PyTorch 2.5 introduced a new CuDNN backend for SDPA, torch.compile region compilation, and the TorchInductor CPP backend. These enhancements improve GPU efficiency, reduce startup times, and boost model performance on advanced hardware, reinforcing PyTorch's leadership in machine learning frameworks.

Google Chrome for Android enhances third-party password manager support in beta version 131, fixing autofill bugs. The native autofill feature for passwords, addresses, and payments will be enabled by default in the stable release on November 12.

Anthropic releases major Claude AI updates for iOS, Android, and iPad, introducing a cleaner interface, enhanced customization, searchable chat history, and tailored instructions to boost productivity.

Tesla's latest Optimus video showcases advances in autonomous navigation, human-machine interaction, and self-charging. Equipped with neural networks and Tesla's AI technology, Optimus navigates factory settings, performs precise tasks, and interacts with humans, positioning it as a potential flagship product.

Microsoft partners with Be My Eyes to enhance AI for 340M visually impaired users by closing data gaps with unique video datasets, ensuring privacy, and advancing inclusive, responsible AI services.

Google's NotebookLM now features customizable audio overviews, enabling users to tailor AI-generated summaries by topic and expertise. It also supports background playback for multitasking. Additionally, Google launched a NotebookLM Business pilot integrated into Google Workspace, targeting organizations with enhanced privacy and security.

OpenAI is testing its Windows ChatGPT app for premium subscribers via the Microsoft Store. Features include file and image uploads, the o1 reasoning model, and multitasking capabilities. A prior security flaw storing chats in plain text was fixed with encryption. Currently available to Plus, Enterprise, Team, and Edu users, with broader release planned later this year.

Google CEO Sundar Pichai appoints Prabakar Raghavan as CTO, succeeding his role in Search and Ads. Nick Fox will oversee Search, Advertising, Maps, and Business. Google also integrates Gemini into DeepMind and Assistant into Platform and Devices amid AI growth and US regulatory scrutiny.

Endor Labs launches Endor Scores for AI Models on Hugging Face, enabling developers to safely select open-source AI models. The tool evaluates security, quality, popularity, and risks using 50 checks, and offers user-friendly search and ranking to enhance efficiency and mitigate vulnerabilities.

Google has petitioned the Ninth Circuit and Judge Donato to suspend changes in the Epic v. Google case, arguing that mandating third-party app stores and removing Google Play billing undermine security, user trust, and developer control, with implementation set for November 1st.

Amazon invests in small modular nuclear reactors with partners to power data centers and AI, aligning with peers like Microsoft and Google seeking carbon-free energy to meet rising demands.

Apple launches the 7th-gen iPad Mini with the A17 Pro chip, boosting CPU by 30% and GPU by 25%. It features Apple Intelligence, Apple Pencil Pro support, enhanced gaming capabilities, and a compact 8.3-inch design. Pre-orders begin October 23, starting at $499.

SearchGPT is gradually expanding its availability to include free users in the US, though access remains limited. Despite initial access issues, the feature is set for broader rollout, promising wider user access soon.

NVIDIA unveiled Llama-3.1-Nemotron-70B-Instruct, outperforming OpenAI and Anthropic models in benchmarks. Based on Meta’s Llama 3.1 with RLHF, it marks NVIDIA’s expansion into AI software, offering powerful, cost-effective solutions via Hugging Face and free hosted services.

Microsoft Research and Tsinghua University unveil Diff Transformer, a new LLM architecture using differential attention to enhance focus on relevant contexts and minimize noise. It outperforms classical Transformers in long-text handling and reduces hallucinations. The scalable design supports multiple modalities, with code released for broader use.

OpenAI research shows usernames with cultural, gender, or racial cues slightly affect ChatGPT responses, especially in older or creative tasks. Reinforcement learning in newer models reduces these biases to below 0.2%, ensuring more equitable interactions.

Pika Labs raised $55M to advance its AI video model Pika 1.5, introducing effects like shatter and "magic disappearing." The web-based tool offers various subscription plans and distinguishes itself amid competition from Runway, Luma, and Adobe’s Firefly Video.

Mistral AI launched Ministral 3B and 8B models, pioneering efficient edge computing for devices. These models enhance real-time AI applications, ensure data privacy, support sustainability, and differentiate Mistral in a competitive market.

Anthropic CEO Dario Amodei presents an optimistic AI outlook, predicting AI-driven advancements in medicine, economics, and global challenges. He emphasizes mitigating risks through robust policies to ensure AI fosters human progress and upholds rights.

At Tech World 2024 in Seattle, Lenovo unveiled AINow AI assistant, Poseidon liquid cooling, and launched key partnerships with NVIDIA, Intel, AMD, Qualcomm, and Meta, advancing AI, data center efficiency, and mixed reality technologies.