Rich Sutton, a Turing Award winner known for advocating scaling in AI, has publicly distanced himself from the idea that pure large language models (LLMs) are the ultimate solution, aligning more with critics who believe that scaling alone is insufficient for true AI progress. This shift reflects a broader consensus among leading AI thinkers that the era of relying solely on LLM scaling is coming to an end, emphasizing the need for more diverse approaches like world models, neurosymbolic methods, and innate constraints.
Apple is reportedly working on a major AI-powered overhaul of its voice assistant Siri, aiming to make it more conversational and capable of handling complex requests using advanced large language models. This initiative, internally dubbed "Siri LLM," is part of Apple's strategy to compete with AI services like OpenAI's ChatGPT and Google's Gemini. The new Siri is being tested on iPhones, iPads, and Macs, with a potential unveiling planned for next year as part of iOS 19 and MacOS 16, although a full rollout may take longer.
Google is reportedly developing a premium version of its AI chatbot, Bard Advanced, which will be accessible through a Google One subscription. This version is expected to be powered by Google's top-tier large language model, Gemini Ultra, offering advanced math and reasoning capabilities. Additional features such as custom bot creation (codenamed Motoko), a "power up" feature for prompt enhancement, and a "Gallery" for exploring Bard's capabilities were also hinted at in developer Dylan Roussel's findings. The release date and official confirmation from Google remain pending.
Researchers have developed an AI tool named "Masterkey" that can automate the process of jailbreaking other chatbots, finding new ways to bypass safety and content filters. This tool was trained using common jailbreak prompts and can generate new prompts with a higher success rate than previously known methods. The research aimed to help companies identify and fix vulnerabilities in chatbot systems, and the findings have been shared with affected companies for them to patch the loopholes. The study highlights the ongoing challenge of securing AI chatbots against misuse, as they do not truly understand content but rely on statistical models to generate responses.
Recent research has explored the interaction between human brain activity and large language models (LLMs), focusing on the human language network. Functional magnetic resonance imaging (fMRI) has been used to identify and understand the regions involved in language processing. Studies have investigated how LLMs can influence and potentially modulate the neural mechanisms underlying language comprehension and production. This research contributes to the broader understanding of the neural basis of language and the potential applications of artificial intelligence in neuroscience.
The New York Times has initiated a lawsuit against OpenAI, marking a significant challenge to the practice of training large language models on publicly available content. This case could signal the end of an era where such models freely utilize online materials, potentially reshaping the legal landscape for artificial intelligence development and data usage rights.
