powered weather prediction: NVIDIA Research has unveiled StormCast, a groundbreaking generative AI model designed to emulate high-fidelity atmospheric dynamics and enable reliable mesoscale weather prediction. Developed in collaboration with Lawrence Berkeley National Laboratory and the University of Washington, StormCast represents a significant advancement in AI-driven weather forecasting technology. The model focuses on mesoscale prediction, which covers an area larger than individual storms but smaller than cyclones, filling a crucial gap in current weather forecasting capabilities. StormCast builds upon NVIDIA's existing CorrDiff model by adding hourly autoregressive prediction capabilities, enhancing its ability to provide detailed and accurate short-term forecasts....
read Aug 19, 2024AI Humor Gap: Why Markov Chains Outshine LLMs in Comedy
The big picture: Markov chains, despite their simplicity, can produce more humorous content than advanced Large Language Models (LLMs) due to their unpredictable nature and ability to create unexpected combinations of words and phrases. What are Markov chains? Markov chains are primitive statistical models that predict the next word based on the current context, without considering semantics or complex vector math. They can be described as very small, simple, and naive LLMs Markov chains are commonly used in phone keyboards for next word suggestions While less accurate than LLMs for specific tasks, Markov chains excel in generating unexpected and potentially...
read