Today:

• Apple's Generative Intelligence

• Musk Threatens Apple Over AI

• OpenAI Announces Key Appointments

• In-House AI Gives Google Edge

• Adobe's New Terms Ensure Privacy

• Hugging Face, Pollen Unveil Robot

Introducing Apple Intelligence, the personal intelligence system that puts powerful generative models at the core of iPhone, iPad, and Mac

Apple introduced Apple Intelligence, a new personal intelligence system for iPhone, iPad, and Mac, integrated into iOS 18, iPadOS 18, and macOS Sequoia. Utilizing generative models and Apple silicon, it enhances language and image processing, aiding tasks like writing, email management, and notifications. New features include Image Playground for creative image creation, and Genmoji for personalized emojis. Siri gains advanced language understanding and contextual actions. 

Apple ensures privacy with on-device processing and Private Cloud Compute for complex tasks. ChatGPT integration enhances capabilities without compromising user data. Available in beta this fall, it promises a transformative user experience.

APPLE

Elon Musk threatens Apple ban over OpenAI integration, cybersecurity experts raise alarms

Elon Musk, CEO of Tesla and SpaceX, has threatened to ban Apple devices from his companies if Apple integrates OpenAI's technology into its operating systems, citing security concerns. This announcement follows Apple’s partnership with OpenAI, revealed at the Worldwide Developers Conference. Musk's criticism reflects the growing rivalry in the AI industry. Security experts also express concerns over potential vulnerabilities in Apple’s integration of OpenAI. 

Despite Apple's assurances of on-device processing and strict data protection, skepticism remains about the safety of such integration. Musk's history with OpenAI and his new AI venture, xAI, adds to the tension.

VENTUREBEAT

OpenAI welcomes Sarah Friar (CFO) and Kevin Weil (CPO)

OpenAI has appointed Sarah Friar as Chief Financial Officer (CFO) and Kevin Weil as Chief Product Officer (CPO). Sarah, formerly CEO of Nextdoor and CFO at Square, will lead the finance team to support OpenAI's core research and scale operations. Kevin, previously President of Product and Business at Planet Labs and a senior product leader at Facebook, Instagram, and Twitter, will guide the product team to apply research to consumer, developer, and business products. 

These appointments aim to enhance OpenAI’s mission to build and deploy AI products safely and effectively.

OPENAI

Google’s in-House AI Development May Be Competitive Edge

Microsoft is outsourcing its top AI research to OpenAI, raising concerns that it could weaken its own AI innovation, according to Okta CEO Todd McKinnon. This move may benefit Google, which relies on its in-house AI research. 

While Microsoft has invested heavily in OpenAI, McKinnon warns that this could limit Microsoft’s control over AI development and turn it into a consultancy. Google, despite some setbacks, remains a strong competitor with its DeepMind research. The AI industry is under scrutiny from the Justice Department and FTC for potential competition issues, as major companies dominate the field.

PYMNTS

Adobe overhauls terms of service to say it won’t train AI on customers’ work

Adobe is updating its terms of service to ensure users that their work will not be used to train AI models. This change comes after backlash from customers who misinterpreted Adobe’s previous terms, fearing their data could be exploited for AI training. Adobe clarified that they have never used customer content for AI training and only access it when legally required. 

The new terms, effective June 18, aim to enhance transparency and trust. Despite previous issues with perceived monopolistic practices and subscription models, Adobe emphasizes its commitment to protecting user content and maintaining ethical standards in AI development.

THE VERGE

Hugging Face and Pollen Robotics show off first project: an open source robot that does chores

Hugging Face and Pollen Robotics have unveiled their first collaborative project, an open-source humanoid robot named Reachy 2, designed to perform household chores and safely interact with humans and pets. Initially controlled via virtual reality by a human, Reachy 2 learned tasks through machine learning, processing 50 videos of teleoperated sessions. 

Hugging Face has released the dataset and model used for training, allowing anyone to replicate the process on smaller robots. This collaboration emphasizes accessibility and open-source development in robotics, aiming to democratize advanced robotics technology and promote ethical practices.

VENTUREBEAT

🧠RESEARCH

Mixture-of-Agents Enhances Large Language Model Capabilities

Researchers developed a Mixture-of-Agents (MoA) method to enhance large language models (LLMs). The MoA architecture layers multiple LLM agents, each using the outputs of previous agents. This approach outperformed GPT-4 Omni, achieving top scores in benchmarks like AlpacaEval 2.0, showcasing the potential of collaborative LLMs.

Large Language Model Confidence Estimation via Black-Box Access

Researchers propose a new framework to estimate the confidence of large language model (LLM) responses using black-box access. By engineering features and training a logistic regression model, this method outperforms existing approaches on benchmark datasets, offering over 10% improvement in some cases. It generalizes well across different LLMs without retraining.

GenAI Arena: An Open Evaluation Platform for Generative Models

GenAI-Arena is an open evaluation platform for generative models, leveraging user feedback to assess model performance in text-to-image, text-to-video, and image editing tasks. Covering 27 models, it has gathered over 6000 community votes. The platform aims to provide reliable evaluation metrics and has released GenAI-Bench for further research, highlighting gaps in current model-based evaluations.

WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild

WildBench is an evaluation framework for large language models (LLMs) using 1,024 real-world tasks from user queries. It introduces two metrics, WB-Reward and WB-Score, for systematic, interpretable assessments. Results correlate highly with human ratings, achieving Pearson correlations of 0.98 and 0.95 respectively, outperforming previous benchmarks in evaluating LLM performance.

CRAG -- Comprehensive RAG Benchmark

CRAG (Comprehensive RAG Benchmark) evaluates Retrieval-Augmented Generation (RAG) for factual question answering with 4,409 Q&A pairs and mock APIs. Covering five domains and eight question categories, CRAG reveals low accuracy in dynamic, unpopular, or complex questions. Current LLMs achieve ≤34% accuracy, while RAG boosts it to 44%, highlighting areas for future research.

🛠️TOP TOOLS

Mental AI - create and operate agents that feature loops, memory, and various tools, all through straightforward markdown syntax

Site Profile - The Simplest AI-Powered API to Access the Most Comprehensive Website Information.

Asana AI Teammates - set company-wide goals, manage strategic plans, and get work done on a single platform.

Replicate - Run and fine-tune open-source models. 

EverLearns - Create, organize, and visualize your ideas effortlessly with our mindmap generator tool.