Early Testing of Apple’s AI

Apple has introduced its AI system, Apple Intelligence, through the developer beta of iOS 18.1, iPadOS 18.1, and macOS Sequoia 15.1, despite the official release of iOS 18 being pending. This beta includes early versions of features revealed at WWDC, akin to writing aids for text rephrasing, proofreading, and summarization, alongside an updated Siri interface that supports each voice and text input.

Additional features include automatic call recording, email prioritization, and natural language photo search. The system utilizes AFM-on-device, a model with about 3 billion parameters for on-device processing, and AFM-server, which leverages Apple’s private cloud for more intensive tasks.

Early reports indicate that while the system’s performance is slower than some competitors, it’s efficient, managing tasks like rephrasing emails with minimal power usage. The total rollout of Apple Intelligence is anticipated in early 2025, with current access limited by a waiting list as a consequence of high demand and hardware requirements. Read more.

OpenAI has granted the US AI Safety Institute at NIST early access to its latest AI model. This collaboration, as stated by OpenAI CEO Sam Altman, goals to boost AI evaluation methods. Altman reaffirmed the corporate’s commitment to allocate a minimum of 20% of its computing resources to safety initiatives, addressing concerns raised by former safety researcher Jan Leike about unmet guarantees.

In May, OpenAI also revised its policies to remove clauses that allowed the revocation of stock options and imposed silence on employees. These changes intend to foster a culture where employees can freely express concerns. The initiative underscores OpenAI’s dedication to advancing AI safety and transparency while promoting a supportive work environment. Read more.

Argentina’s Ministry of Security has launched an AI unit, harking back to the “Minority Report” concept, to anticipate and forestall future crimes. Named the Artificial Intelligence Unit Applied to Security, this initiative will utilize machine learning to research past crime data and forecast potential criminal activities. Its duties encompass crime prevention, drone surveillance, social media monitoring, and facial recognition technology.

Security Minister Patricia Bullrich asserts that the AI system will improve the responsiveness and precision of federal police and security forces. Nevertheless, critics raise concerns about privacy and freedom of expression, pointing to prior issues with government surveillance and facial recognition. Read more.

Neura Robotics has showcased its humanoid robot 4NE-1, demonstrating capabilities like ironing and box-moving in a brand new video. This comes as Nvidia introduces its latest humanoid development tools on the Siggraph conference. Neura, amongst select firms with early access to Nvidia’s tools, has generated buzz with the 4NE-1, which had previously been in development behind closed doors.

The video highlights 4NE-1’s advanced functionalities, though such promotional material often includes selective editing. Neura plans to boost its robot’s performance through collaboration with Nvidia, integrating advanced computing and simulation technologies. Despite the thrill, experts urge caution, noting that real-world testing and widespread deployment of humanoids are still within the early stages. Read more.

Runway has introduced the Gen-3 Alpha Turbo, an upgraded version of its AI video model, boasting seven times the speed of its predecessor, Gen-3 Alpha, while maintaining similar performance standards. The brand new model can produce a 10-second video in roughly 11 seconds, enhancing real-time interactivity.

Runway plans to supply Gen-3 Alpha Turbo at a reduced price, potentially lowering the price per second of generated video in comparison with previous models. This move is designed to spice up usage and subscription revenues, despite the model being computationally lighter. The corporate continues to refine its technology amidst ongoing scrutiny over its data training practices, including allegations of using copyrighted content from YouTube.

While Runway has not confirmed specifics about its training data, the broader AI industry faces similar concerns regarding data usage and mental property. Read more.

Microsoft says OpenAI is now a competitor in AI and search: This development follows OpenAI’s recent announcement of SearchGPT, a brand new search engine that directly challenges Microsoft’s Bing. Although OpenAI stays a major partner, the inclusion on Microsoft’s competitor list indicates an evolving competitive landscape. This modification comes amid a yr of upheaval for OpenAI, including leadership changes and Microsoft’s establishment of a brand new AI unit led by Mustafa Suleyman. Read more.

Reddit CEO says Microsoft must pay to look the location: Reddit CEO Steve Huffman has called out Microsoft, Anthropic, and Perplexity for using Reddit’s data without permission. Huffman criticized these firms for scraping Reddit’s content and noted the challenges of blocking them. Reddit’s recent update to its robots.txt file blocks web crawlers that lack agreements, which has led to Reddit search results appearing only on Google.

Amazon introduces latest AI model to make checkout-free ‘Just Walk Out’ stores even higher: The brand new model integrates advanced machine learning capabilities with cameras, shelf sensors, and RFID technology to streamline in-store transactions. The updated system consolidates data processing, replacing the previous method which required multiple models to handle different tasks sequentially. It’s designed to handle various retail environments, from airports to hospitals, and is meant to cut back theft by accurately tracking purchases. Despite using extensive visual data, Amazon emphasizes that the system doesn’t collect biometric information, focusing solely on transaction-related data. Read more.

Midjourney releases V6.1 model, which is quicker and “generally more beautiful”: This update enhances overall image quality, reduces artifacts, and hastens standard image generation by 25%. It refines details in human extremities, plants, and animals while introducing latest upscaling tools. The addition of the –q 2 parameter offers increased texture but on the expense of image coherence. Despite these upgrades, V6.1 doesn’t support inpainting or outpainting, features still reliant on V6. The brand new model is now the default option, with a V6.2 update scheduled for August to deal with user feedback.