300 million users already? Baidu stepping up big time with Ernie 4.0 Turbo
++ Google Launches Groundbreaking Gemini 1.5 Pro, Innovative Living Skin for Robots, Space-Based Data Centers, AI Integration in Audi Vehicles & More Tech Advances..
Today's highlights:
🚀 AI Breakthroughs
Google Enhances Gemini 1.5 Pro with 2 Million Token Context, Code Execution, and Introduces Gemma 2
• Gemini 1.5 Pro now offers developers a 2 million token context window, enhancing long-context data handling capabilities
• Code execution feature enabled on Gemini 1.5 Pro and Flash, allows dynamic Python code generation and execution within a secure sandbox
• Gemma 2 model becomes available in Google AI Studio, providing developers with advanced AI experimentation tools
• Significant speed and affordability improvements are noted with the deployment of Gemini 1.5 Flash in various technological applications
• Continuous innovations and updates in Gemini 1.5 Flash cater to a broad array of real-world applications, from AI-assisted visual aids to legislative analysis.
Baidu Unveils Ernie 4.0 Turbo Amid Tightened AI Market Competition in China
• Baidu launched Ernie 4.0 Turbo, enhancing functionalities over its predecessor Ernie 4, introduced just months earlier in October 2023
• The upgraded AI model on the Qianfan AI platform will be accessible through both web and mobile applications, according to Baidu;
• Baidu’s PaddlePaddle AI ecosystem now supports over 14.65 million developers and serves around 370,000 businesses and institutions;
• Following OpenAI's restriction on API access in China, Baidu, along with other local AI players, is offering free migration services to attract those affected;
• Ernie’s user base has grown rapidly, now boasting a reach of 300 million users since its initial launch.
Japanese Researchers Develop Durable Living Skin for Robots, Enhancing Human-like Features
• Researchers in Japan develop a new method to bind living skin tissue to robotic surfaces, enhancing durability and human-like appearance
• The innovative anchoring technique, inspired by human skin ligaments, uses collagen gel for a seamless attachment to the robot's structure
• Humanoid robots such as Ameca employ advanced AI to interact naturally with humans, leveraging eye movements for expressive communication
• Unlike industrial robots, humanoids are designed for roles in hospitality and healthcare, where realistic facial expressions are critical for engagement
• The University of Tokyo's latest advancements may lead to robots with enhanced sensory capabilities, mimicking human skin's sensitivity to environmental stimuli
• Future research aims to integrate a vascular system into the robot's skin, enhancing longevity and maintaining the skin's health through nutrient supply.
Thales Alenia Space Leads Feasibility Study for Space-Based Data Centers
• Thales Alenia Space announced results from the ASCEND feasibility study, aimed at deploying space-based data centers to support the EU Green Deal’s net-zero goals
• The ASCEND project, funded by the European Commission, explores the environmental and technological feasibility of data centers in orbit, promising reduced CO2 emissions
• Findings suggest space data centers could eliminate the need for water cooling, offering a drought-resistant solution to data processing's environmental impact
• ASCEND could enhance Europe's digital sovereignty, aiming to establish substantial data center capacity in space by 2050, ensuring secure data handling
• The study indicates economic viability with potential multi-billion euro ROI by 2050
• Partnerships within the ASCEND project include key European space and technology companies, aiming to utilize modular space infrastructures and advanced robotics for in-orbit assembly.
Audi Integrates ChatGPT into Vehicles with Enhanced Voice Control Features
• Audi has started using the AI-based chatbot ChatGPT in their vehicles through Microsoft Azure OpenAI Service to enhance voice control capabilities
• Starting July, approximately two million Audi models from 2021, equipped with MIB 3, will incorporate natural language processing for in-car interactions
• The new Audi Q6 e-tron and future models with E3 1.2 electronic architecture will integrate ChatGPT via Cerence Chat Pro to augment the Audi assistant
• ChatGPT allows Audi drivers to query general knowledge questions, control infotainment, navigation, and AC systems using voice commands
• Audi emphasizes data security, ensuring that all queries are deleted post-processing and that ChatGPT does not access vehicle data
• Audi is exploring broader AI applications to enhance customer experiences and optimize business processes, ensuring compliance with EU regulations.
Luma Keyframes Enables Morphing of Characters and Environments in Dream Machine
• Luma Labs unveils Dream Machine Keyframes, enabling users to morph characters and settings effortlessly
• Through Dream Machine, individuals can set start and end frames, letting AI interpolate the rest for smooth transitions
• The process involves uploading two images, one as a starter and one as an end frame, with optional transformative instructions
• Dream Machine provides an option to extend the video beyond the last keyframe, facilitating longer animations
• Users can experiment with various images and detailed text inputs to guide AI in crafting unique sequences
• This AI-powered feature supports continuous extension of animations by adding new frames and instructions after initial generation.
Goldman Sachs to Implement Generative AI Tool for Developers Company-Wide
• Goldman Sachs completes the rollout of its AI-based code generation tool across the company by month's end
• Goldman's centralized AI approach involves exclusive use of its internal GS AI Platform, banning external tools like ChatGPT
• GS AI Platform allows secure data manipulation and model customization while complying with strict regulations
• Developers can quickly create new applications using the GS AI Platform, significantly reducing development time from months to weeks
• The platform supports various models, enabling tailored applications across different financial services
• Goldman's AI coding assistant reported a 20% boost in developer efficiency, marking its most extended AI deployment.
⚖️ AI Ethics
Microsoft Updates AI Systems to Mitigate New Skeleton Key Jailbreak Technique
• Skeleton Key jailbreak attack allows users to bypass AI guardrails, enabling harmful or forbidden behaviors in models under attack
• Microsoft has updated Azure AI-managed models with Prompt Shields to detect and mitigate Skeleton Key and similar attacks
• Microsoft Defender for Cloud now integrates with Azure OpenAI Service for enhanced security alerts and protection against AI-targeted threats
• Microsoft encourages AI developers to employ red teaming techniques using updated tools like PyRIT to defend against Skeleton Key attacks
• Azure AI Content Safety features like Prompt Shields set to high severity help safeguard against unsafe content generation in AI systems
• Microsoft shares new security insights and tools through platforms like Microsoft Threat Intelligence Blog and the Microsoft Threat Intelligence podcast.
YouTube Seeks Permission to Clone More Artists' Styles Amidst Legal Controversies with AI Music
• YouTube seeks permissions from major labels like UMG and Sony to legally use their music for AI tool training
• Financial Times reports Google's platform offering lump sums to record labels for song licensing, not royalty-based payments
• AI music tools under development by YouTube aim to incorporate styles of "dozens" of artists in new features to be launched this year
• Record labels and artists express concerns over AI technology infringing on rights, with over 200 artists advocating for a halt in AI-generated music
• Recent RIAA lawsuits target generative AI music companies Suno and Udio for mass, unlicensed use of sound recordings, seeking hefty damages per infringement
• Sony Music and UMG pose significant resistance, highlighting past disputes over AI-generated content on platforms like TikTok.
Study Reveals High Undetected Rate of AI-Completed Exams in UK University
• A blind study conducted at the University of Reading revealed 94% of AI-written exam submissions went undetected in the assessment process
• AI-generated content, submitted blindly in five undergraduate psychology modules, consistently outperformed human students, scoring higher by half a grade boundary on average
• The study highlighted a significant challenge in educational integrity, showing that AI can successfully infiltrate university grading systems without detection
• Results indicated that AI submissions outperformed real student answers in 83.4% of cases, raising concerns about the fairness and reliability of current assessment methods
• This real-world 'Turing Test' underscores the urgent need for academia to adapt assessment strategies to effectively identify AI-generated content.
AI-Enhanced "CriticGPT" Outperforms Humans in Detecting Code Errors and Writing Critiques
• CriticGPT, a GPT-4-based model, has shown to outperform humans in detecting bugs within model-generated code critiques
• In evaluations, text critiques from CriticGPT were preferred over those from human contractors in 63% of instances
• Human and CriticGPT collaborations produce more detailed critiques with fewer errors, compared to either working alone
• The study discovered that CriticGPT can identify numerous errors in data previously rated as flawless, challenging earlier assessments
• Despite their effectiveness, CriticGPT models can generate false positives, creating a risk of misleading human evaluators
• A new sampling strategy, Force Sampling Beam Search, enhances critique quality by balancing detection of real and spurious issues.
🎓AI Academia
Improving LLM Performance on Long-Context Tasks Through Synthetic Data Finetuning
• Recent studies indicate Large Language Models struggle with long-context information retrieval and reasoning a synthetic dataset fine-tuning method may improve this. ;
• Experiments on models like GPT-3.5 Turbo and Mistral 7B show the new methodology significantly boosts performance in long-context tasks. ;
• The synthetic, numerically-based dataset is devoid of factual content, reducing the risk of data hallucination commonly seen in other datasets. ;
• Results display a marked 10.5% improvement in 20 document MDQA evaluations at position 10 for GPT-3.5 Turbo, highlighting the effectiveness of the approach. ;
• Despite enhancements in long-context capabilities, general benchmark performances on models remain largely stable, preserving overall utility. ;
• Compared to other baseline long-context datasets, the proposed solution does not compromise performance and avoids issues like performance drops up to 6.19%. ;
New Guide Advances Standardized Evaluation of Generative AI's Diverse Social Impacts
• A new guide details standard approaches to evaluating generative AI systems across text, image, video, audio, and multimodal setups
• Seven social impact categories for AI evaluation include bias, cultural sensitivity, privacy, financial and environmental costs, and labor issues
• The evaluation framework differentiates impact assessments inherent to AI systems from broader social context evaluations
• Existing AI system evaluations may overlook crucial nuances and localized impacts, necessitating more comprehensive methods
• Recommendations target mitigating harms in various subcategories like trustworthiness, inequality, authority concentration, labor, and environmental impacts
• The study calls for substantial investment in improving social impact evaluations and developing globally applicable standards.
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!
