Generative AI

Ant Group's Robbyant has released LingBot-VLA 2.0, an open-source Vision-Language-Action (VLA) foundation model specifically engineered for robots. This release includes a technical report, an Apache-2.0 licensed codebase, and a 6 billion parameter checkpoint, aiming to bridge the gap between VLA models that perform well in laboratory settings and those that succeed in practical deployment. LingBot-VLA 2.0 enhances its predecessor by focusing on three key areas: improved generalization capabilities, an expanded action space for robots, and more sophisticated predictive dynamics modeling. LingBot-VLA 2.0 functions as a generalist robot policy, leveraging a vision-language backbone to interpret camera images and textual instructions, translating them into executable robot actions. The publicly available model, lingbot-vla-v2-6b, is a 6B parameter checkpoint that utilizes Qwen3-VL-4B-Instruct as its vision-language model backbone. Training is further refined through distillation from two teacher models: LingBot-Depth and DINO-Video. For inference, a single call takes approximately 130 milliseconds on an NVIDIA GeForce RTX 4090D, utilizing 10 denoising steps. The action expert component is built with a Mixture-of-Experts (MoE) architecture to facilitate scaling. The model's generalization is significantly supported by its extensive training data. The research team curated approximately 60,000 hours of pre-training data, comprising 50,000 hours of robot trajectories and 10,000 hours of egocentric human videos. This robot data encompasses 20 distinct robot configurations, ranging from single-arm setups to full humanoid robots. The initial data pool was even larger, with around 90,000 robot hours and 20,000 egocentric hours. A refined data pipeline filters out noisy samples to ensure a high-quality dataset for training. This filtering process involves computing third-order jerk, velocity, and acceleration Z-scores for each embodiment, and episodes exhibiting abnormal smoothness or over 95% static signals are discarded. Videos are cross-referenced with replayed states using each robot's URDF, and annotators remove issues like blur, occlusion, dropped frames, and multi-view misalignment. Egocentric clips undergo VLM filtering, followed by egocentric SLAM and MANO hand-pose reconstruction.

SpaceXAI released Grok 4.5 this week, a new artificial intelligence model designed for coding, agentic tasks, and knowledge work. The company stated that Grok 4.5 is its smartest model to date and was trained in conjunction with Cursor, an AI-powered coding editor. This collaboration suggests a focus on enhancing developer productivity and complex problem-solving capabilities. Grok 4.5 demonstrates significant improvements in token efficiency, reportedly using approximately 4.2 times fewer output tokens than Opus 4.8 on the SWE Bench Pro benchmark. The model is priced at $2 per million input tokens and $6 per million output tokens, with a serving speed of 80 tokens per second. SpaceXAI also highlighted Grok 4.5's #1 ranking on Harvey’s Legal Agent Benchmark, indicating its strength in office-related tasks and legal applications. It is now the default model within the Grok Build environment. SpaceXAI trained Grok 4.5 on extensive datasets covering coding, science, engineering, and mathematics, utilizing tens of thousands of NVIDIA GB300 GPUs. The training process incorporated advanced techniques for large-scale runs, including rigorous data filtering, deduplication, quality scoring, and domain-specific selection. Reinforcement learning was scaled to hundreds of thousands of tasks, with a particular emphasis on multi-step software engineering and technical problem-solving. The training stack supports highly asynchronous operations, allowing agentic rollouts to continue learning for extended periods. Benchmark performance data released by SpaceXAI shows Grok 4.5 performing competitively against leading models. While the company's internal charts indicate that Fable (max) achieved the highest scores across four coding benchmarks, Grok 4.5 was noted as being the closest competitor on the Terminal Bench 2.1. The model's reasoning capabilities are described as both intelligent and efficient, positioning it as a powerful tool for specialized engineering and knowledge-intensive work.

Google AI Studio launched an 'import from GitHub' feature within its Build mode on July 8, 2026, enabling developers to convert existing code repositories into runtime-compatible applications. This new functionality allows users to import a GitHub repository directly into AI Studio, where it is transformed into a format compatible with the platform's runtime. Once imported, developers can continue to iterate on the application within AI Studio, deploy it, and further refine its functionality. This feature aims to streamline the development process by providing a pre-existing codebase as a starting point, rather than requiring users to begin with a blank prompt. The 'import from GitHub' feature is integrated into Build mode, which is Google AI Studio's environment for 'vibe coding,' where users can describe an app in a prompt, and Gemini generates a full-stack application with a live preview. Developers can then refine the generated app through chat or annotation modes. For applications utilizing the Gemini API, Google AI Studio automatically configures the GEMINI_API_KEY as a server-side secret, ensuring that API keys are not exposed in client-side code. This approach emphasizes security by keeping sensitive credentials on the server. The process involves importing the repository, followed by iteration and deployment within the AI Studio environment. While the specific internal steps of the importer have not been fully detailed by Google, the general flow involves reading the repository, adapting it to the AI Studio runtime, and then opening it in the Build interface for further development. The company shared this update through its official Google AI Studio account and via Logan Kilpatrick, who leads the product.

OpenAI released GPT-Live, a new generation of voice models designed for natural, real-time conversation, powering the ChatGPT Voice experience globally starting today. The initial rollout includes two versions: GPT-Live-1 and GPT-Live-1 mini. These models feature a full-duplex architecture, allowing them to listen and speak simultaneously, incorporating natural interjections like 'mhmm' or 'yeah' and maintaining conversational flow even during complex queries. When a conversation requires web search, deeper reasoning, or intricate tasks, GPT-Live delegates these operations to a frontier model operating in the background. At launch, this background model is GPT-5.5. While GPT-5.5 processes the request, GPT-Live continues the conversation, preventing conversational pauses and maintaining engagement. Human tests indicated that GPT-Live-1 and GPT-Live-1 mini were strongly preferred over the previous Advanced Voice Mode. Previous voice systems faced limitations. Cascaded voice systems chained multiple models (speech-to-text, LLM, text-to-speech) for each turn, leading to potential information loss and slow, stilted responses. Turn-based models, such as ChatGPT's Advanced Voice Mode, processed audio within a single model to reduce latency and improve conversational smoothness. However, these still operated in discrete turns, requiring users to complete their speech before the AI responded. GPT-Live's full-duplex capability addresses these limitations by enabling continuous listening and speaking. This allows for more dynamic and natural interactions, moving beyond the constraints of turn-based communication. While video, screen sharing, and full multilingual parity are not included at launch, OpenAI plans to release an API for GPT-Live in the near future. The company aims for GPT-Live to facilitate more seamless and human-like interactions with AI.

Researchers have introduced novel frameworks designed to enhance the data processing capabilities of artificial intelligence agents. These new systems aim to equip AI agents with a more sophisticated understanding of complex data structures, thereby improving their efficiency and accuracy in performing a wider range of tasks. The development focuses on enabling agents to interpret and utilize information that is not presented in simple, linear formats, such as unstructured text, images, and sensor data. One key aspect of this advancement involves the integration of advanced natural language processing (NLP) and computer vision techniques. These allow AI agents to not only read and comprehend textual information but also to analyze visual content and extract relevant data points. This multimodal understanding is crucial for agents that need to interact with the real world or process diverse datasets. For instance, an agent equipped with these capabilities could analyze a financial report that includes charts and graphs, extracting both numerical data and visual trends. The improved data understanding is expected to unlock new applications for AI agents across various sectors. In scientific research, agents could sift through vast amounts of experimental data, identifying patterns and anomalies that human researchers might miss. In customer service, agents could better understand user queries, even when phrased ambiguously or accompanied by visual aids, leading to more effective problem resolution. The goal is to move beyond simple command-response interactions towards more autonomous and context-aware AI behavior. These advancements are part of a broader push in the artificial intelligence community to create more capable and versatile AI systems. By focusing on the fundamental ability of agents to process and interpret data, developers are laying the groundwork for more sophisticated AI applications. The ongoing research aims to refine these frameworks, making them more robust, scalable, and adaptable to new data types and task requirements, ultimately leading to AI agents that can operate with greater autonomy and intelligence.