11/14/2025 AI Express | Biggest AI Hackathon Ever Launched, Several New Technologies Unveiled

This is probably the largest and most influential AI hackathon to date. The hackathon is co-organized by Anthropic and Gradio and supported by 10 sponsors.

The event attracted more than6,300participants, will be held from November 14 to 30 for two weeks. All participants can enter online and join the tech extravaganza wherever they are.

The contest not only provides$20,000 in cash prizesThe value of the program is more than $10,000.Sponsor Credit for $3.5M.. Participants will have the opportunity to present their innovative projects and network with leading experts in the industry.

• Activity period: November 14-30
• Number of participants: over 6,300
• Prize: $20,000 cash + $3.5M + sponsor credit
• Participation: Open online

Today at 10am (PT), the organizers will hold a kickoff event where a number of guests from Hugging Face, OpenAI, Google DeepMind, Modal, Blaxel AI, SambaNova AI, and Nebius TF will share their insights and experiences. This is not only a competition, but also a great opportunity to learn and network.

Those interested can watch the live stream at the link below:https://t.co/Mu3KwZuTTPThe

Google has just launched Deep Research, a new feature in NotebookLM.

This feature is capable of browsing hundreds of websites and generating well-organized reports based on user queries. In addition, it provides an annotated list of sources for more in-depth research. Users can add this information directly to their notebooks.

Deep ResearchA variety of new file types are supported, including Google Sheets, Microsoft Word documents, PDFs, and images accessible via Google Drive URLs. This feature not only simplifies the research process, but also makes information management and sharing more efficient.

For professionals who need to work with large amounts of data and informationNotebookLMDeep Research is undoubtedly a powerful tool. Whether it's for academic research or business analysis, it helps users quickly access and organize relevant information to improve work efficiency.

The Claude API has recently introduced structured output functionality, which is an important technical update.

Developers can now ensure that Claude's responses always conform to the specified JSON schema or custom tool definition, avoiding retries and parsing errors. This feature is particularly suited for application scenarios that require a high degree of consistency and accuracy, such as financial analytics, healthcare data processing, and large-scale enterprise systems integration.

Key benefits include:

• Improved consistency and accuracy of data processing
• Reduced development and maintenance costs
• Simplify processing of complex data streams

In addition, this update provides developers with more flexibility and control, enabling them to better manage and optimize API responses.

The Claude team says this feature will significantly enhance the user experience and provide robust support for a wider range of industries.

NotebookLM now supports an image input feature that allows users to take images of notes, textbook pages or diagrams and turn them into summaries, insights and answers. The introduction of this feature marks the further integration of image processing technology in the field of natural language processing.

Whether it's a photo of handwritten notes, a textbook screenshot or a diagram on a web page, theNotebookLMBoth can extract key information from it and generate corresponding text output. This technology not only improves learning efficiency, but also provides great convenience for researchers and students.

• Supports multiple image sources: including handwritten notes, textbook screenshots, web charts, etc.
• Automatically extract key information and generate summaries
• Provide insights and detailed answers

By combining computer vision and natural language processing technologies, NotebookLM is able to understand and parse image content more accurately, thus providing users with richer and more comprehensive information.

This innovative feature will greatly enhance the user's learning experience, making knowledge acquisition more convenient and efficient.

Recently, ChatGPT announced the launch of a group chat feature where users can now collaborate with others and ChatGPT in the same conversation. This new feature is currently being piloted in Japan, New Zealand, South Korea and Taiwan.

With this innovative feature, users can not only communicate with friends, family or colleagues in real time, but also invoke ChatGPT's powerful language processing capabilities directly in the chat window. Whether discussing projects, sharing ideas or seeking advice, this multi-participant interaction will greatly enhance communication efficiency and user experience.

Key features include:

• Support multi-user online communication at the same time
• Integration of ChatGPT Intelligent Assistant for instant feedback and information assistance
• Suitable for a variety of scenarios, such as team meetings, family gatherings, etc.

This update marks another major breakthrough in ChatGPT's interaction model, providing users with a richer and more flexible experience.

At the AI Dev 25 x NYC conference, renowned AI expert Andrew Ng delved into the reasons why artificial intelligence continues to accelerate.He noted that with the significant increase in coding speedThe development team is able to prototype faster than ever before. This rapid iteration makes technology innovation and product optimization more efficient.

However, Ng emphasized thatThe real bottleneck at the moment is collecting user feedback. Although technological advances have made the development process much faster, the challenge remains to obtain and understand the actual needs of users in a timely manner.

Ng encouraged attendees to actively network, collaborate, and work together to build future AI applications. He mentioned that it was a conversation with Kirsty Tan at the last AI Dev 25 conference that inspired the AI Aspire project. Such interactions and collaborations are crucial to advancing AI.

Overall, Ng's insights shed light on the core issues and directions of the current AI field, providing valuable guidance for developers and researchers.

The Google AI Developer Team shares how to use the Code Execution tool in the Gemini API to run Python code in the AI Studio environment to solve complex math problems, perform efficient data analysis, and enable code-based reasoning.

@nikitanamjoshi details how, with this tool, users can easily write and execute Python code in AI Studio to quickly handle a variety of computational tasks. For example, in terms of solving mathematical problems, users can utilize Python's powerful libraries (such as NumPy and SciPy) to perform complex numerical calculations and symbolic operations.

Specific application scenarios include:

• Risk assessment and quantitative analysis in the financial sector
• Data modeling and simulation in scientific research
• Teaching and experimenting with online programming for the education industry

In addition, Code Execution tools support data visualization to help users understand data and results more intuitively.

With these features, developers and researchers are able to accomplish their work tasks more efficiently and increase productivity.

Meta recently made a major breakthrough in the area of retrieval-enhanced generation (RAG), solving the biggest problem that has long plagued the technology. Traditional RAG systems typically retrieve large chunks of documents, even if the user only needs a small portion of them. This practice not only wastes computational resources, but also results in users paying unnecessary fees.

To address this problem, Meta introduced REFRAG, a new approach to RAG.REFRAG significantly improves the performance of the system by compressing and filtering the context at the embedding layer:

• Time to first token speeds up30.85 times
• Context window expanded16 times
• The number of tokens processed is reduced2-4 times
• Outperforms LLaMA model in 16 RAG benchmarks

REFRAG works as follows:

1. Encoding and storing documents in a vector database
2. When a query arrives, retrieve the relevant document block in the usual way
3. Evaluating compressed embeddings and selecting the best document block using a strategy trained in reinforcement learning
4. The selected document block is expanded into a full token embedding
5. Unchecked document blocks remain compressed or are filtered out completely.
6. All content is sent together to the Big Language Model for processing

This approach enables REFRAG to process more context at a higher speed with zero accuracy loss. This is important for application scenarios that require efficient processing of large-scale text data, such as enterprise-level knowledge management and intelligent customer service.

We are very excited to share the newly released SophontAI results - theOpenMidnight, which is our first pathology-based model.

OpenMidnight has been able to achieve the same results after using only 12,000 full-slice images and1.6 thousand dollars in computing resourcesIn the case of the case, it reached theLatest performance levels. This efficient training method opens up new possibilities for research and applications in the field of pathology.

Pathology-based models play an important role in medical image analysis and can help doctors improve diagnostic accuracy and efficiency. openMidnight's success not only demonstrates the feasibility of training high-performance models at low cost, but may also drive the development of more similar technologies to further enhance the intelligence of the medical field.

We expect OpenMidnight to be useful in real clinical applications and to provide strong support for future medical research.

Stefano Pasquali, Head of Financial Services at DomynAI, spoke at the AI Dev 25 Summit in New York, emphasizing DomynAI's commitment to building a transparent, auditable and autonomous AI ecosystem. He stated.DomynAI believes in combining innovation with responsibilityThis will lead to a more secure and reliable solution for the financial industry.

Currently, AI technology is increasingly used in the financial sector, from risk management to intelligent investment advice to anti-fraud detection, all of which require a high degree of trust and transparency. domynAI meets the stringent requirements of both regulators and customers for data privacy and security through its advanced technology framework, which ensures that AI systems are not only efficient, but also traceable and controllable in their operation.

Additionally, Pasquali mentioned that DomynAI is driving the widespread adoption of AI technology in financial services, aiming to enhance the user experience through technological innovation while ensuring system reliability and security.