Google is creating an AI framework for Universal Music Group (UMG) that raises questions and concerns about copyright law and data scraping. The deal between the two companies was made in response to a controversial track on YouTube, “Heart on My Sleeve” by Ghostwriter977, which featured AI-generated voices of Drake and The Weeknd. Streaming services Apple and Spotify complied with UMG’s request to take the track down, but YouTube did not, exposing a conflict between big music labels and AI companies.
Google must find a way to keep UMG happy, as YouTube cannot operate without blanket licenses from the labels. Therefore, YouTube has agreed to extend Content ID, its system to ensure copyright holders are paid for their work, to AI-generated content. However, this creates problems, as no distinct legal system decides when AI content infringes copyright and removes accountability from YouTube.
The deal with UMG highlights Google’s immense bargaining power over publishers, as the company can scrape their websites for free while publishers rely on Google referrals for traffic. Google is also rolling out its Search Generative Experience to answer search queries directly using AI.
The Alliance of Motion Picture and Television Producers (AMPTP) has proposed to the Writers Guild of America (WGA) to end the ongoing strike. The proposal includes banning written works produced by artificial intelligence from being considered “literary material” under the contract and providing quarterly reports on streaming view hours per project. The AMPTP president stated that the offer meets the writers’ concerns and expressed hope for a resolution. The companies have also offered certain protections and guidelines regarding AI-produced writing. The proposal includes a wage rate increase and adjustments to residuals and offers more data on streaming platform performance. The AMPTP has also addressed minimum durations and staffing for writers’ rooms. These latest developments in the negotiations occurred on the 114th day of the WGA strike, which has now surpassed the previous strike’s duration. Talks between the management and the WGA have been ongoing since August 11.
Meta, the company behind Facebook, has unveiled SeamlessM4T, an AI model capable of performing speech-to-text, text-to-speech, speech-to-speech, and text-to-text translations in “up to 100 languages.” The model, which combines neural networks for audio and text processing, aims to enhance communication between individuals speaking different languages. The company has released SeamlessM4T under a research license, allowing developers to build on the technology. Meta has also made available SeamlessAlign, claimed to be the largest open multimodal translation dataset thus far, consisting of 270,000 hours of mined speech and text alignments. The dataset is expected to support other researchers’ training of future translation AI models. Meta’s SeamlessM4T offers various features in multiple languages, including speech recognition, text translation, and speech output functions. The model represents a step towards creating a universal translator, similar to the fictional Babel Fish from “The Hitchhiker’s Guide to the Galaxy.”
AI models have traditionally been specialized tools designed for specific applications. However, recent advancements in large language models (LLMs), such as ChatGPT, have transformed generative AI into a versatile “anything tool.” Previously, organizations would invest time and resources into creating custom AI models tailored to their specific needs. For example, Google’s AlphaFold model focused solely on predicting protein folding.
Contrary to expectations, many industries, including robotics, have discovered the power of using off-the-shelf ChatGPT for various applications without requiring specialized training. This shift in AI usability results from changes made to LLMs like GPT3 to enhance their responsiveness to human interaction. These advancements have unintentionally allowed newer models like GPT3.5 and GPT4 to be utilized as general-purpose information-processing tools.
To better understand the mechanics behind generative AI, it is essential to explore the probabilistic nature of LLMs. These models predict the probability of words and phrases based on input, generating an output that is most likely to be appropriate. While generative AI can be unpredictable, its stochastic nature offers unprecedented power and flexibility when harnessed effectively.
LLMs can be continuously adjusted to produce more accurate and meaningful outputs by leveraging techniques like gradient descent during model training. This has allowed AI to be used across various domains, including text, images, and DNA sequences. The transformation of generative AI into an “anything tool” marks a significant milestone in the evolution of artificial intelligence.
Scientists and philosophers have developed a checklist of criteria to evaluate if artificial intelligence (AI) systems are conscious. The 19-strong group proposed 14 criteria that, while they “could suggest but not prove” consciousness, offer a “systematic methodology” to evaluate AI systems. The researchers drew on theories of human consciousness to propose the criteria and then applied them to existing AI architectures. They concluded that none met the criteria, but co-author Robert Long said the work offered “a framework for evaluating increasingly humanlike AIs”. The researchers comprised computer scientists, neuroscientists, and philosophers, with one contemporary calling the inquiry “a deep and nuanced exploration.” The framework is a work in progress and part of a CIFAR-funded project into a broader consciousness test.
This New York Times piece explores the possibility of AI technology replacing human interaction in healthcare and the potential effects on doctor-patient relationships. While AI may help with administrative tasks and triaging patients, the article emphasizes the importance of maintaining human-to-human communication for delivering delicate news and fostering compassionate care. While AI can potentially improve communication with the public and reduce physician burnout, it cannot replace the empathy and art of practicing medicine. The authors argue that teaching empathy and communication skills should remain a priority in medical education and that healthcare professionals should embrace AI while staying true to their values and commitment to patient-centered care. The article calls for a balance between technology and human connection to provide the best care for patients.
Researchers at MIT have developed a new system that could revolutionize machine-learning programs. The technology featured in Nature Photonics uses light instead of electrons to perform computations, resulting in greater efficiency and reduced energy consumption. The system achieved more than a 100-fold improvement in energy efficiency and a 25-fold improvement in compute density compared to current systems. The process leverages hundreds of micron-scale lasers, allowing cellphones and other small devices to handle programs currently only possible at large data centers. Given the scalability and compatibility with existing fabrication processes, commercial use may be possible within a few years. The new technology would bypass the current limitations seen in systems such as ChatGPT and surpass machine-learning models that may otherwise take years to develop. Various institutions supported the research, including the U.S. Army Research Office, NTT Research, and the U.S. National Defense Science and Engineering Graduate Fellowship Program.