5/25/2025 AI Express | New AI Developments: Google Releases Gemma Model, Microsoft Open Sources Qlib

Langflow is a powerful visualization tool for building and deploying AI-based agents and workflows without writing any code.

The tool supports all major large-scale language models (LLMs) and vector databases, providing developers with great flexibility. With an intuitive drag-and-drop interface, users can easily create complex AI application scenarios covering a wide range of domains from customer service to automated decision making.

Langflow's core strengths include:

• Completely open source with over 62,000 stars of support.
• Supports multiple mainstream technology stacks to ensure seamless integration with existing systems.
• Provides rich predefined components to accelerate the development process.

Langflow is certainly an option to consider for businesses and individuals looking to quickly implement an AI solution.

Link to project: https://t.co/v3Gv4smWbv

Infinite tool use is a thought-provoking perspective.

Openai o3, which only realizes "full tool use" and can only call 8 internal tools, already makes every day work more efficient. If this vision is realized in the future, its potential is unlimited.

Core propositionIt is the ideal large-scale language model (LLM) that only outputs tool calls and their parameters during reasoning or authoring; all specific states (e.g., memories, drafts, intermediates, goals, etc.) should be externalized into dedicated tools. In this way, the model is responsible for immediate reasoning and the tool is responsible for persistent, structured task states, realizing the division of labor and complementarity between the two.

**Why "infinite instrumentalization"? **

• Externalized Intelligence: Handing off complex operations to efficient, domain-specific programs significantly reduces contextual window pressure and improves the quality of results.
• Explicit forgetting & backtracking: With editing tools, the model can be deleted and refined at any time, without having to generate a long text positively all at once, thus reducing the accumulation of errors.

**Example Scenarios** include:

• Text writing/editing: Human writing is often "jump-revise-back"; with a full-featured text editor, LLMs can alternate between multiple levels of granularity as humans do, avoiding the limitations of "forward once" writing.
• 3D Generation: With CAD libraries + viewing/rotation tools + note files, models can be iteratively shaped into 3D objects at any scale, rather than being constrained by a fixed voxel context.
• Long video comprehension: The model can repeatedly "watch" clips, record/edit notes, and analyze hours or even days of video in a limited window.

In addition, unlimited instrumentation has security and interpretability advantages. Full version control and editing trajectories are visible to both humans and models, making potential spoofing more costly in terms of "inference budget" and increasing the credibility and auditability of the output.

**Training Ideas** relies heavily on reinforcement learning to train the model to learn to plan and correct in the tool space. Since the model has a finite context, it is possible to generalize to an "infinite" number of contexts as long as they can be recovered and edited from a variety of starting points.

**Architecture conception** favors structures with constant or upper inference budgets/tokens, such as sliding window attention, SSM, RWKV, Mamba, etc., and encourages "forgetfulness" to promote role separation from tools. The authors propose a "reverse hybrid" concept: a localized high-precision sliding-window attention layer, followed by a long-range SSM layer, to achieve a dual macro + micro perspective.

Currently, Tool-use is already available in o3, Agentic RAG, etc., but it is still limited to short contexts and partial processes. The authors call for all interactions with the external world to be done through tools, and to move towards true "infinite tooling" using a model of cheap storage + limited reasoning costs.

Recently, Google launched three variants based on the Gemma model:MedGemma,SignGemmarespond in singingDolphinGemma. These models are optimized for different application scenarios respectively.

MedGemma is a model designed for the medical field and is offered in two versions:

• 4B Multimodal Model: This model is capable of handling combined image and text tasks and is particularly suitable for medical data. It has been pre-trained for chest X-rays, dermatology images, ophthalmology images and pathology slides.
• 27B Textual Inference Model: this version has more powerful plain text processing capabilities for in-depth inference tasks. It is also pre-trained for medical text and image data, covering a wide range of medical data such as radiological images, pathology slides, ophthalmology images, dermatology images, etc.

SignGemma, on the other hand, is a sign language interpreter model that supports multiple languages and enables sign language to text or speech conversion, which greatly improves the communication efficiency of the deaf community.

DolphinGemma is a unique sound synthesis model focused on generating dolphin sounds. This model has potential applications in marine life research and conservation.

The launch of these three models not only demonstrates Google's technical strength in the field of artificial intelligence, but also further promotes innovation and development in related fields.

On GitHub, an organization calledAgenticSeek's open source project has attracted a lot of attention. The project provides a 100% localized intelligent agent system designed to replace Manus AI.

Based on local inference models and intelligent agent technology, AgenticSeek is able to autonomously browse the web, write code, plan tasks, and support voice interaction functions, bringing users an experience similar to that of Jarvis in the Iron Man movies.

• Fully localized operation:All data is retained on the device with zero cloud dependency.
• Smart Web Browsing:Automatically search for information and fill out forms.
• Autonomous Programming Assistant:Support for Python, C, Go, Java and many other languages.
• Intelligent agent selection system:Automatically matches the most appropriate task processing.
• Voice Interaction Features:Supports voice input and text-to-speech output.
• Complex mission planning execution:Ability to break down large tasks into steps for completion.

AgenticSeek provides a detailed installation and configuration tutorial, and it is recommended to use at least DeepSeek 14B or higher parameter model, which requires some computer performance. For developers and researchers, AgenticSeek is a valuable tool that can greatly improve work efficiency and protect data privacy.

When writing automation scripts using Selenium, developers often encounter element loading timing issues that cause tests to fail. In addition, manually handling various browser drivers can be a tedious and error-prone task.

In order to address these challenges.SeleniumBasewas born. It's a powerful Python-based framework designed to simplify the process of web automation testing. It makes complex browser operations simple and reliable by deeply encapsulating Selenium with several built-in utilities.

The main features of SeleniumBase include:

• Intelligent waiting mechanism that automatically solves the problem of unstable page loading;
• Support a variety of authoring methods , such as BaseCase class inheritance and SB context manager , to meet different development needs ;
• UC Mode stealth mode to help bypass anti-crawl detection such as Cloudflare and improve test success;
• Automatically generate detailed test reports and real-time dashboards to visualize test results for team collaboration and issue tracking;
• Automatically manages browser drivers without manual intervention, reducing maintenance costs.

For professionals engaged in web application development or testing, using SeleniumBase can significantly improve work efficiency and reduce errors caused by improper environment configuration.

GitHub project address: https://github.com/seleniumbase/SeleniumBase

Microsoft's open source on GitHubQlib Project, probably the most complete AI quantitative investment platform in the industry at present, has wildly gained 20k+ stars as of now. Based on a powerful data processing engine and a rich model library, Qlib covers a variety of paradigms such as supervised learning, reinforcement learning, etc., helping developers quickly build a complete quantitative research process from idea to production.

The main features of Qlib include:

• Efficient data servers that process dozens of times faster than traditional databases;
• Integrate 30+ state-of-the-art quantization models, including Transformer, LSTM, LightGBM, and more;
• Reinforcement learning trading strategies are supported, capable of modeling a continuous investment decision-making process;
• Covering the complete investment chain: factor mining, risk modeling, portfolio optimization, order execution;
• Provide automated research tool qrun, one click to complete the whole process of modeling to backtesting;
• Supports market dynamic adaptation and concept drift processing to improve model stability.

Qlib not only provides a detailed deployment and installation tutorial, but is also particularly suitable for developers who want to work in the field of financial AI for research and learning.

Qlib is a valuable tool for fintech companies and individual investors to better understand and apply AI technology to real-world scenarios in the financial markets.

UQLM is a library for detecting illusions in large language models (LLMs), quantifying uncertainty through multiple scoring methods. This technique is crucial for improving the reliability and accuracy of AI applications.

UQLM integrates seamlessly with the LangChain platform, making it easier for developers to build and deploy more reliable AI applications. The library supports a variety of scoring methods including, but not limited to, probability-based methods, calibration-based methods, and confidence-based methods.

application scenario::

• Error Detection in Natural Language Processing (NLP) Tasks
• Increased Reliability of Chatbots and Virtual Assistants
• Quality control of automated text generation

The emergence of UQLM provides a powerful tool for solving the uncertainty and illusion problems of AI models in practical applications, and helps to promote the further development of AI technology.

For more information, please visit: UQLM official website

It's only been four days since Google unveiled the Veo 3, and the new product is already making a huge splash around the world.

Users are already creating incredible content with Veo 3. Here are 13 of these amazing examples:

• Self-aware AI characters: these characters are capable of self-learning and evolving, demonstrating unprecedented interaction capabilities.

Veo 3's groundbreaking technology not only revolutionizes visual effects, but also reveals new possibilities in the field of artificial intelligence and machine learning. The availability of this tool allows creators to express their creativity in ways never before possible, whether it's game development, virtual reality or filmmaking.

application scenarioIncluding but not limited to:

• Highly intelligent NPCs (non-player characters) in game development
• Immersive Interaction in Virtual Reality Experiences
• Realistic Character Generation for Movie Effects

As more users begin to explore the features of the Veo 3, there is reason to believe that this product will continue to lead the way in innovation for years to come.

Microsoft's classic apps Notepad and Paint have recently received new artificial intelligence (AI)-powered features that offer users more creativity and efficiency gains.

Notepad now integrates an AI assistant called "Write" to generate and optimize text content. This feature not only helps users to write documents quickly, but also improves the quality and efficiency of writing through natural language processing.

PaintThe new "AI Stickers" feature allows users to automatically generate personalized stickers through simple input or sketches. This feature uses image recognition and model generation to enable users to quickly realize their ideas and customize them according to their needs.

In addition, the Snipping Tool has been upgraded with a "Perfect Screenshot" feature that automatically optimizes the quality of the screenshot, as well as a convenient built-in color picker that makes it easy to select and use colors in the screenshot.

These new features not only enhance the user experience, but also demonstrate the potential of AI technology to be widely used in everyday applications.

In exciting news for business and personal users, it is now possible to connect Box and Dropbox accounts to ChatGPT's Deep Research feature. This new feature allows users to access files and data stored in Box and Dropbox directly while using ChatGPT.

With this integration, users can more easily leverage ChatGPT's powerful natural language processing capabilities to analyze, summarize and extract data stored in the cloud. This is especially useful for professionals who need to perform extensive text analysis and research.

Application scenarios include:

• Market Research
• Analysis of Academic Literature
• Internal Document Management

This seamless integration not only improves productivity, but also provides users with a new perspective on data analysis.

Both business and individual users can better manage and utilize their data resources with this feature. For more details, please visit:https://t.co/owNtgrPSV8