Elevenlabs is bringing back the voices of famous dead celebrities. Listen to how it sounds.
++Kyutai Labs Unveils Free Moshi AI, Cloudflare's New Anti-AI Easy Button Shakes Up Web Security, Ola Cabs Ditches Google Maps, YouTube Tightens Privacy, and More Tech Innovations Unleashed
Today's highlights:
🚀 AI Breakthroughs
Kyutai Labs Launches Moshi AI: A Free, Real-Time, Emotive Voice Chatbot
• Kyutai Labs unveiled Moshi AI, a real-time verbal response AI developed entirely in-house, capable of expressing emotions and diverse speaking styles
• Moshi AI is publicly available for free use with a maximum conversation duration of five minutes, featuring minimal latency in response
• The AI's architecture will eventually be open-sourced, allowing users to download and install it for offline use on their own devices.
ElevenLabs partners with estates of iconic stars to bring their voices to the Reader App
• Legendary voices of Judy Garland, James Dean, and others have been added to the Reader App, enabling users to hear digital texts in familiar tones
• Liza Minnelli endorses the use of her mother Judy Garland's voice in the app, highlighting its potential to attract both new fans and please long-time admirers
• The Reader App launched last week transforms text from articles, e-books, and more into context-aware, emotionally rich voiceovers using AI technology.
Perplexity Enhances Pro Search Tools for Advanced Problem Solving and Comprehensive Research
• Pro Search has been enhanced to handle complex queries and advanced computations in math and programming
• The new version features multi-step reasoning, improving its ability to plan and execute searches that involve multiple stages
• Pro Search now differentiates itself from Quick Search by providing detailed analyses and comprehensive answers for intensive research needs.
Cloudflare Launches Easy Button to Block All AI Bots for Enhanced Web Security
• Cloudflare launches a new feature allowing all customers, even on the free tier, to block all AI bots with just one click
• Top AI bots like Bytespider, Amazonbot, ClaudeBot, and GPTBot dominate request volumes on Cloudflare, indicating significant scraping activity
• Cloudflare's advanced machine learning models and global signals are used to continuously detect and block new evasive AI bots to protect content creators.
GraphRAG Now Available on GitHub with Enhanced Question-Answering Capabilities
• GraphRAG, a graph-based retrieval-augmented generation technology, is now available on GitHub, enhancing question-answering capabilities over diverse datasets
• The technology features a solution accelerator hosted on Azure, enabling users to deploy it without coding and in just a few clicks
• GraphRAG outperforms naive RAG methods by using community summaries that consider all data from a dataset, providing more comprehensive and diverse answers.
⚖️ AI Ethics
Ola Cabs Switches to In-House Ola Maps, Exits Google Maps to Save Costs Annually
• Ola Cabs has transitioned to its proprietary Ola Maps, significantly reducing operational costs by nearly ₹100 crore annually
• The shift follows Ola Group's move to Krutrim, its in-house AI firm, after detaching from Microsoft Azure
• Future enhancements for Ola Maps include street view, indoor images, NERFs, drone maps, and 3D mapping capabilities.
Google's Rising Greenhouse Gas Emissions Challenge AI-Centric Climate Goals
• Google's greenhouse gas emissions surged by 48% since 2019, complicating its goal to halve emissions by 2030
• The company attributes this increase largely to its energy-intensive data centers and expanding AI-driven projects
• Despite the challenges, Google aims to power its operations with carbon-free energy 24/7 by 2030 to reduce its environmental impact.
YouTube Updates Privacy Policy to Allow Removal of AI-Generated Content Mimicking Users
• YouTube updated its privacy policies to allow requests for the removal of AI-generated content mimicking a person's appearance or voice
• The platform evaluates several criteria before removing content, including realism, disclosure, and potential public interest value
• YouTube's policy mandates "first-party claims" for privacy violations, with exceptions for legal guardians or representatives.
Figma Suspends New AI Design Tool After Mimicking Apple’s Weather App Interface
• Figma launched new AI-powered tools last week, aimed at overcoming creative barriers and boosting efficiency
• The 'Make Design' feature was disabled after it inadvertently copied the Apple Weather app interface
• The issue was identified when a user requested a mock-up design for a weather app, which repeatedly mirrored Apple’s design.
🎓AI Academia
New Defense Technique Uses Self-Evaluation to Protect LLMs from Adversarial Attacks
• A new defense mechanism against adversarial attacks on LLMs leverages self-evaluation to classify interactions as safe or unsafe, requiring no model fine-tuning
• This method has proven more effective than existing solutions like Llama-Guard2 and content moderation APIs, drastically reducing attack success rates on both open and closed-source LLMs
• Despite potential targeting, this strategy remains robust even under adaptive attack conditions, maintaining overall lower attack success rates compared to undefended models.
New InternLM-XComposer-2.5 Model Excels in Long-Contextual Vision-Language Tasks
• InternLM-XComposer-2.5 introduces three new upgrades, enhancing vision-language comprehension and text-image composition capabilities
• Tested on 28 benchmarks, IXC-2.5 outperforms existing open-source models in 16 of them and matches or surpasses top models like GPT-4V in others
• Available for public use, InternLM-XComposer-2.5 supports long-contextual inputs and outputs, significantly expanding application possibilities in various domains.
SOS: New Training Time Attack Exposes Vulnerabilities in Open-Source Large Language Models
• SOS, a new training time attack technique, exploits soft prompt tuning to insert adversarial embeddings into open-source large language models without altering model weights
• The attack demonstrates potential through various scenarios, including backdoor attacks, jailbreak attacks, and prompt stealing, highlighting its versatility and threat level
• Researchers introduce a ‘copyright token’ strategy within the SOS framework to protect copyrighted content from being learned or utilized by compromised models. Read more
PharmaGPT Establishes New Standards in Bio-Pharmaceutical and Chemical NLP Modeling
• PharmaGPT, developed by PatSnap Co., LTD., is a domain-specific large language model tailored for the bio-pharmaceutical and chemistry sectors
• The models, featuring 13 billion and 70 billion parameters, surpassed general models on benchmarks like NAPLEX, demonstrating advanced NLP capabilities
• PharmaGPT's training involved a diverse corpus of billions of tokens, preparing it to handle intricate terminologies and specialized knowledge effectively.
Guided Deferral Systems in Healthcare Enhance Collaboration Using Large Language Models
• A novel human-AI collaboration system utilizes large language models to enhance decision-making in healthcare by providing intelligent guidance when AI defers cases
• The system significantly improves classification and deferral performance, demonstrating the success of blending verbalised and hidden-state predictions from the models
• A pilot study highlights the deferral system's efficacy in practical healthcare applications, maintaining both high performance and strict data privacy.
Study Exposes Flaws in Ranking Large Language Models Through Leaderboards
• Sensitivity of leaderboard rankings for Large Language Models (LLMs) can shift significantly with minor benchmark alterations, changing rankings by up to eight positions
• Systematic experiments reveal that even small changes, like order of answer choices, can destabilize LLM evaluation leaderboards
• Recommendations from the study stress the need for hybrid scoring methods in answer selection to ensure more stable evaluative benchmarks.
About us: 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!
