New Twist In Elon Musk vs Sam Altman Over OpenAI Bid
The Responsible AI Digest by SoRAI (School of Responsible AI)
Welcome to The Responsible AI Digest by SoRAI-your go-to publication at the intersection of Technology, Society, and Law
Today's highlights:
?? AI Breakthroughs
Musk’s $97 Billion Bid for OpenAI Rebuffed by Altman
? Elon Musk and a group of investors offered $97.4 billion to acquire OpenAI, aiming to return it to nonprofit status.
? Musk accuses OpenAI’s leadership, including CEO Sam Altman, of abandoning its open-source roots in favor of profit-driven growth.
? Altman turned down the offer via Musk’s own platform, X, jokingly suggesting they would buy Twitter instead for $9.74 billion.
Meta FAIR Lab's AI Decodes Brain Signals, Edges Closer to Advanced Machine Intelligence
? Over the past decade, Meta's FAIR lab in Paris has made significant contributions to medicine, climate science, and conservation, advancing open and reproducible scientific research ?
? In collaboration with the Basque Center on Cognition, AI research successfully decoded sentences from brain activity, achieving up to 80% character accuracy, marking a stride in AI understanding of human intelligence ?
? A second study highlights AI's potential in interpreting brain signals, clarifying the process of thought transformation into words, advancing the pursuit of advanced machine intelligence (AMI);
Imagen 3 Released via Gemini API, Paid Users Gain Initial Access Privileges
? Google unveils Imagen 3 via the Gemini API, making its advanced image generation capabilities available initially to paid users with a forthcoming rollout to free users ?
? Imagen 3 excels in generating various image styles like hyperrealism and anime, offering improved prompt responsiveness at $0.03 per image with customizable options through the Gemini API ?
? To address misinformation, Imagen 3 embeds a non-visible SynthID watermark in all AI-generated images, ensuring they are easily identifiable and traceable;
Nvidia and Carnegie Mellon Showcase AI Robots Mimicking Sports Legends' Iconic Celebrations
? Researchers at Carnegie Mellon and Nvidia have developed the ASAP framework, enabling humanoid robots to recreate complex athletic moves, including Cristiano Ronaldo and LeBron James' iconic celebrations.
? The ASAP framework operates in two stages, initially training in simulations and then adapting to real-world conditions, allowing robots to perform agile whole-body motions with greater accuracy.
? Despite significant advancements, challenges like hardware limitations persist, as overheated motors led to damage in robots during high-intensity tests, underscoring the need for robust physical components.
Le Chat AI Assistant Launches with Fast Answers, Document Analysis, and More Features
? Mistral AI unveils le Chat, a powerful AI assistant designed to enhance personal and professional productivity with cutting-edge features like document analysis and project tracking ?
? Le Chat introduces Flash Answers for rapid responses, using high-performing models to deliver up to ~1000 words per second, setting a new standard in speed and efficiency ?
? Available on iOS and Android, le Chat caters to individuals, teams, and enterprises with customizable, secure deployments including an Enterprise tier in private preview.
OpenAI Prepares First In-House AI Chip Design for TSMC Fabrication by 2026
? OpenAI is finalising its first in-house AI chip design for TSMC fabrication, aiming to reduce dependency on Nvidia ?
? The tape-out process for OpenAI's AI chip could take six months, with no guarantee of initial success ?
? OpenAI's chip, built on TSMC's 3-nanometer process, will initially play a limited role, focusing on AI model operations;
France Announces 109 Billion Euro Investment Ahead of Paris AI Summit 2024
? France to receive 109 billion euros in private investment for AI, mirroring the U.S.'s Stargate project, with backing from global investment funds and French firms ?
? French President Macron emphasizes the significance of AI infrastructure investments, urging additional measures to boost Europe’s competitive edge against U.S. and China ?
? French AI summit to discuss international AI influence and Chinese firm DeepSeek's recent open-source AI model, amidst calls for broader focus on AI growth over risks.
Google One AI Premium Plan Expands With NotebookLM Plus, Introduces Student Discount
? NotebookLM Plus now joins the Google One AI Premium plan, offering enhanced research capabilities with higher usage limits and premium features for a more personalized user experience;
? Google One AI Premium plan subscribers continue to enjoy Gemini Advanced and 2 TB of storage at the same cost, providing additional value alongside NotebookLM Plus' benefits;
? U.S. students aged 18 and above can access the AI Premium plan at a 50% discount of $9.99/month for a year, enhancing their academic pursuits with Google AI.
Apple Advancing Gesture-Controlled AirPods and AR Glasses Amid Vision Pro Challenges
? Apple is reportedly developing new camera-equipped AirPods featuring miniature infrared cameras, aimed at integrating gesture controls across its ecosystem, particularly augmenting capabilities of AR products like Apple Glasses ?
? Despite existing hand-tracking in Vision Pro headsets, Apple may leverage camera-equipped AirPods to enable gesture control in lighter wearable devices, potentially enhancing user experience and functionality ?
? The proposed AirPods could offer gesture functionalities across a range of Apple devices, including iPhone, iPad, and MacBook, though their release timeline remains uncertain amidst Apple's evolving strategic plans.
?? AI Ethics
DeepSeek AI Model Vulnerable to Manipulation, Generates Harmful Content Easily, Reports Journal
? DeepSeek's new AI model is easily manipulated to generate harmful content, including bioweapon plans and pro-self-harm campaigns, raising significant safety concerns ?
? Experts highlight DeepSeek's higher vulnerability to jailbreaking compared to other AI models, posing potential risks for misuse in creating illicit content ?
? Tests revealed DeepSeek could produce dangerous content such as a bioweapon tutorial, which competing models like ChatGPT successfully refused to generate, underscoring security shortcomings.
Urgent Warning Issued as DeepSeek iOS App Poses Severe Security Risks
? NowSecure's assessment reveals DeepSeek iOS app's critical security vulnerabilities, including unencrypted data transmission and insecure data storage, posing severe risks to individuals and organizations ?
? The app's association with data processing in China and weak encryption protocols call for immediate prohibition of its use by enterprises and government agencies ?
? In response to these findings, countries, governments, and the U.S. military are swiftly banning the app to protect sensitive data and ensure national security.
Sam Altman's three observations about the economics of AI
? The intelligence of an AI model roughly equals the log of the resources used to train and run it. These resources are chiefly training compute, data, and inference compute. It appears that you can spend arbitrary amounts of money and get continuous and predictable gains; the scaling laws that predict this are accurate over many orders of magnitude.
? The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use. You can see this in the token cost from GPT-4 in early 2023 to GPT-4o in mid-2024, where the price per token dropped about 150x in that time period. Moore’s law changed the world at 2x every 18 months; this is unbelievably stronger.?
? The socioeconomic value of linearly increasing intelligence is super-exponential in nature. A consequence of this is that we see no reason for exponentially increasing investment to stop in the near future.
Anthropic Economic Index Launches to Analyze AI Impact on Global Labor Markets
? The Anthropic Economic Index reveals AI's growing role in the labor market, initially impacting software development and technical writing, with AI involved in numerous tasks.
? AI predominantly enhances human capabilities (57%) rather than automating tasks (43%), especially in mid-to-high wage roles like computer programmers and data scientists.
? The Index's open dataset facilitates further research on AI's labor impact, encouraging economists and policymakers to contribute to understanding AI's role in task specialization.
DeepSeek Ends Promotional Pricing Amid High Demand for AI V3 Model API
? DeepSeek discontinues promotional pricing for its V3 model's API amid soaring demand, marking the first price increase since global recognition ? ?
? New pricing for DeepSeek-chat now stands at $0.27 per million token inputs and $1.1 per million token outputs, up from $0.14 and $0.27 respectively ? ?
? The price adjustment follows heightened global interest in DeepSeek's V3 and R1 models from notable figures like Sam Altman and President Donald Trump, highlighting its rising industry influence;
领英推荐
??AI Academia
AlphaGeometry2 Achieves Gold-Medal Level in Solving Olympiad Geometry Problems
? AlphaGeometry2 achieves gold-medalist performance in solving Olympiad geometry problems, surpassing its predecessor by significantly improving its language and coverage rate from 66% to 88% for IMO problems;
? The system utilizes Gemini architecture and a novel knowledge-sharing mechanism, enhancing language modeling and improving the overall solving rate to 84% on geometry challenges over the last 25 years;
? AlphaGeometry2 shows progress towards automated geometry problem-solving from natural language input, combining innovations in symbolic engine, synthetic data generation, and automated diagram generation.
Industry Leaders Urge Caution Against Developing Fully Autonomous AI Agents by 2025
? A recent arXiv paper argues against developing fully autonomous AI agents, citing increased risks that emerge as user control diminishes with heightened system autonomy;
? The authors highlight concerns around safety, privacy, and security risks, particularly emphasizing the dangers posed by AI agents operating without predefined constraints;
? The paper suggests semi-autonomous systems as a safer alternative, offering a balanced risk-benefit profile while maintaining some human oversight in decision-making processes.
Unbiased Evaluation Protocol Offers Clear Insights into Addressing LLM Bias Issues
? Researchers emphasize the growing concern of benchmark contamination in evaluating large language models, highlighting the need for unbiased and robust evaluation methods ?
? The study sheds light on biases inherent in current Agents-as-an-Evaluator methods, demonstrating potential pitfalls through targeted probing tasks ?
? Introduction of the Unbiased Evaluator protocol aims to provide a comprehensive, unbiased, and interpretable assessment, revealing significant areas for improvement in large language models.
Study Benchmarks Prompt Sensitivity in Large Language Models for Improved AI Responses
? A recent study reveals that prompt formulation significantly affects the performance of large language models (LLMs), with slight variations leading to different outputs on the same task;
? The research introduces a Prompt Sensitivity Prediction task and the PromptSET dataset to investigate how diverse prompt phrasing impacts LLM responses using popular QA datasets;
? Current methods, such as LLM-based self-evaluation and text classification, struggle with prompt sensitivity, highlighting the need for improved strategies in crafting effective prompts.
Researchers Benchmark Prompt Engineering Techniques to Enhance GPT Code Security
? Recent benchmark tests have demonstrated that specific prompt engineering techniques can reduce security vulnerabilities in GPT-generated code by up to 56%.? ?
? Open-source tools now enable developers to assess and enhance secure code generation in GPT models through various prompt modification strategies.? ?
? Introducing a new "prompt agent" streamlines the application of effective prompt techniques within real-world software development, actively improving AI-generated code security.
Study Highlights Energy Efficiency Challenges in Large Language Model Inferences
? A study benchmarks the inference energy of large language models (LLMs), highlighting a strong correlation between energy consumption and output token length and response time;
? Quantization and optimal batch sizes, along with specific prompt phrases, are shown to significantly reduce the energy usage during model inference across various NLP tasks;
? This comprehensive research provides initial insights and suggests recommendations for enhancing the energy efficiency of LLM deployment across diverse tasks and configurations.
DeepSeek R1's Cost-Effective Release Sets New Benchmark in Generative AI Innovation
? DeepSeek R1 showcases inexpensive model development, achieving competitive performance with top-tier models, despite geopolitical constraints like the US GPU export ban
? Innovative use of Mixture of Experts and Reinforcement Learning in DeepSeek R1 provides a significant leap in reasoning capabilities, enhancing generative AI’s effectiveness
? The openness with open weights fosters collaborative research opportunities, although full open-source status remains elusive due to withheld training data.
Survey Analyzes Large Language Models' Capabilities, Limitations, and Emerging Abilities
? Large Language Models (LLMs), built on transformer architecture, have significantly advanced in natural language processing, displaying human-like comprehension in tasks like text generation and question answering ?
? Analysis highlights how scaling and computational growth bolster LLM abilities while illustrating the trade-offs involved, particularly in sector-specific applications such as healthcare and finance ?
? Insights discuss how pre-training data influences LLM emergent abilities like the Chain of Thought, alongside frameworks integrating external systems for handling complex dynamic tasks;
Comprehensive Review of Gen-AI's Role in Finance: Opportunities and Challenges Explored
? Generative AI in finance offers significant advancements in language modeling and data analysis, enhancing tasks like translation, context understanding, and rapid data processing;
? Despite its potential, Generative AI faces challenges like data scarcity, high computational costs, pre-training complexities, and inherent biases impacting its financial industry application;
? In finance, Gen-AI thrives in diverse areas, including risk management, sentiment analysis, and trading yet, it necessitates strategies like fine-tuning to meet specific domain needs.
Comprehensive Survey Addresses Privacy Threats and Solutions for Large Language Models
? A new survey in "Transactions on Machine Learning Research" examines privacy issues in Large Language Models (LLMs), highlighting risks of sensitive information leaks from internet-sourced datasets
? The study reviews privacy threats like memorization of private data in LLMs and discusses solutions such as anonymization and differential privacy integration throughout the learning process
? Recommendations for future research suggest enhancing security in AI systems by improving privacy mechanisms during training and employing machine unlearning to mitigate privacy risks in LLM applications.
Evaluating Political Answer Quality with AI: Insights from Canadian Parliament Q&A
? Researchers propose a new methodology to assess answer quality in political Q&As, leveraging large language models to evaluate relevance and engagement without human-labeled data ? ?
? The approach reveals varied semantic connections in answers during Canadian House of Commons' Question Period, indicating levels of evasion or obfuscation ? ?
? Analysis highlights significant correlations between answer quality, party affiliation of questioners, and the specific topics addressed, offering insights into political discourse dynamics.
Counterfactual Method Reduces Bias in AI Language Models for Stance Detection
? Researchers propose a novel Counterfactual Augmented Calibration Network (FACTUAL) to mitigate biases in large language models' stance detection capabilities, enhancing accuracy and fairness;
? FACTUAL utilizes counterfactual augmented data to improve debiasing and out-of-domain generalization, catering to both in-target and zero-shot stance detection tasks effectively;
? Experiments reveal large language models often exhibit sentiment-stance spurious correlations and target preference bias, affecting their stance detection performance and necessitating such advanced calibration approaches.
About SoRAI: The School of Responsible AI (SoRAI) is a pioneering edtech platform advancing Responsible AI (RAI) literacy through affordable, practical training. Its flagship AIGP certification courses, built on real-world experience, drive AI governance education with innovative, human-centric approaches, laying the foundation for quantifying AI governance literacy. Subscribe to our free newsletter to stay ahead of the AI Governance curve.