TLDR AI 2023-10-24

Apple plans to start implementing generative AI technology in iOS in late 2024 at the earliest. The company will likely build hundreds of AI servers over the next two years. It will probably offer a combination of cloud-based and on-device AI. The time frame seems to line up with the release of iOS 18.

A team of researchers from Stanford, MIT, and Princeton has developed a transparency index to evaluate the level of transparency in commercial foundation models. The index, known as the Foundation Model Transparency Index (FMTI), assesses 100 different aspects of transparency, and the results indicate that there is significant room for improvement among major foundation model companies.

Google is offering its customers two indemnities to mitigate the risks associated with generative AI. The first covers training data, protecting users against third-party intellectual property infringements. The second, generated output indemnity, covers IP claims made against content created using Google products, assuming responsible AI practices are followed.

This study assesses the ability of large language models (LLMs) to engage in human-like multi-turn conversations.

This paper introduces SALMONN, a unique AI model that combines text understanding with the ability to process various audio inputs, from speech to music. Not only does SALMONN perform well in trained audio tasks like speech recognition, it also shows impressive emergent skills like translating speech into untrained languages.

This project presents a novel detection-centric network for lidar panoptic segmentation (LPS) and tracking, moving away from the traditional segmentation-focused methods.

Weakly supervised object localization helps find objects using only simple picture labels. The new Background Activation Suppression (BAS) method does this better by focusing on important parts of a picture and ignoring the rest, showing amazing results on several datasets.

Tuning language models requires creating training data in a specific format. This is often frustrating and slow. This post dives into easy ways to automate that process.

AI Pause organizer Holly Elmore and AI Safety YouTuber Rob Miles have a long discussion on AI safety.

The AI boom has led to an oversaturation of startups with some raising massive funds on just a pitch deck, potentially indicating overhype and overvaluation. Experts caution that not all startups are innovators; many simply imitate. There should be more authenticity and clarity in AI branding and utility and less marketing of subpar AI tools as "magic".

Eureka is an open-ended agent that designs reward functions for robot dexterity at a super-human level - it can learn pen-spinning tricks, something that is difficult for CGI artists to animate frame by frame.

Framework for writing language model plugins that gives you full control.

Pi now has access to data from the web like news and events.

An AI platform that turns qualitative data into actionable insights.

Reddit may block search crawlers if it can't reach deals with generative AI companies to pay for its data. More than 535 news organizations have opted to block their content from being scraped by AI companies. X's new reading limits are due to data scraping by AI startups.