Microsoft's Phi-3.5 AI Models Outshine Rivals: Google, Meta, and OpenAI Left Behind!
Anybody Can Prompt (ABCP)
ABCP: Making Generative AI easy for everyone. Join us, because Any Body Can Prompt!
Today's highlights:
?? AI Breakthroughs
Microsoft Releases Phi-3.5 AI Models with Specialized Capabilities for Reasoning and Vision Tasks
? Microsoft has unveiled the Phi-3.5 series, featuring specialized AI models for reasoning and vision tasks ?
? Phi-3.5-MoE-instruct, with 42 billion parameters, is designed for advanced reasoning and supports multiple languages ?
? Phi-3.5-vision-instruct excels in image and video analysis, demonstrating superior performance in visual benchmarks.
Grok-2 and Grok-Mini Climb to Top Spots in LMSys Chatbot Arena Leaderboard
? xAI's Grok-2 captures the #2 spot on the LMSys Chatbot Arena leaderboard, tying with Gemini, driven by over 6,000 votes ?
? Grok-2 excels in mathematical tasks, claiming the #1 position, and shows strong performance in coding and instruction-following tasks ?
? Following a major upgrade using SGLang, Grok-2-Mini now operates at double its previous speed, improving performance and efficiency.
Midjourney Rolls Out Free Trial, Enhances AI Image Generation Features
? Midjourney announced a temporary free trial to attract more users, enhancing access to its AI-driven image generation platform ?
? The platform update includes user interface improvements for easier navigation and advanced custom keywords for personalized image creation ?
? New features in version 6.1, such as improved image realism and faster processing speeds, reinforce Midjourney's leadership in AI-generated art.
OpenAI Forms Partnership with Condé Nast for Enhanced Content in ChatGPT and SearchGPT
? Condé Nast content from brands such as Vogue and Wired will be displayed in ChatGPT and the new SearchGPT prototype ?
? SearchGPT features new search capabilities, combining conversational models with web information, providing quick access to reliable sources and news links ?
? Feedback from news partners on SearchGPT's design and performance is being collected to enhance future updates and integration into ChatGPT.
Significant Price Reductions for Turbo AI Audio Models and New Business Plan
? Turbo Models, including Turbo v2 and v2.5, now offered at a 50% reduced price, enhancing affordability for high-quality, low-latency audio generation ?
? Updated pricing structure introduced with Self-Serve Plans starting at $50 per million characters and Enterprise Plans as low as $15 per million characters ?
? A new Business Plan launched, providing 11M credits per month, three personal voice clones, and priority support for $1,100 monthly.
Google AI Edge Enhances On-Device AI Performance with MediaPipe LLM Experiment
? Google AI Edge's MediaPipe recently launched an experimental API for on-device LLM inference using device GPUs across various platforms ?
? Initially, this system supported small LLMs up to 3 billion parameters, making them operable on Android, iOS, and web browsers ?
? A recent update has expanded this capability, now supporting models like Gemma 1.1 7B with 7 billion parameters, improving response quality significantly.
Google's HeAR Project Aims to Transform Health Monitoring Using AI and Bioacoustic Data
? Google Research introduces HeAR, a bioacoustic foundation model utilizing AI to potentially revolutionize disease screening using sounds like coughs ?
? Salcit Technologies adopts HeAR for enhancing TB detection capabilities in their Swaasa? product, aiming for broader accessibility and early diagnosis ?
? Global health organizations like The StopTB Partnership support HeAR for its potential to offer low-impact, widely accessible TB screening tools.
OpenAI Hires Former Meta Executive to Lead Strategic Initiatives
? OpenAI has appointed former Meta executive Irina Kofman to lead strategic initiatives, focusing initially on safety and preparedness ?
? The recruitment signifies a trend of AI startups, like OpenAI, hiring experienced leaders from major tech firms to bolster their competitive edge ?
? Recent high-profile departures from OpenAI include co-founder John Schulman and researcher Jan Leike, both joining competitor Anthropic.
How Computer Vision Aids Mosquito Identification and Malaria Prevention
? Computers are now capable of effectively understanding visual inputs, revolutionizing fields such as autonomous driving and medical imaging ?
? The VectorCam app, supported by Johns Hopkins University and the Gates Foundation, enables rapid identification of mosquito species using just a smartphone ?
? Uganda is testing VectorCam's effectiveness, streamlining mosquito data collection and aiding quicker strategic responses in malaria control efforts.
D-ID Launches AI Video Translate That Clones Voices, Syncs Lips in Multiple Languages
? D-ID launches AI Video Translate technology, allowing videos to be automatically translated into 30 languages, enhancing global reach for creators ?
? The new technology clones the speaker's voice and matches lip movements to the translated speech, aiming to streamline video localization for marketing and social media ?
? Available through D-ID Studio and its API, the service offers a one-month free trial with more demonstrations available on the company's website.
Salesforce Launches Autonomous AI Sales Agents to Transform Customer Engagement
? Salesforce introduces two autonomous AI sales agents: Einstein SDR Agent and Einstein Sales Coach Agent, set for release in October ?
? Einstein SDR Agent autonomously nurtures leads 24/7 with capabilities such as handling objections and booking meetings, enhancing seller productivity ?
? Einstein Sales Coach Agent provides real-time, AI-driven role-plays for sales reps, aiming to improve pitch and negotiation skills.
?? AI Ethics
Trump Uses AI-Generated Images of Taylor Swift and Fans for Endorsement Claims
? Donald Trump used AI to create fake endorsements from Taylor Swift and her fans, misleadingly suggesting their support for his presidential campaign ?
? Despite some fakes being tagged as "satire," Trump ambiguously accepted them as real on his Truth Social account, blurring the lines between humor and deception ?
? Trump’s posts raise concerns about the ease of generating convincing AI fakes, potentially enabling politicians to manipulate public perception without clear disclosures.
Authors Sue AI Firm Anthropic for Copyright Infringement Over Claude Models Training
? Three authors sue AI startup Anthropic for allegedly using pirated books to train language models ?
? Anthropic faces accusations of large-scale copyright infringement, potentially jeopardizing writer incomes and copyright norms ?
? The lawsuit could reshape legal boundaries for AI companies using copyrighted material without obtaining licenses.
OpenAI Opposes California AI Safety Bill Citing Stifled Innovation and National Security Concerns
? OpenAI opposes California's SB 1047, arguing it could stifle innovation and urging regulation at a federal level ?
? SB 1047, aimed to enforce safety in AI development, has sparked debate over potential talent exodus and stifling of innovation ?
? California's state assembly is set to vote on SB 1047, which could affect the state’s position as a global AI leader if passed.
Spotify and Meta CEOs Discuss Europe's Urgent Need for Open-Source AI Adoption
? Spotify and Meta advocate for open-source AI, asserting it democratizes technology and spurs economic opportunities globally ?
? European regulatory inconsistencies are criticized for stifling innovation and complicating access to new AI technologies ?
? A call for Europe to simplify and unify regulations to harness the full potential of AI and remain competitive on the global stage.
??AI Academia
Assessment of Data Privacy Risks in Large Language Models Introduced
? LLM-PBE is a new toolkit designed to evaluate data privacy risks in Large Language Models (LLMs), addressing data leakage concerns ?
? The tool analyzes privacy throughout the lifecycle of LLMs, employing various attack and defense strategies for comprehensive assessment ?
? Detailed experimentation with multiple LLMs helps identify critical influencing factors such as model size and data characteristics in privacy issues.
Study Evaluates Large Language Models on Diverse Clinical Tasks in Medicine
? MedS-Bench extends existing medical LLM benchmarks by covering 11 complex clinical tasks beyond typical MCQA ?
? The study reveals that even top-performing LLMs like GPT-4 and Claude-3.5 struggle with advanced medical tasks when tested ?
? The MedS-Ins dataset, a new large-scale instruction tuning set for medical applications, is now fully accessible for research purposes.
领英推荐
GenderCARE: New Framework to Assess, Mitigate Gender Bias in AI Models
? GenderCARE introduces novel benchmarks that include overlooked gender groups like transgender and non-binary people, aiming for comprehensive gender bias assessment in large language models (LLMs) ?
? The framework features new debiasing techniques using counterfactual data augmentation and fine-tuning strategies, promising over 90% peak reduction in gender bias without impacting overall LLM performance ?
? GenderCARE evaluation shows that reductions in gender bias across 17 different LLMs achieve an average of 35% while maintaining under 2% variability in mainstream language tasks.
New Whitepaper Details Ethical AI Certification and Trustworthy AI Implementation Guidelines
? The whitepaper provides normative guidelines for the integration of ethical principles into AI development to establish Trustworthy AI ?
? Six ethical principles including fairness, privacy, and sustainability have been detailed with actionable recommendations for AI systems ?
? It aligns with the EU AI Act requirements, offering a framework for AI system risk assessment and certification.
New Study Evaluates Large Language Models in Data-Driven and Text-Based Feature Selection
? Researchers categorize LLM-based feature selection into two main types: data-driven and text-based, each using unique approaches for semantical associations ?
? Extensive tests show that text-based feature selection via LLMs effectively enhances classification and regression tasks in diverse applications, including medical ?
? Challenges remain in applying LLMs for feature selection, yet the field holds promising future prospects for advancing data-centric feature selection methods.
Efficient Detection of Toxic Prompts in Large Language Models Achieves High Accuracy
? ToxicDetector, a new greybox approach, achieves a 96.39% accuracy in detecting toxic prompts in large language models, outperforming current methods ?
? With a processing time of just 0.0780 seconds per prompt, ToxicDetector offers a real-time solution, enhancing the practicality for applications involving chatbots and automated content creation ?
? Despite high diversity in toxic prompts and the use of jailbreaking attempts to evade detection, ToxicDetector maintains a low false positive rate of 2.00%.
New Method Combines Language Models with RAG for Enhanced Carbon Footprint Accounting
? A novel approach, LLMs-RAG-CFA, utilizes large language models and retrieval-augmented generation to advance real-time carbon footprint accounting;
? Experimental results show LLMs-RAG-CFA surpasses traditional methods in carbon footprint precision across industries like aluminum and new energy vehicles;
? The method balances efficiency with cost-effectiveness, providing a sustainable solution for managing real-time carbon emissions.
Challenges and Advanced Responses in Large Language Model Practices Analyzed
? Hongyin Zhu's paper reviews key AI challenges across industry trends, academic research, and business applications ?
? Explores cloud-edge-end architecture for efficient computing resource integration in scenarios like IoT and smart cities ?
? Discusses China's Xinchuang Plan impact on IT enterprise innovation and domestic market positioning.
Study Reveals AutoJailbreak Boosts Attack Success Rate on GPT-4V Exceeding 95%
? AutoJailbreak, a new technique for bypassing GPT-4V's safety mechanisms, achieves over 95% Attack Success Rate ?
? The method utilized advances prompt optimization and employs a strategic early stopping protocol to reduce time and resource use ?
? Researchers highlight the need for heightened security measures in response to potential privacy risks associated with GPT-4V's facial recognition capabilities.
Paper on Generative AI and Large Language Models' Development
? Recent advances in Generative AI and LLMs significantly enhance NLP capabilities, enabling new applications across multiple sectors ?
? The study addresses crucial challenges in Generative AI, including bias, fairness, and data security, urging responsible technology integration ?
? The research identifies and proposes solutions for research gaps in AI, aiming to foster ethical, and impactful technological advancements.
Study Explores How Informative Prompts Reduce Uncertainty in AI Responses
? Uncertainty in responses from large language models decreases with more informative prompts, aligning with principles of epistemic uncertainty ?
? A newly proposed prompt-response concept model by researchers shows how LLMs infer and generate output, clarifying prompt-response dynamics ?
? Experimental validation using real datasets confirms the proposed model's effectiveness in predicting and managing response uncertainty in LLMs.
Enhancing Honesty and Helpfulness in Large Language Models: A New Study and Dataset
? A recent study introduced HONESET, a novel dataset with 930 queries to test Large Language Models' (LLMs) honesty across six categories ?
? Two new methods to improve LLMs' honesty and helpfulness were presented, including a training-free approach using curiosity-driven prompts and a fine-tuning process based on curriculum learning ?
? Significant enhancements in honesty aligning were observed in LLMs like Llama3-8b and Mistral-7b, recording improvements of 65.3% and 124.7% respectively in the H2 assessment.
Large Language Models Prove Effective as Zero-Shot Predictors for Future Locations
? Large Language Models (LLMs) show potential as zero-shot predictors for future location visits with up to 36.2% accuracy ?
? LLMs outperform traditional models in predicting human mobility, marking an improvement of almost 640% ?
? Researchers highlight LLMs' capability to act as text-based explainers in explaining their predictions in next-location forecasting.
Survey Examines Retrieval-Augmented Text Generation in Large Language Models
? Retrieval-Augmented Generation (RAG) combines retrieval methods with deep learning to enhance the adaptability of large language models by incorporating real-time external data ?
? The study categorizes RAG into four stages—pre-retrieval, retrieval, post-retrieval, and generation—providing a structured approach to understanding its mechanisms ?
? Future research directions and evaluation methods for RAG are discussed, addressing current challenges and aiming to improve the accuracy and reliability of language model outputs.
Elephants Never Forget: Learning and Memorization in LLMs with Tabular Data
? Large Language Models (LLMs) demonstrate significant memorization of tabular data, impacting their performance on seen datasets ?
? The newly released 'tabmemcheck' Python package allows testing of LLMs for memorization of tabular datasets ?
? Despite advances, LLMs still lag in sample efficiency for few-shot learning compared to traditional statistical learning algorithms.
LLM2Vec: Transforming Large Language Models into Superior Text Encoders, Published at COLM 2024
? LLM2Vec transforms decoder-only large language models into effective text encoders using a simple unsupervised method ?
? The approach includes bidirectional attention, masked next token prediction, and unsupervised contrastive learning ?
? LLM2Vec achieves unprecedented performance on the Massive Text Embeddings Benchmark using only publicly available data.
Survey Reveals Extent of Contamination in Large Language Models and Introduces LLMSanitize Library
? Concerns rise over contamination in Large Language Models (LLMs), potentially jeopardizing their reliability and effectiveness in fields like medical and legal advice ?
? The LLMSanitize library, now available on GitHub, offers tools for the community to detect and analyze contamination levels in LLMs ?
? The survey differentiates between open-data and closed-data contamination detection, highlighting unique challenges in unearthing tainted data in undisclosed datasets.
New ChatSpamDetector Uses Large Language Models to Combat Phishing Emails
? ChatSpamDetector, developed by NTT Security Holdings, employs large language models to enhance phishing email detection accuracy ?
? The system provides detailed explanations for phishing alerts, helping users make informed decisions about suspicious emails ?
? In tests, the ChatSpamDetector demonstrated a 99.70% accuracy rate, surpassing several other large language models and baseline systems.
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!
Join our growing community of over 30,000 readers and stay at the forefront of the Generative AI revolution.
Yoga teacher & Occupational therapy student
3 个月Go D-ID! ??
Freelance Web Designer | Social Media Strategy | Florist
3 个月Breaking down language barriers-D-ID's video translation is a game-changer!
Chief Executive Officer at Sree Sastha Institute of Engineering and Technology
3 个月Hai