MiroMind team released the open source bAgent model MiroThinker v1.0, and proposed the concept of "Deep Interaction Scaling". Google develops multimodal voice conversation and Agent import function for Gemini Enterprise.Grok makes a breakthrough in image and video generation technology and releases 4.1 Beta.GMI announces a $500 million investment in an AI data center in Taiwan.
- This summary is generated by AI analyzing the content of the article for informational purposes only.



The MiroMind team recently launched a new open source bAgent model: the MiroThinker The most innovative feature of this model is a new concept: "deep interaction". Scaling (Interactive Scaling)."
Traditional AI models usually follow the linear growth law of "the larger the model size → the better the performance", but MiroThinker breaks through this bottleneck and emphasizes that "the depth and frequency of the model's interactions with the environment" are the key factors for smart growth.
MiroThinker supports multiple interactions and reasoning with external tools (e.g., search engines, Linux sandboxes, speech recognition, etc.), and is able to flexibly use these tools to obtain information and complete tasks. In addition, it has a 256K context length, can process and memorize large amounts of information (hundreds of thousands of words), and can perform 600 tool calls in a single session, enabling the AI to continually use external tools such as search, code execution, computation, translation, etc., as well as to perform complex reasoning and long-duration tasks.
What is "Deep Interaction Scaling"?
In short, performance ∝ depth of model-environment interaction × frequency of reflection. This means:
- Rather than passively absorbing knowledge, models actively interact with their environment;
- Each trial and error and reflection allows the model to "evolve" in the strategy space;
- The more "hands-on" the AI, the more it can correct errors and improve the quality of reasoning.
Just as humans learn complex things through trial and error and hands-on practice, when learning to cook, for example, it's not enough to read a recipe - you have to try, fail, fix, and try again. For AI, multiple rounds of environmental interaction + feedback = the real fuel for intelligence evolution. Each interaction is a "learning" experience, and the intelligence gets stronger and stronger.
As a result, MiroThinker pushes the limits of both "context length" and "number of interaction rounds" to form a true "Thinking Loop", resulting in more efficient and intelligent task processing. "for more efficient and intelligent task processing.
In the latest news, Google is working on a real-time feature called Agentspace for Gemini Enterprise. This feature will provide a multimodal voice experience that will enable users to have voice conversations with agents.
It was revealed that Agentspace will not only support voice conversations, but may also integrate other interaction methods such as text input and image recognition to provide a more comprehensive user experience. In addition, Gemini Enterprise will also introduce "Recommended for You" personalized homepage features to further enhance the user experience.
The launch of these new features signifies Google's continued innovation and technological advancements in the enterprise application space. By integrating multimodal technologies and personalized recommendation systems, Google aims to provide smarter, more efficient and convenient services to business users.
Google is working on a new for Gemini EnterpriseAgent import function, allowing users to integrate external agents into their own multi-agent workflows. The introduction of this feature will greatly enhance the flexibility and scalability of Gemini Enterprise, enabling users to better manage and coordinate multiple automation tasks.
In addition, Google has launchedAgent Assessment Toolthat helps users test and optimize the performance of their agents. Also, the new version will supportscheduled executionfunction, users can set a specific time to trigger the task to improve work efficiency.
The update also includes a series of new connectors that further enrich the integration capabilities of Gemini Enterprise. These new connectors will help organizations more easily interact with a variety of third-party services and systems, enabling more efficient workflows.
With these new features, Google aims to provide users with a more comprehensive and powerful enterprise-grade automation solution that improves the operational efficiency and competitiveness of organizations.
Kosmos is a revolutionary tool for AI scientists, capable of doing the amount of research a human scientist does in about 6 months in a single run.Kosmos comes with a wide range of features:
- Read and integrate thousands of scientific papers;
- Automatically run data analysis code;
- Build hypotheses, validate experiments, and output research reports;
- Even reproducing or discovering brand new scientific research.
Ultra-long contexts are key to its continued operation, and Kosmos can maintain logical coherence across a span of more than 10 million tokens, extracting knowledge from thousands of papers and tens of thousands of lines of analyzed code.
The core innovation is Structured World Models (SWM). Traditional AI scientific tools use a linear way of thinking, input-analysis-output. Kosmos introduces a framework that is closer to human scientific thinking, i.e. Structured World Models. This means that Kosmos can build an ever-expanding knowledge graph in "memory", integrating information from hundreds of autonomous reasoning trajectories to form a dynamic, iterative world model of a specific scientific goal.
For example, if the goal of the study was "to understand the mechanisms of neuronal degeneration in Alzheimer's disease," Kosmos would:
- Search for relevant papers (about 1500)
- Extraction of proteomic, genetic and single-cell data from publicly available databases
- Automatically run over 42,000 lines of analysis code
- Reasoning about critical molecular pathways and generating reports with citations
Kosmos can analyze tens of millions of tokens, far beyond the "memory limit" of existing LLM. This technology not only improves research efficiency, but also provides scientists with new research perspectives.
In the latest news, Google is working on a multi-intelligence system designed to help users refine research ideas through a tournament-like evaluation mechanism. Each run takes about 40 minutes and can generate 100 detailed ideas for a given research topic.
On the Gemini Enterprise platform, Google is developing two new multi-intelligence capabilities:">Idea Generation - "Creating a Multi-Intelligent Body Innovation Session."Partners in Science - "Advancing New Scientific Discoveries with Science Partners."
- Explain your research program to science partners, point to relevant data, and set evaluation criteria.
- A set of intelligences will generate research ideas for you based on the available data.
- The Intelligence Society will evaluate and rank these ideas based on your criteria, using a tournament-style approach.
This advancement marks a new stage in research that will greatly enhance research efficiency and innovation.
UPDATE: The AI Studio platform has added a new Gemini 3 prompting tool. Preparations are currently underway.
For Gemini 3.Best results with default setting of 1.0. Decreasing this value may affect inference performance.
This update is designed to help developers and researchers better utilize the power of Gemini 3 to improve the performance of their models in a variety of tasks. For example, in areas such as natural language processing, image recognition, and speech synthesis, Gemini 3 can provide more accurate results.
- **Natural Language Processing**: improving text generation and comprehension.
- **Image Recognition**: Enhanced image classification and target detection accuracy.
- **Speech Synthesis**: improves speech quality and naturalness.
For more detailed information, please refer to the official documentation.
Recently, an anime video generated by Grok has attracted a lot of attention on the Internet. In the video, the details of the interaction between the character's hands and the utensils and food are barely noticeable, only the movements are slightly stiff.
This technological advancement heralds the coming era of using AI to create film and television-quality animation.Grok Imagine generates cooking scenes that resemble the classic 1990s anime style, presented in an ASMR style that is relaxing and enjoyable.Some seemingly simple actions (such as slicing carrots, kneading dough, beating eggs, stirring or sautéing) that would have been difficult to accomplish not so long ago are now done perfectly.
However, the video has also sparked some controversy. Some viewers gave negative feedback and even cursed, arguing that it was AI copying Miyazaki's work and that the characters lacked soul. This sparked a discussion about the role and limitations of AI in art creation.
- **Technical Advancements**: AI is being used more and more extensively in animation production, and the technical details continue to improve.
- **Artistic Controversy**: some viewers felt that the AI-generated work lacked soul and questioned its artistic value.
- **Future Outlook**: Despite the controversy, the potential of AI in animation production is huge, and more innovative applications are likely in the future.
GMI recently announced that it will invest $500 million in a high-density AI data center in Taiwan. The data center will be equipped with 7,000 NVIDIA Blackwell GB300 GPUs spread across 96 racks and is expected to be able to process nearly 2 million tokens per second. This facility is scheduled to be operational by March 2026.
The data center is expected to consume about 16 megawatts of power.GMI also revealed that they have planned to build a new 50 megawatt facility in the United States to meet the growing demand.
Key features include:
- 96 racks
- Processes nearly 2 million tokens per second
- Electricity demand of about 16 MW
- Scheduled to go live in March 2026
The construction of these facilities not only demonstrates GMI's strong investment in AI infrastructure, but also reflects the growing global demand for high-performance computing resources. With the rapid development of AI technology, these types of data centers will become an important infrastructure for driving technological innovation and adoption.
Grok has made impressive progress in the field of image and video generation. Its latest technological achievements have achieved stunning results, not only generating images and videos of high quality, but also generating them very quickly.
Behind this technological breakthrough is the continuous optimization of deep learning and artificial intelligence algorithms.Grok utilizes advanced neural network models such as Transformer and GAN (Generative Adversarial Network) to achieve high-quality image and video synthesis.
Wide range of application scenarios, including but not limited to film and television production, advertising creativity, virtual reality, etc.. The application of these technologies can greatly improve the efficiency of content creation and provide more possibilities for creators.
Video from @Diesol showing the latest from Grok:
Update: Grok 4.1 Beta is now officially available on the Grok website. In this update, Grok 4.1 Beta is available as a standalone option alongside the existing Grok 4 mode.
Grok is a powerful development tool designed to help developers write and debug code more efficiently. The new release marks further enhancements to Grok's functionality and performance. Users now have the option to test with Grok 4.1 Beta to experience the latest feature improvements and optimizations.
Key features include:
- Enhanced code editor to support more programming languages
- Improved debugging tools with more detailed error messages
- Optimized user interface to enhance user experience
Developers can visit here for more details and to start testing with Grok 4.1 Beta.


Comments are closed.