This week's top AI news and research stories featured OpenAI's high-level guidelines for use by human labelers to steer model behavior, all about AlphaFold 3, Saudi Arabia's $100 billion investment AI, and a method that commands a robot to perform practical tasks via signals from an electroencephalogram (EEG). But first:
IBM releases Granite code models to make coding easier for developers (IBM)
IBM open-sourced a family of Granite code models, ranging from 3 to 34 billion parameters, to the developer community. The models, trained on code from 116 programming languages, automate many programming tasks like code generation, bug fixing, and documentation to boost developer productivity. In testing, Granite models matched or outperformed other open-source and closed-source code models (including much larger models) on a range of benchmarks, demonstrating strong performance in code fixing, synthesis, explanation, editing, and translation across programming languages.
Stack Overflow partners with OpenAI, but many contributors revolt (The Verge)
The two companies’ API partnership aims to provide accurate and vetted data to OpenAI users and integrate Stack Overflow’s technical knowledge into ChatGPT. However, the deal has faced significant backlash from Stack Overflow users, who are attempting to remove or edit their posts, leading to bans and content reversals by moderators, who cite the site’s Terms of Service. The controversy highlights the ongoing debate about who owns user-generated content on online platforms and whether users have the right to control how their data is used, especially in the context of AI training.
ElevenLabs teases song generator with realistic vocals (Tom’s Guide)
ElevenLabs, known for its AI voice cloning, announced a new music generator that creates full-length songs (between two and three minutes) from a single text prompt. The AI-generated tracks, spanning various popular genres, feature natural-sounding vocals that compare favorably to competitors’. ElevenLabs’ product is still not publicly available, but it follows releases by Udio and Suno that have shown rapid advances in AI models’ ability to generate longer and more structured songs.
Microsoft deploys air-gapped AI for spies (Ars Technica/Bloomberg)
Microsoft created a version of OpenAI’s GPT-4 language model that operates in a secure, isolated environment for use by U.S. intelligence agencies. The system, which has been in development for 18 months, allows agencies to analyze classified data without the risks associated with internet connectivity, like data breaches or hacking attempts. The need for such a specialized system shows the growing collaboration between major tech companies and government entities in the realm of AI and national security, and the potential for similar applications for business intelligence and other sensitive information.
Apple to announce generative AI-powered Siri at WWDC in June (The New York Times)
According to this report, after Apple executives realized Siri had fallen behind chatbots like OpenAI’s ChatGPT, the company made generative AI a top priority, reassigning hundreds of engineers to the project and exploring the creation of its own AI-optimized servers. The new Siri will be more conversational and versatile, processing requests on-device for greater privacy. Meanwhile, other reports claim that Apple is finalizing a deal to bring ChatGPT and other chatbots to its devices in its next mobile OS.
Elon Musk's xAI set to close funding round at $18 billion valuation (Forbes/Bloomberg)
The round is expected to raise approximately $6 billion from investors including Sequoia Capital. It was reportedly marketed using a pitch deck highlighting Musk’s successes at Tesla and SpaceX and the potential for xAI to leverage high-quality data from Musk’s social network X (formerly Twitter) to train its AI models and Grok chatbot With the infusion of cash, xAI can hire more talent and invest in infrastructure, including a reported $10 billion deal to rent servers from Oracle.
Voice cloning technology brings country singer Randy Travis back to the recording studio (The Verge)
Travis’s new song, “Where That Came From,” is his first since a 2013 stroke made it impossible for him to properly sing. By training an AI model on 42 of Travis’s vocal-isolated recordings and having fellow country singer James DuPre provide the base vocals, producers were able to create a convincing rendition of Travis’s signature baritone style. This process demonstrates the potential for AI to assist artists in creating music that might otherwise be impossible, but also raises concerns about the future of music production and the ethical use of AI-generated vocals.
OpenAI develops AI detection tools to identify DALL·E 3 generated images (OpenAI)
The company introduced a detection classifier that predicts the likelihood an image was generated by DALL·E 3. Early testing shows the classifier correctly identifies roughly 98% of DALL·E 3 images while incorrectly tagging less than 0.5% of non-AI generated images. OpenAI is opening applications for outside researchers to access and test the classifier to assess its effectiveness and real-world application. The tool is part of tech companies’ broader efforts to help identify the provenance of AI-generated media, particularly to stem its potential use in fraud and propaganda.
AI-powered audiobooks gain traction despite concerns (Bloomberg)
Amazon’s “virtual voice” tool, which allows self-published authors to easily convert their ebooks into audiobooks, has been used to create more than 40,000 audiobooks on Audible since its beta launch last year. While AI narration enables more readers to enjoy audiobooks and authors to save hundreds or thousands of dollars on production costs, some listeners have expressed concerns about the influx of AI-narrated content and the current inability to filter these titles out when browsing. As voice technology continues to develop and major publishers strike their own deals with AI companies, tensions continue between publishers, authors, voice talent, and readers regarding the future of audiobook narration.
Vidu, a Chinese counterpart to Sora (SCMP)
Chinese start-up Shengshu Technology, in collaboration with Tsinghua University, unveiled Vidu, a text-to-video AI tool similar to OpenAI’s Sora. While Vidu can generate 16-second 1080p videos based on text prompts, it still lags behind Sora’s 60-second video generation capability. Vidu’s launch highlights Chinese companies’ ongoing efforts to catch up with leading global AI players, despite facing challenges such as limited access to advanced computing hardware due to U.S. export restrictions.
AI system to protect Olympic athletes from online abuse (IOC)
The International Olympic Committee (IOC) will implement a new AI-powered monitoring service to safeguard athletes and officials from online abuse during the Paris 2024 Olympic and Paralympic Games. The system, which was successfully piloted during Olympic Esports Week, will monitor thousands of social media accounts in real-time across multiple platforms and languages, flagging abusive messages for swift action by the relevant platforms. This is the first time AI will be used to protect such a large number of athletes across various sports at a major international event and could set a precedent for future sporting events and other industries.
U.S. Homeland Security tests AI in training officers to interview refugee applicants (Reuters)
The Department of Homeland Security’s pilot program will use AI to simulate refugee characteristics, such as reticence in discussing trauma, to help officers practice interviewing applicants, but will not make the immigration decisions themselves. It is unclear whether outside companies will help build the model or how it will be trained, other than that it will include country-specific conditions. By using AI to simulate refugee characteristics during training, junior officers can better prepare for real-world interviews and make more informed decisions, while freeing senior officers to conduct more interviews themselves.
TikTok to label AI-generated content with digital watermarks (Reuters)
The network announced plans to label images and videos uploaded to its platform that were created using AI, even if they were generated outside the app. The company will use a digital watermarking system called Content Credentials, developed by a coalition co-founded by Adobe and Microsoft and also adopted by OpenAI and YouTube, to identify and label AI-generated media, while unlabeled AI-generated media may be removed from the platform. TikTok aims to prevent the spread of misleading images and videos, particularly in light of concerns about potential interference in the upcoming U.S. elections.
Apple developing in-house AI chip for data centers (Reuters/Wall Street Journal)
Apple reportedly has been working for several years on its own AI chip, codenamed Project ACDC, to power artificial intelligence software in its data centers. The move aims to leverage the company’s chip design expertise for its server infrastructure, focusing on running AI inference models rather than training them. Developing an in-house AI chip for inference could allow Apple to reduce its dependence on other companies like Nvidia, which currently dominates the market for AI training chips.
X launches Stories, AI summaries of trending news for Premium subscribers (TechCrunch)
X, formerly Twitter, will now use Grok to generate summaries of trending stories on the platform’s “For You” page in the Explore section. The summaries are based on conversations around the news on X, rather than the actual news articles, which has raised concerns about potential misinformation. The feature helps showcase Grok and is part of X’s push to drive Premium subscriptions.