OpenAI announces a Search Engine called 'SearchGPT' to directly compete with Google
OpenAI challenges Google with SearchGPT, while advanced AI systems achieve silver medal status in the International Mathematical Olympiad. Gemini 1.5 Flash enhances productivity with faster responses
Today's highlights:
🚀 AI Breakthroughs
OpenAI Challenges Google's Search Supremacy with New AI-Powered Engine, SearchGPT
• OpenAI challenges Google with the debut of SearchGPT, an AI-driven search engine integrating web data and partner information.;
• SearchGPT, built on OpenAI's generative AI, competes against Google and Microsoft's Bing, aiming to rejuvenate the search engine landscape.;
• The prototype of SearchGPT, aimed at refining user interactions with AI, allows real-time, context-aware responses and sources citing, available for a limited tester base.
Advanced AI Achieves Silver Medal Standard in International Mathematical Olympiad
• AlphaProof and AlphaGeometry 2 AI systems solved four out of six problems at the International Mathematical Olympiad, achieving silver-medal status
• AlphaProof utilizes a combination of a pre-trained language model and AlphaZero reinforcement learning to prove complex mathematical statements
• AlphaGeometry 2, improved from its predecessor, now solves advanced geometry problems faster with high accuracy.
Gemini 1.5 Flash Enhances Productivity with Faster Responses and Expanded Features
• Gemini upgrades its free-tier service with 1.5 Flash, featuring enhanced quality, reduced response times, and improved reasoning and image understanding
• The platform now supports a larger context window of 32K tokens, enabling more complex queries and extended conversations without additional costs
• Gemini expands functionality for file uploads from Google Drive or direct from devices, facilitating the creation of customized practice questions and data analysis.
Apple's AI Features Delayed, Missing Initial iPhone and iPad Software Launch
• Apple's new AI feature, Apple Intelligence, is delayed and will miss the initial software release of iOS 18 and iPadOS 18 in September
• The Apple Intelligence AI feature is expected to be rolled out with software updates targeted for October, following the release of iOS 18 and iPadOS 18
• Apple plans to release early versions of these AI capabilities for developers by launching the iOS 18.1 and iPadOS 18.1 betas as early as this week.
⚖️ AI Ethics
Runway's AI Video Tool Secretly Trained with YouTube, Pirated Content: Report
• Runway's AI video generation tool, Gen-3, was trained using videos scraped from YouTube and pirated movies, causing ethical and legal concerns
• The internal training dataset reportedly includes content from major brands like Disney and Netflix, and from piracy sites like KissCartoon
• Despite claims of using "curated, internal datasets," Runway's actual data sourcing practices for Gen-3 remain under scrutiny, unconfirmed by top executives.
Equity Joins SAG-AFTRA in Strike Against Major Video Game Studios Over AI Concerns
• British union Equity expresses full support for SAG-AFTRA's strike against U.S. video game studios over AI and contract issues
• Equity calls for greater transparency and regulation in the use of AI in the entertainment industry to protect performers
• Despite significant tax relief to game companies, Equity highlights stagnant pay in a rapidly growing industry, urging dialogue and resolution.
🎓AI Academia
UC Berkeley and Google Research Achieve Breakthrough in 4D Scene Reconstruction from Single Video
• UC Berkeley and Google Research collaboratively developed a new method for 4D reconstruction from single video sources
• The technique uses SE(3) motion bases and data-driven priors to capture long-range 3D/2D motion and perform novel view synthesis on dynamic scenes
• This research provides a solution to the monocular dynamic reconstruction challenge, which has long been limited by the requirement for additional sensors or multiple camera setups.
Comparative Study Reveals MLP Outperforms KAN in Multiple AI Tasks Except Symbolic Representation
• KAN and MLP models were compared based on equal parameters and FLOPs revealing that MLP consistently outperforms KAN across most tasks in ML, CV, NLP, and audio processing
• The one exception noted was in symbolic formula representation, where KAN demonstrated superior performance, attributed to its unique B-spline activation function
• Potential limitations of KAN were highlighted, including a severe forgetting issue in standard class-incremental continual learning settings, contrasting earlier reports.
Study Reveals Gender Bias in AI-Generated Images of Construction Workers
• AI models like DALL-E demonstrate gender bias by sexualizing female figures and presenting males as more authoritative
• Research highlights the importance of considering ethical implications in AI's role within creative industries
• Study emphasizes AI's tendency to reflect and potentially exacerbate existing societal gender biases.
Survey Explores How Large Language Models Can Safely Refuse to Answer
• Abstention in large language models (LLMs) is studied for its potential to reduce hallucinations and increase safety
• The survey introduces a framework evaluating abstention from perspectives of the query, the model, and human values
• Future research will aim to apply abstention across various tasks and contexts to improve AI system robustness.
Enhancing Stock Return Predictions Through Fine-Tuned Large Language Models on Financial Newsflow
• Large language models (LLMs), fine-tuned for stock return forecasting using newsflow, show enhanced portfolio performance
• Decoder-only LLMs outperform others in large investment universes, while no consistent leader emerges in smaller ones
• LLM-generated text representations for return predictions outstrip conventional sentiment scores in portfolio construction.
Large Language Models Outperform Analysts in Financial Statement Predictions
• GPT4 outperforms financial analysts in predicting future earnings from standardized financial statements, without narrative or industry-specific context;
• Large Language Models (LLMs) show high accuracy, matching narrowly trained state-of-the-art machine learning models in financial statement analysis;
• Trading strategies based on GPT4’s financial predictions yield higher Sharpe ratios and alphas than other model-based strategies.
Generative AI Faces Challenges: Autophagy and Data Integrity Issues Explored
• Generative AI now frequently uses self-produced synthetic data to reduce training costs, complicating the balance between real and synthetic data use
• Increased contamination of web-scraped datasets by unlabelled synthetic data raises concerns about the reliability of training datasets
• The emerging challenge is maintaining control over the integration of real and synthetic data, with risks of uncontrollable outcomes.
Survey Highlights Large Language Models in Automating Table-Centric Tasks
• A new survey reviews the use of Large Language Models for automating table-centric tasks such as database queries and spreadsheet manipulations
• The study highlights the development of training techniques and prompt engineering for both LLMs and Visual Language Models tailored to enhance table processing
• Challenges such as processing implicit user intentions and extracting information from diverse table sources are emphasized in recent findings.
About ABCP: We are dedicated to reducing Generative AI anxiety among tech enthusiasts by providing timely, well-structured, and concise updates on the latest developments in Generative AI through our AI-driven news platform, ABCP - Anybody Can Prompt!
