On June 4-5, 2025, several new developments were released in the AI space: the Gemini-2.5-Pro topped several AI competition lists again; Anthropic launched a new web search tool to improve retrieval efficiency; OpenAI added enterprise features to ChatGPT, such as an internal tool connector and meeting logs; and Morph V2 optimized code modification efficiency with 1600 per second token per second to optimize code modification efficiency. Also of note are Google's release of the Gemma 3 target detection fine-tuning project, an enterprise-grade spreadsheet proxy tool, and @xai's real-time search capabilities.
- This summary is generated by AI analyzing the content of the article for informational purposes only.
In the latest news, Google DeepMind's Gemini-2.5-Pro (version 06-05) is once again at the top of all competitive fields.
The model performs well in many fields such as text processing, visual recognition, web development, etc., especially in the more difficult tasks of coding, math, creative writing and multi-round conversations.first placeThe performance of the program. This update not only improves its performance in programming skills, logical reasoning and creative writing, but also further enhances its ability to handle complex instructions and long queries.
Gemini 2.5 ProIt is one of the smartest models available and excels in several key areas:
- Better programming skills and ability to handle more complex coding tasks.
- Improved logical reasoning skills to better understand and solve complex problems.
- Creative writing skills are enhanced to generate more creative content.
This update marks another important breakthrough for the Gemini series of models in the field of artificial intelligence, providing more possibilities for future applications.
Congratulations to the Google DeepMind team on this significant achievement!
---
Claude Code combined with Opus 4 has been hailed as one of the finest programming tools on the market today, offering an unprecedented perspective on the future of programming.
As an all-encompassing development tool, Claude Code has a superior sense of product design that enables developers to easily tackle a variety of complex application scenarios. Its features include:
- Full support for full-stack development
- Excellent sense of product design
- Delightful user experience
- Simple and easy-to-use interface
Whether you're a beginner or an experienced developer, Claude Code offers an efficient and easy programming experience. If you haven't started using this tool yet, you may already be behind the technological curve.
To get a better visualization of the power of Claude Code and tips on how to use it, you can watch a 72-minute build demo video for more useful tips and best practices.
---
OpenAI recently launched a series of new business features for ChatGPT designed to boost productivity and efficiency in organizations. These new features include internal tool connectors, meeting recording, custom integrations, and an updated enterprise pricing model.
In this update, ChatGPT has added support for a variety of internal tools such as Google Drive, SharePoint, Dropbox and Box. These connectors enable real-time access and provide references for teams, enterprises and education customers, and administrators can control access to the connectors in the settings.
In-depth research functionNow generates long and cited responses, combining internal and web resources for comprehensive analysis, and supports multiple platforms such as Google Drive, SharePoint, Dropbox, Outlook, Gmail, and more for teams, enterprises, education clients, and Pro and Plus users globally (excluding HubSpot users and users in Switzerland, the EEA and the UK).
In addition, the Google Cloud Drive Sync Connector indexes an organization's cloud content in advance, enables fast, high-quality responses to documents, slides, tables, etc. through semantic search, and supports models such as o4-mini, o3, and GPT-4o, which allows administrators to assign permissions based on user needs.
The flexible pricing model for enterprise customers has also been tweaked to introduce a credit-based access mechanism for advanced ChatGPT models and features, where enterprise workspaces can purchase a shared pool of credits from which all users will be deducted when using advanced features, while maintaining unlimited access to core ChatGPT models and features.
---
Our latest release is the Production Ready Spreadsheet Agent tool!
Industries such as audit firms, tax teams, insurance companies and corporate finance departments need to spend more than 10 hours a week manually working through hundreds of spreadsheet files, just copying and pasting numbers. This inefficient way of working not only wastes time, but is also prone to errors.
Our proxy tool utilizes a "parse first, reason later" architecture., the ability to understand visual structure and context, and the following capabilities:
- Converting confusing spreadsheets into a standardized format
- Provide natural language quizzes directly on spreadsheet content
- Handling merged cells and complex layouts
- Maintaining data relationships during conversion
In testing, the agent tool outperformed even the human baseline with an accuracy of 961 TP3T, compared to only 751 TP3T for OpenAI Code Interpreter (GPT 4.1).
For more details, please visit our blog: https://t.co/zO0ZXvsjZ7
The Spreadsheet Agent tool is currently in private preview. For a trial, please contact us at https://t.co/YFCL9weLWU
Also, you can try our state-of-the-art document processing technology, LlamaCloud, by signing up now: https://t.co/D2LIAXZZB0
---
Recently, the Google AI developer team released a new open source project designed to enable image target detection tasks by fine-tuning the Gemma 3 model.
Gemma 3 is a multimodal model developed by Google DeepMind with visual processing capabilities, but in its initial state is unable to perform image-specific tasks such as target detection and semantic segmentation. Inspired by the PaliGemma (@giffmana et. al.) project, this project shows how image-specific tasks can be converted into autoregressive tasks for more efficient task processing.
Key contributors to the project include Sergio Paniego and others, who together created this lightweight codebase to make it easy for users to quickly adjust parameters and begin the fine-tuning process. Specific steps include:
- Create a target detection dataset labeled with ";
- Download the Gemma-3 model from Hugging Face Hub;
- Fine-tune the model;
- Visualize the predicted results.
Although the quality of the images generated so far is not ideal, Gemma 3 has been able to generate location labels with only a small number of data points, which is enough to construct basic bounding boxes. This result prompted the team's decision to make this project public.
future plansIn the middle, the team hopes to further improve the results through optimization and welcomes community members to contribute code or suggest improvements.
---
Morph V2, a program that can process1600 tokensThe code revision model is now available on the playground platform. The model is designed to improve the efficiency of code modification and is applicable to a wide range of domains, including software development, code review, and automated testing.
Morph V2 utilizes advanced natural language processing technology to quickly understand and generate high-quality code snippets. Its efficient processing speed enables developers to complete a large number of code modifications and optimizations in a short period of time.
Key Features:
- Processes 1600 tokens per second, dramatically increasing the speed of code modification;
- Multiple programming languages are supported, including Python, JavaScript, Java, and more;
- Provides an intuitive user interface that is easy to get started.
This new tool will greatly enhance developer productivity and is expected to become an important tool for code management and optimization.
---
Anthropic has recently released a new web search tool designed to provide users with a more efficient and accurate information retrieval experience. The tool utilizes advanced natural language processing technology to understand complex query statements and quickly sift through massive amounts of data to find the most relevant results.
This search tool is not only for general users, but also particularly suitable for researchers and business users. It optimizes the quality of search results through deep learning algorithms that can significantly improve productivity.
Key features include:
- Intelligent semantic understanding: able to accurately parse the user's query intent
- Multi-language support: covering a variety of common languages to meet the needs of global users
- Real-time updates: Continuously updating the data source to ensure the timeliness of information
This innovative tool from Anthropic breaks new ground in modern information retrieval and deserves attention.
---
Recently, artificial intelligence technology company @xai announced the launch of its latest real-time search feature. This new feature provides users with more accurate and instant search results by combining advanced natural language processing and machine learning algorithms.
Real-time search not only responds quickly to the user's query needs, but also makes personalized recommendations based on the user's search history and behavioral patterns, greatly enhancing the user experience.
Key features include:
- Real-time feedback: After the user inputs keywords, the system immediately returns relevant results.
- Intelligent recommendation: provide personalized suggestions based on users' historical search records and behavioral data.
- Multi-modal support: Supports multiple content forms such as text, image and video search.
The function has a wide range of application scenarios, covering e-commerce, online education, news and information, and many other fields. For enterprises, real-time search can not only improve user satisfaction, but also stand out in the competitive market.
With the continuous progress of technology, the real-time search function will be further optimized to provide users with more intelligent services.
---
v0 The latest version now supports inline code generation.
This update allows users to press the ⌘+K shortcut to directly modify segments of code known to be in need of change when making quick and precise code edits. This not only improves productivity for developers, but also reduces the likelihood of manual typing errors.
Key features include:
- Real-time code generation: no need to switch to other tools or pages, code changes are done directly in the current editor.
- Intelligent Suggestions: The system is able to provide relevant code snippets based on the context, further increasing the speed of development.
- Compatible with a wide range of programming languages: Whether it's JavaScript, Python, or one of the other major languages, v0 provides support for it.
Additionally, the v0 team has released a video showing how to use the feature in action. Interested developers can learn more at the following link: watch the demo video.
---
In this hands-on example, we will show how to automate SEC Form 4 extraction using LlamaExtract and agent workflows.
SEC Form 4 plays an important role in market transparency by requiring company officers, directors, and large shareholders to disclose their stock transactions. Not only does this help regulators monitor insider trading, but it also provides an important source of information for investors.
In this Colab, we will:
- Creating an extraction agent that can extract structured information from unstructured Form 4 files
- Build a scalable workflow to extract trading information from SEC Form 4 filings for all companies in the Dow Jones Industrial Average
In this way, we can efficiently process large amounts of data and ensure the accuracy and consistency of the information.
Explore the full example: https://t.co/fJwoSHYfvo
Comments are closed.