AI in Politics? Elon Musk shares a Deepfake video of Kamala Harris on X
Meta has released SAM 2 for advanced video and image segmentation, while a new AI model shows promise in predicting breast cancer progression. AI Studio's US launch enables users to create shareable..
Today's highlights:
🚀 AI Breakthroughs
Meta Releases SAM 2 for Advanced Video and Image Segmentation Under Open Source License
• Meta has unveiled SAM 2, an advanced AI model for real-time, promptable object segmentation in both images and videos
• SAM 2 operates under an Apache 2.0 license with its code, weights, and SA-V dataset shared openly to foster further innovation in AI
• The model promises substantial advancements in video segmentation, requiring three times less interaction time than its predecessors and enabling new applications in fields such as autonomous vehicles and video editing.
Gemma Scope: New Tools to Decode Language Models with Sparse Autoencoders
• Gemma Scope, a new suite of open sparse autoencoders, enhances the interpretability of language models by shedding light on their complex inner workings
• Researchers can now access hundreds of sparse autoencoders with Gemma Scope for free, facilitating the study of advanced AI features in Gemma 2 9B and Gemma 2 2B models
• By employing the sophisticated JumpReLU SAE architecture in Gemma Scope, the decoding of layered, complex algorithms in larger models is now more achievable and precise.
AI Startup Friend’s $99 necklace help combat loneliness
• Avi Schiffmann launches "Friend," a neck-worn AI companion device aimed at combating loneliness, with preorders starting at $99 for January 2025 delivery
• Schiffmann's startup pivots from a $600 pendant aimed at transcription to "Friend," focusing on emotional support, without storing user conversations.
• Mixed responses for AI hardware startups with Rabbit and Humane facing disappointing launches, while a16z-backed Limitless and Exor-backed Bee AI develop new AI wearables
AI Studio Launches in the US, Enabling Users to Create and Share AI Characters
• AI Studio is now rolling out in the US on ai.meta.com/ai-studio and Instagram, supporting the creation and sharing of personalized AI characters
• Users can craft AI extensions for direct interaction on platforms like Instagram and WhatsApp, enhancing fan engagement and content personalization
• Meta introduces step-by-step guides and customization tools within AI Studio to ensure creators maintain control and transparency in AI interactions.
Swiggy's Hermes AI Enhances Data Accessibility and Decision-Making Speed
• Swiggy introduces Hermes, an in-house generative AI workflow enabling natural language queries within Slack to streamline data-driven decision making
• Hermes leverages advanced AI to produce SQL queries and swiftly execute them, significantly boosting efficiency and reducing dependency on technical analysts
• Sophisticated metadata management in Hermes V2 enhances model accuracy, helping Swiggy to customize data queries specific to distinct business units.
New AI Model Predicts Breast Cancer Progression Using Simple Tissue Images
• An AI model developed by MIT and ETH Zurich researchers accurately identifies stages of DCIS from affordable breast tissue images
• The model's innovative approach highlights the significance of both the state and spatial arrangement of cells for accurate cancer stage diagnosis
• Potential future applications of the model include aiding clinical diagnosis processes and adapting for other cancer types or neurodegenerative diseases.
⚖️ AI Ethics
Elon Musk Criticized for Sharing Deepfake Video of Kamala Harris, Violating Platform Policies
• Elon Musk reposted a deepfake video of Kamala Harris, gaining over 130 million views, without indicating it was a parody
• This post appears to breach X’s policies against deceptive media, which aim to prevent harm and confusion
• Critics, including tech campaigners and political figures, argue that Musk's actions could fuel AI-driven political disinformation ahead of the US election.
Commerce Department Releases New NIST Guidance for AI Safety Evaluation and Mitigation
• New NIST draft guidance focuses on evaluating and mitigating risks from generative AI and dual-use foundation models
• U.S. AI Safety Institute's new software package tests AI systems against adversarial attacks to assess performance degradation
• Updated USPTO guidance on patent eligibility aims to foster innovation within AI and other emerging technologies.
OpenAI May Need to Raise Funds as It Faces $5 Billion in Potential Losses This Year
• OpenAI projected to spend $7 billion this year on AI development, including $4 billion on Microsoft server rentals
• An additional $3 billion estimated for training AI models with newly acquired copyrighted content
• OpenAI may seek further funding within 12 months due to significant operational and labor costs.
🎓AI Academia
Google's ShieldGemma Advances AI Safety with State-of-the-Art Moderation Models
• ShieldGemma, a new AI content moderation suite, reportedly surpasses existing models like Llama Guard by 10.8% in accuracy on public benchmarks
• Google releases a novel LLM-based data curation pipeline with ShieldGemma, aimed at enhancing safety in AI-mediated environments
• ShieldGemma is demonstrated to effectively moderate harmful content types, including sexual content and hate speech, across different digital platforms.
New Automated Review Generation Method Employs Large Language Models for Swift Literature Analysis
• A new automated review generation method using Large Language Models streamlines literature analysis, significantly reducing cognitive load
• The method efficiently synthesizes insights from hundreds of scientific articles, showcasing deep analysis of catalysts’ composition and performance
• Enhanced with a multi-layered quality control system, this new tool demonstrates highly reliable output with minimized error rates below 0.5%.
AI-Generated Image Detection Faces Challenges Due to New Adversarial Attack Techniques
• Recent study uncovers vulnerabilities in AI-Generated Image Detectors, susceptible to novel adversarial attacks using frequency domain perturbations
• A new attack method called Frequency-Based Post-Train Bayesian Attack (FPBA) significantly challenges existing detectors across various models and defense methods
• FPBA proves effective in black-box scenarios, indicating a significant risk of adversarial attacks evading cross-generator AI-generated image detection.
Survey on Machine Unlearning in Generative AI Reveals New Techniques and Challenges
• Machine Unlearning (MU) is developed to help Generative AI models forget undesirable data, enhancing data privacy and security
• The study presents a new problem formulation for MU, assessment techniques, and discusses the benefits and constraints of various MU methods
• Research highlights crucial challenges and promising future directions for MU in Generative AI, including a curated list of key readings available online.
StockAgent: Advanced AI Simulates Real-World Stock Trading Scenarios for Enhanced Market Analysis
• StockAgent, a multi-agent AI system, uses LLMs to simulate real-world stock trading scenarios and analyze investor responses to market changes
• This AI-driven platform offers insights into how external factors like macroeconomics and global events affect trading behaviors and stock market dynamics
• Designed to avoid test set leakage, StockAgent provides a more reliable tool for financial research and investment strategy development, enhancing market prediction accuracy.
AttackEval Framework Assesses Jailbreak Attack Effectiveness on Large Language Models
• AttackEval introduces a framework to evaluate jailbreak attack effectiveness on large language models by assessing the attacking prompts themselves
• The method features both coarse-grained and fine-grained evaluations, utilizing a scoring range from 0 to 1 to offer detailed insights into different attack scenarios
• A comprehensive ground truth dataset tailored for jailbreak prompts has been developed, providing a crucial benchmark for current and future research in prompt injection security.
Comprehensive Survey on Model Compression Techniques for Large Language Models
• Recent survey highlights advanced model compression techniques for Large Language Models, contributing to efficiency in resource-constrained settings
• Discussion includes quantization, pruning, knowledge distillation, essential for evaluating compressed LLM performance metrics
• Study advocates for community collaboration towards sustainable and efficient future developments in LLM technology.
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!
