By building @v0's full stack application on the Vercel platform, we were able to experience the rapid development process of a successful startup on our platform. This process not only validated our platform, but also taught us many valuable lessons.
The key is to optimize the speed of iteration. We discovered that Large Language Models (LLMs) can write not only poems and assignments, but also React and Tailwind code, and Jared Palmer came up with an interface for "Text to React Components", which was quickly rolled out.
- **Model innovation**: a continuously improving language model
- **Better Cognitive Architecture**: Improving the ability of models to understand and handle complex tasks
- **Prompting techniques**: more effective prompting methods
- **Retrieval solution**: Enhanced information retrieval capabilities for models
- **Tool call**: enabling more efficient tool integration
- **Integrated program**: providing richer third-party service support
To accommodate the future of AI, we have developed @aisdk as a common interface for AI models to ensure that the future of intelligence is not monopolized by a single vendor.
Currently, we are still using a single Postgres instance, but expect to have to shard and replicate within 4 weeks to cope with increasing demand. Looking back, we made the right decisions to ensure the continued growth of the product by quickly iterating on the user experience and replacing some of the backend components in the process.
This was the underlying philosophy when we started @nextjs and @vercel, and the reason we open sourced the tools and best practices in the first place. Stay tuned for further evolution of our database architecture.
---

The latest news shows that the Gemma 3 has passed theUnslothAIThe platform has been fine-tuned and utilizes the same features provided by theFirecrawl_devCustomized instruction datasets provided by the team.
This fine-tuning process allowed Gemma 3 to be used in the processing of the same as theOpenAI Agents SDKThe performance of document-related queries is even better. This not only improves the accuracy and responsiveness of the model, but also provides developers with more convenient tool support.
Specific application scenariosIncluding but not limited to:
- Quick access to key information in the SDK documentation
- Solve technical problems encountered in the development process
- Improve code writing efficiency
In addition, with this fine-tuning, Gemma 3 is able to better understand natural language, thus providing more precise assistance to users.
Interested developers can view the demo at the following link: https://t.co/cgVETkbixC.
---

The Anthropic team recently released a detailed report summarizing their experience using Claude Code internally. This report explores in detail the best practices for effectively utilizing the Large Language Model (LLM) for programming.
Here are a few of the most effective patterns they found:
- Clear and specific instructions: Providing clear task descriptions helps improve the quality of code generation.
- Progressive refinement: Starting with high-level concepts and progressively refining them to specific implementation details, which allows LLM to better understand the context.
- Multiple Iterations: Continuously improve the accuracy and efficiency of your code through multiple feedback and adjustments.
- Code Samples: Provide relevant code snippets as reference to help the model better understand and generate the required code.
- Bug Fixing: Utilize LLM for code debugging and bug fixing to improve development efficiency.
These methods are not only applicable to Claude Code, but also to other LLM-based programming tools.Anthropic hopes that by sharing these experiences, it will help developers more effectively utilize AI technology to improve programming efficiency.
---

BREAKING: Google has just launched a new tool called Gemini 2.0 Flash, which makes traditional image editing software such as Photoshop look obsolete.
With Gemini 2.0 Flash, users can realize complex image editing with simple text commands. No need for layers, brushes and other tedious operations, just enter a text description and the powerful AI technology will automatically complete the photo modification and adjustment.
This new way of image processing not only greatly simplifies the editing process, but also lowers the threshold for users, making it easy for more people to perform high-quality image editing.
Here are 12 amazing examples:
- Changing the weather in a photo
- Adding or removing background elements
- Adjusting the character's expression and posture
- Replacing objects in a photo
These examples fully demonstrate the power and potential of Gemini 2.0 Flash for a wide range of applications in image editing. Both professional designers and ordinary users can benefit from them.
With the continuous development of artificial intelligence technology, image editing will become more intelligent and convenient in the future, and Gemini 2.0 Flash is a reflection of this trend.
---

Microsoft recently open-sourced an efficient inference framework: bitnet.cpp, which is capable of running 1-bit model inference, such as BitNet b1.58, directly on the CPU.
Provides a variety of optimized kernels to support fast and lossless inference on the CPU, enabling model speedups of 1.37-6.17x while reducing energy consumption by up to82.2%. This allows a single CPU to run models up to 100B in size.
GitHubThe address is https://github.com/microsoft/bitnet.
The official 2B parameter model is provided for direct use and supports a wide range of 1-bit models on Hugging Face. In addition, complete example code and a detailed usage guide are provided for Windows, Linux and macOS.
In practice, this low-precision computation not only improves the inference efficiency of the model, but also dramatically reduces the demand for hardware resources. This is a very attractive option for enterprises that need to deploy AI models on a large scale.
The video demo shows the effects of bitnet.cpp running the BitNet b1.58 3B model on an Apple M2, further demonstrating its performance on real hardware.
---

For the first time in history, Artificial Intelligence (AI) has surpassed the Intelligence Quotient (IQ) of the average human. This breakthrough was made possible by the Claude series of models developed by Anthropic.
Claude-1, when it was released in March 2023, performed almost indistinguishably from a randomized level, answering six questions correctly with an IQ of about 64. However, subsequent releases have made significant progress:Claude-2Released in July 2023, the score increased by 6 points per test (approximately 18 IQ points).Claude-3Released in March 2024, it again adds 6.5 points and about 19 IQ points, putting its IQ above the human average.
These symmetrical increases have sparked speculation as to whether Anthropic is basing its releases on internal benchmarks, which happen to correlate closely with this IQ measure.
Based on the current trend of progress, it is expected thatClaude-4will be released in 12 to 16 months and will be able to answer approximately 25 questions correctly per test with an IQ score of 120. after that, in another 16 to 32 months, theClaude-5will be able to answer about 31 questions correctly at a time, with an IQ of about 140. eventually, in another 20 to 64 months.Claude-6Should be able to answer all questions correctly and have an IQ that surpasses almost all humans. Add it all up and it would take about 4 to 10 years total.
While such progress is not inevitable, and Anthropic may encounter issues such as budgetary constraints, energy constraints, and regulatory limitations, the continued advancement of Moore's Law suggests that there is a good chance that these models will be sustained.
This achievement not only marks a major breakthrough in AI technology, but also provides new possibilities for future applications of AI in healthcare, education, finance and other fields.
---

When building AI intelligences, interacting with external tools and data sources is a complex and time-consuming process that usually requires writing a lot of code. Now, a utility called LangGraph Agents + MCP has been created to effectively address this pain point.
The tool provides a Streamlit interface that allows users to dynamically add, remove and configure MCP tools, easily access various data sources, and interact with LangGraph ReAct intelligences in a visual interface. This greatly simplifies the development process and improves development efficiency.
Key features include:
- Supports real-time streaming display of smart body response and tool invocation process for developers to debug and optimize.
- Supports complete historical conversation tracking and management for easy backtracking and analysis.
- Provides a Docker one-click deployment method without complex configuration steps and supports multiple mainstream models such as Claude and GPT.
In addition, the tool comes with a detailed Jupyter tutorial to help users understand the principles of MCP and LangGraph integration.
In short, LangGraph Agents + MCP not only simplifies the development process of AI intelligences, but also provides rich functionality and convenient usage experience, making it an indispensable tool for AI developers.
---

Recently, the open source community launched a new program calledFastAPI-MCPa new library designed to simplify the connection between legacy FastAPI applications and modern AI agents. By using the Model Context Protocol (MCP), FastAPI-MCP makes it easier for developers to integrate existing FastAPI applications with the latest AI technologies.
The release of FastAPI-MCP provides developers with a powerful tool to quickly implement complex AI functionality without having to build complex interfaces from scratch. This innovation not only improves development efficiency but also reduces integration costs.
- **Simplified Integration**: FastAPI-MCP reduces compatibility issues between legacy applications and AI agents by standardizing the MCP protocol.
- **Improve development efficiency**: developers can utilize off-the-shelf libraries and tools to quickly implement AI functionality without having to delve into the underlying technical details.
- **Reduced maintenance costs**: Future updates and maintenance are made easier through harmonized protocols.
FastAPI-MCP has gained widespread attention and has been used in several real projects. For more details, please visit:https://t.co/wOJGFNcctb
---

Recently, a project to redraw "Qingming Shanghe Tu" using artificial intelligence technology has attracted widespread attention. By using the GPT-4o tool and combining it with different art style cue words, the researchers succeeded in presenting this famous painting from the Song Dynasty to the audience in a new perspective.

The program explores a variety of visual styles, including3D Q style,3D Style, 3D miniature style, realistic style and Disney style, Pixar style and Studio Ghibli style in animation.

- 3D Q style: Make the characters and scenes more cartoonish and add interest.
- 3D Style: Gives the image a stronger sense of three-dimensionality and spatial depth.
- 3D miniature style: makes the whole picture look like an exquisite handicraft, with every detail coming to life.
- Realistic style: retain the artistic characteristics of the original work as much as possible, while adding modern painting techniques.
- Disney/Pixar/Ghibli animation styles: classic animation aesthetics from different cultures are incorporated into ancient Chinese paintings.
This innovative attempt not only demonstrates the boundaries of AI's capabilities in the field of image generation, but also provides new ways of interpreting traditional artworks.
---

Just released! A few minutes ago.@seldo Provides a complete code-level tutorial on how to build intelligent agents in TypeScript using LlamaIndex.
Learn in this detailed tutorial:
- **What is LlamaIndex**: a library for building Retrieval Augmented Generation (RAG) systems.
- **What is an Intelligent Agent**: Intelligent agents are automated systems capable of performing complex tasks.
- **What is RAG**: Retrieval-Augmented Generation (RAG) is an approach that combines retrieval and generation techniques to improve model performance.
- Common intelligent agent patterns: including chained calls, routing, parallelization, etc.
- Build the intelligent agent RAG in LlamaIndex.TS.
- Build a full-stack React app with built-in workflows.
The tutorial not only covers the basic concepts but also provides practical code examples and detailed steps.
Click on the following link to watch the live video and learn more: https://t.co/lAHowy41mz

