After a weekend hackathon, your home is a mess. However, when you come home on Monday night, you find a spotless living room and a candlelit dinner on the dining room table.It's impossible to tell if it's human or machine.. This may seem simple, but it is extremely difficult.
It's a story aboutPhysical Turing Testvision, it is seen as the next big thing in AI. This dream has inspired researchers to work through the night in their labs, working on new computing platforms that can automate the processing of matter rather than limiting it to information.
Sergey CorporationRecently hosted an event called AI Ascent where I shared how to address fundamental principles in general purpose robotics through data strategies and the law of scale. This talk will help you dive deeper into the field and I'm sure you'll find it well worth the 17 minutes.
The robots of the future will not only be able to perform complex tasks in home environments, but will also be able to play important roles in a variety of fields, including industry and medicine. Realizing this vision will require interdisciplinary collaboration, including cutting-edge technologies in areas such as computer vision, natural language processing, and mechanical engineering.
---
Remember Reinforcement Fine-Tuning (RFT)? We've been working hard on this technology since last December and officially launched the OpenAI o4-mini version today.RFT Utilizes Chained-Thinking Reasoning and Task-Specific Scoring Mechanisms to Improve Model Performance, especially for complex domains. For example, @AccordanceAI used RFT to fine-tune their tax and accounting models to the current optimum (SOTA).
Additionally, in terms of supervised fine-tuning, you can now fine-tune the GPT-4.1 nano. By training for specific use cases, you can get even more value out of our fastest and most cost-effective model.
Today, we previewed Enhanced Fine-Tuning, a new approach to model customization that enables organizations to build expert models for specific complex tasks such as coding, scientific research, or finance.
---
I am pleased to announce that AI Fund has successfully completed an oversubscribed round, raising $190 million for our new fund. We look forward to working with more entrepreneurs to create new companies that benefit humanity.
AI Fund is different from traditional venture capital firms in that we not only provide financial support, but also participate directly in the creation and operation of companies. Our team is actively involved in code writing, customer feedback collection, product design iteration, market analysis, etc., and is committed to driving the rapid development of multiple AI products. We continue to summarize best practices for emerging AI startups through practice.
Among the many factors that contribute to a startup's success, speed is crucial. Here are some important strategies to accelerate the growth of AI startups:
1. Focus on a single specific idea: A small team focused on a specific idea can move a project forward quickly. Be able to redirect quickly if initial assumptions prove problematic, rather than making multiple attempts at once.
2. Trusting the intuition of experts in the field: While data and user research are important, in some cases the intuition of a domain expert can help make quick decisions that can be implemented and validated quickly.
3. Utilization of AI-assisted coding: AI-assisted coding dramatically speeds up prototyping, especially in stand-alone prototyping projects that don't require high reliability and security. This allows us to prototype and test at an extremely fast pace.
4. Accelerating access to user feedback: As the speed of prototyping increases, getting user feedback becomes a new bottleneck. Teams that have the ability to quickly find potential customers and get feedback in a short period of time will have an advantage.
In addition to speed, deep technical understanding is key to success. Rapidly evolving AI technology requires teams to have a deep technical background in order to make better decisions and avoid detours.
Thanks to all of our investors, team members, and startup partners for their support, we still have a lot of work to do.
---
OpenAI recently announced that ChatGPT now supports connecting GitHub repositories to deep research capabilities.Users can ask questions about specific codebases, the in-depth research agent reads and searches the repository's source code and Pull Requests (PRs), generating a detailed report with citations.
This new feature is particularly suitable for software development teams and open source project maintainers to help them understand and work with complex code more efficiently. In this way, developers can not only quickly access code-related information, but also identify potential problems and improvement points. In addition, for researchers, this feature also helps to deeply analyze the internal structure and historical evolution of open source projects.
To use this feature, select the "Deep Research" option in ChatGPT and click the "GitHub" button to get started.
---
NVIDIA has just announced that it has open sourced three different sizes of code inference models, 32 billion parameter, 14 billion parameter and 7 billion parameter versions, and all of these models are under the Apache 2.0 license.
According to official test results, these models outperform other comparable models such as O3 mini and O1 (low) in LiveCodeBench benchmarks. In addition, based on the OCR dataset support, these models outperform other equivalent inference models by about 30% in token efficiency.
Key features include:
- Support for multiple frameworks such as llama.cpp, vLLM, transformers and TGI.
- Suitable for a variety of application scenarios such as code generation, debugging and optimization.
- Suitable for academic research and commercial development.
Developers can start using these models immediately to explore the potential for their application in real projects.
The open source not only provides powerful tools for researchers, but also opens up new possibilities for developers in general, promising to advance the field of code generation and automation.
---
SDV is an open source Python library that can generate large-scale synthetic data by learning and replicating patterns from real data using machine learning algorithms. This is useful for scenarios that require large amounts of data for training or testing.
The process of generating synthetic data using SDV is divided into three steps:
- train: Point SDV to your real data table and it will capture the underlying data distribution and relationships.
- generating: Run the trained SDV model to generate the required number of similar data rows without exposing any real data.
- validate (a theory): Utilize SDV's quality reports to check how well the generated data matches the real data; if needed, adjust the parameters for more accurate results.
Key features of SDV include:
- Multiple modeling options, from Gaussian conjugation to CTGAN.
- Supports single, multiple and sequential tables.
- Built-in anonymization and logical constraints.
- Everything can be done with just one command:sdv.sample()The
In addition, SDV can effectively address category imbalance and improve data quality.
SDV is applicable to a number of fields, such as finance, healthcare, and market research, and can help users test and train models without revealing sensitive information.
---
We celebrate the one-year anniversary of the release of AlphaFold 3. Over the past year, AlphaFold 3 has been widely used to predict millions of molecular interactions, dramatically accelerating the process of scientific research.
Scientists have used these predictions to make breakthroughs in a number of areas, including crop immunity and bacterial resistance.AlphaFold 3's ability to make accurate predictions, providing researchers with unprecedented tools to gain a deeper understanding of protein structure and function.
Isomorphic Labsis also utilizing AlphaFold 3 in conjunction with its advanced drug design engine to work on transforming the way new drugs are developed. With this innovative approach, they hope to develop new treatments faster and more effectively.
AlphaFold 3 not only plays an important role in academic research, but also demonstrates great potential in the pharmaceutical industry, promising to revolutionize the drug discovery and development landscape in the coming years.
---
A tutorial by @LangChainAI showing how to build a multimodal Retrieval Augmented Generation (RAG) system using Gemma 3 and LangChain for processing mixed-content PDF documents containing text and images. The system combines the powerful natural language processing capabilities of Google's Gemma 3 model with the flexibility of LangChain to efficiently parse and understand complex documents.
The whole implementation also integrates tools such as Streamlit and Ollama to provide a user-friendly interface and powerful computational support. Through this tutorial, developers can learn how to build a RAG system that can process both text and images, thus providing strong support for practical application scenarios such as document analysis and knowledge extraction.
For professionals working in data science, artificial intelligence, or related fields, mastering this technology not only improves productivity, but also opens up new avenues for solving more complex business problems.
---
Meta recently introduced an advanced model called Meta Locate 3D, which is capable of high-precision object localization in three-dimensional environments. This technological breakthrough is expected to significantly improve the ability of robots to navigate in complex scenes and interact naturally with humans.
With Meta Locate 3D, robots can more accurately understand their surroundings, including recognizing and locating various objects. The technology has a wide range of applications, not only for industrial automation and logistics and distribution, but also in home service robots, enabling them to better understand and respond to user commands.
- **Download models and data sets**
- **Read research paper**
- **Attempted online presentation**
More detailed information can be found athttps://t.co/GgJEPXTH8WThe
---
Learn how to utilizeVoyageAIThe multimodal embedding andMongoDBThe multimodal index for efficient multimodal retrieval.
In this tutorial, you will master the following:
- How to useVoyageAIThe multimodal embedding technique. This technique is capable of converting multiple types of data, such as text and images, into a unified vector representation for subsequent processing.
- How to set upMongoDB Atlasas a vector repository for image embedding. By storing image data in MongoDB, efficient data management and querying can be achieved.
- Create a multimodal index that supports both image and text data types. This allows the system to process different types of inputs simultaneously and provide accurate retrieval results.
- Querying images using text prompts (text-to-image retrieval). This technique allows users to search for relevant images through natural language descriptions, greatly improving the user experience.
These techniques are used in a wide range of application scenarios, including product recommendation in e-commerce, content management in social media, and medical image analysis, among other areas. With this tutorial, you can gain insight into the practical applications and technical details of multimodal retrieval.
