OpenAI releases GPT-4.5, its largest AI language model yet, initially available to ChatGPT Pro users. It offers improved writing skills, global knowledge, and a more natural conversational style, though it lacks cutting-edge features compared to other models. Trained with synthetic data and advanced techniques, GPT-4.5 generates fewer fabrications and excels in emotional nuances. OpenAI plans to roll it out to more users soon and hints at GPT-5 later this year, aiming toward AGI development.

China-led IEC 63310 standard for elderly care robots in connected home environments has been officially released. It addresses elderly physiological, psychological, and behavioral needs, providing comprehensive guidelines for design, manufacturing, and testing. The standard emphasizes health monitoring, emergency alerts, communication support, household assistance, and data privacy. This advancement boosts global elderly care robotics development and supports aging populations.

Octave TTS by Hume revolutionizes text-to-speech technology by focusing on context, emotion, and voice customization. Built on advanced LLMs, it goes beyond literal text conversion to deliver nuanced, expressive speech tailored for various scenarios. Featuring voice design, performance directives, and strong evaluation results, Octave sets new standards in expressive TTS. Future updates include voice cloning capabilities.

DeepSeek announced three open-source projects to optimize deep learning model training. DualPipe uses dual-channel parallel processing for efficient resource allocation, EPLB enhances distributed training with smart scheduling, and profile-data provides V3/R1 model performance insights. These initiatives aim to reduce costs, improve efficiency, and lower hardware reliance, advancing AI development in China through an open ecosystem.

IBM unveiled the Granite AI model family with experimental reasoning, vision, and predictive capabilities. Key models include Granite 3.2 Instruct (8B/2B versions) for tasks like summarization and code generation, featuring efficient "chain-of-thought" reasoning. The lineup also introduces Granite Vision 3.2 for visual document understanding, Guardian 3.2 for risk detection with verbalized confidence levels, and compact timeseries models. These models are available on platforms like Hugging Face

Nvidia CEO Jensen Huang remains optimistic about the company's growth, dismissing concerns over DeepSeek's impact. He highlighted R1 as a positive innovation driving demand for inference models, which require significant computing power. Nvidia reported record revenue of $39.3 billion last quarter and forecasts $43 billion for the next. The data center business grew to $115 billion in 2024, with strong demand for the new Blackwell chip. Major tech firms plan massive AI infrastructure investments

ElevenLabs, an AI startup known for audio generation, raised $180 million and launched Scribe, its first standalone speech-to-text model. Valued at $3.3 billion, the company aims to compete in speech recognition with support for 99+ languages. Scribe boasts excellent accuracy (<5% error rate) in 25 languages, outperforming models like Google's Gemini 2.0 Flash. Features include speaker identification, word-level timestamps, and sound event tagging. Currently limited to pre-recorded audio, a real

Microsoft launched two new compact models, Phi-4-multimodal and Phi-4-mini, available on Azure AI Foundry and Hugging Face. Phi-4-multimodal excels in speech, translation, and multimodal data handling, while Phi-4-mini prioritizes speed and efficiency for resource-constrained environments, expanding Microsoft's AI offerings.

Amazon launched Alexa Plus, an upgraded version of its smart assistant, offering features like concert ticket searches and Uber bookings with more natural conversation capabilities. It will be free for eligible U.S. Echo Show users until March 2025, then cost $19.99/month unless bundled with Prime membership. Compatibility includes most Alexa devices, excluding older models and Astro robots.

OpenAI is reportedly set to launch GPT-4.5, with clues found in ChatGPT's Android version under the codename "Orion." Sources suggest the release may occur in the coming days, marking a key advancement in natural language processing, though official confirmation is pending.

Nvidia predicts a future with a billion robotic vehicles. Data center revenue surged 93% YoY, driven by Blackwell and Hopper chips. Blackwell's deployment is the fastest in Nvidia's history. Automotive revenue hit $570 million, up 27% QoQ. Gaming revenue was impacted by supply constraints but expects growth with RTX 50 series launch. AI inference demand accelerates as new models emerge. China sales remain below pre-export control levels.

DeepSeek Open Platform introduces nighttime discounts from 00:30 to 08:30 Beijing Time, reducing API call costs. DeepSeek-V3 prices are halved, and DeepSeek-R1 costs drop to 25% of original. This aims to optimize resource use and enhance efficiency while meeting user cost management needs. Users can plan API usage for better cost-effectiveness.

Alibaba Cloud has open-sourced Wan2.1, an advanced video generation model with text-to-video and image-to-video capabilities. It offers a professional version with 14 billion parameters for complex tasks and a fast version with 1.3 billion parameters for consumer-grade GPUs. Built on Causal 3D VAE and Video Diffusion Transformer architectures, Wan2.1 ensures coherent content generation and supports video editing, text-to-image, and video-to-audio tasks. Available under Apache 2.0 license on GitH

Apptronik, a Texas-based humanoid robotics company, announced a pilot partnership with Jabil to test its Apollo robots in manufacturing settings. Following a $350M Series A funding round, this collaboration aims to scale production and integrate humanoid robots into industrial processes. Jabil will trial Apollo systems for logistics tasks, with potential future in-house robot manufacturing. Apptronik, with roots in UT Austin and NASA projects, plans commercial production by 2026 while competing

Google launched a free version of Gemini Code Companion, offering up to 180,000 monthly code completions and AI-driven code reviews for GitHub. It supports global access with a Gmail account and enhances code quality through automated feedback. Based on Gemini 2.0, it features a large context window and robust IDE integration, aiming to democratize advanced AI coding tools while emphasizing the importance of human oversight in innovation.

OpenAI expands Deep Research to ChatGPT Plus subscribers for $20/month, previously exclusive to the $200/month professional plan. Plus users get 10 queries/month, while professional users' limit increases to 120. Upcoming access for Team, Edu, and Enterprise plans. Enhanced features include image embedding and improved file analysis. Free users will have delayed access due to resource demands.

Microsoft has removed all usage restrictions on OpenAI's o1 inference model for Copilot users, enabling unlimited access to voice features and the "Think Deeper" capability. Previously limited for free users, these enhancements now allow extended AI interactions. Potential delays may occur during high demand or due to security concerns. This change follows Microsoft's recent integration of Office AI into Microsoft 365 and comes two years after Copilot's debut in Bing. A $20/month Copilot Pro sub

DeepSeek unveiled DeepGEMM, a CUDA-based matrix multiplication library optimized for NVIDIA Hopper architecture. Designed for FP8 computations, it offers high performance in regular and MoE group GEMM tasks, achieving up to 2.7x speedup over CUTLASS 3.6 on H800 GPUs. Featuring JIT kernel generation, secondary accumulation for precision, and TMA integration for efficiency, DeepGEMM supports large models like DeepSeek-V3/R1 and is open-sourced under the MIT license. While not surpassing expert-tun

OpenAI now allows all free users to access its premium voice mode using the GPT-4o mini model, providing natural conversations at lower costs. Plus subscribers get five times higher usage limits and retain full GPT-4o features, while Pro users enjoy unlimited access with enhanced capabilities.

Boston Dynamics' AI Institute purchased robots from Chinese innovator Unitree Robotics for testing. Founder Marc Raibert highlights China's AI advancements, noting Unitree's impressive humanoid robots. He stresses the importance of hardware and software in AGI development but cautions that predicting AGI's arrival is complex due to ethical and technical challenges. Collaboration is key to addressing AI issues like model hallucinations.