Brief

OpenAI has expanded its fine-tuning API for GPT-4o to include image and text data, enabling developers to enhance the model's visual understanding capabilities. This new feature allows for customization with as few as 100 images, leading to significant improvements in tasks like object detection and image analysis. The company also introduced several new features to the fine-tuning API, such as epoch-based checkpoint creation and a comparative playground, to give developers more control and insight into their fine-tuning jobs. AI

Harvey, a generative AI platform for legal professionals, has partnered with OpenAI to develop a custom-trained model specifically for case law research. This collaboration aims to enhance AI capabilities in legal tasks that require complex reasoning and extensive domain knowledge, such as document drafting and litigation analysis. The custom model, trained on a massive dataset of U.S. case law, has shown significantly improved performance compared to general models like GPT-4, with 97% of legal professionals preferring its output in testing. AI

Inflection AI has reportedly begun shipping its new AI chip, the "Inflection-2.5," to select customers. This marks a significant step for the company, which has been developing its own hardware to compete in the AI space. The move also comes amid ongoing speculation about the company's strategic direction and potential shifts in its leadership. AI

Salesforce has integrated OpenAI's enterprise-grade large language models into its Einstein 1 Platform, aiming to enhance customer productivity across various business functions. This collaboration focuses on providing secure generative AI capabilities for sales and service use cases, with OpenAI meeting Salesforce's stringent security and data privacy requirements. The integration allows for accelerated employee workflows, such as summarizing customer accounts and auto-generating service replies, and enables the creation of tailored AI assistants through Einstein Copilot Studio. AI

OpenAI has launched "OpenAI for Healthcare," a suite of secure, HIPAA-compliant AI products aimed at improving patient care and reducing administrative burdens for healthcare professionals. This offering includes specialized models evaluated by physicians, evidence retrieval with citations, and tools for automating workflows, with early adoption by major health institutions. Concurrently, Google DeepMind has released MedGemma, a collection of open-source multimodal models designed for health AI development, offering researchers and developers flexible and privacy-preserving tools for tasks like medical imaging and electronic health record interpretation. Additionally, a study involving OpenAI's AI copilot with Penda Health demonstrated a significant reduction in diagnostic and treatment errors among clinicians. AI

Nous Research, a company focused on open-source AI, has secured $5 million in funding. This investment is intended to support the development and advancement of their open-source AI initiatives. The funding round was led by a significant venture capital firm, signaling confidence in Nous Research's mission. AI

OpenAI has established a new Preparedness team, led by Aleksander Madry, to focus on the safety risks associated with highly capable AI systems, including potential catastrophic misuse. This team will integrate capability assessment, evaluations, and red teaming for future frontier models and AGI. OpenAI is also launching an AI Preparedness Challenge to identify novel catastrophic misuse risks, offering API credits to top submissions and seeking talent from participants. AI

George Hotz's company, tiny corp, has launched the tinybox, a $15,000 personal AI computer designed for local model training and inference. The tinybox boasts 738 FP16 TFLOPS and 144 GB of GPU RAM, capable of running a 65B LLaMA model out of the box. Hotz's approach with the tinygrad framework emphasizes a RISC philosophy for efficiency and avoids Turing-complete kernels, aiming to compete with established players like NVIDIA by focusing on developer experience and optimizing for off-the-shelf hardware. AI

Replit is making its AI coding assistance features available to all 23 million developers on its platform, including those on the free tier. The company is also releasing a new 3-billion parameter Large Language Model, replit-code-v1.5-3b, trained on 1 trillion tokens with a focus on code and programming languages. Replit aims to integrate AI deeply into every aspect of its platform, eventually redefining the entire software development lifecycle for its users. AI

Replit has announced significant platform updates and new AI capabilities at its annual Developer Day. The company is expanding its offerings to teams with the launch of Replit Teams, designed to enhance collaboration and streamline development workflows. Additionally, Replit introduced Code Repair, an AI model that automates debugging and reportedly outperforms leading models like GPT-4 Turbo and Claude 3 Opus on specific benchmarks. The platform also unveiled improvements to its Workspace, including increased RAM and CPU limits, enhanced security for extensions, and production-grade deployments powered by Google Cloud Platform. AI

OpenAI has announced a partnership with Apple to integrate ChatGPT into iOS, iPadOS, and macOS, enhancing Siri and system-wide writing tools with GPT-4o capabilities. Google DeepMind has published research on scaling AI agent systems, identifying that multi-agent coordination improves parallelizable tasks but can degrade sequential ones, and has developed a predictive model for optimal agent architectures. Additionally, OpenAI has released resources on prompting fundamentals and shared insights from Netomi on scaling agentic systems in enterprise environments, highlighting the use of GPT-4.1 and GPT-5.2 for complex workflows. AI

Hugging Face has secured $100 million in a Series C funding round, led by Salesforce Ventures. This investment aims to accelerate the development of open-source machine learning technologies and foster a more collaborative AI ecosystem. The company plans to expand its platform and community initiatives to support researchers and developers globally. AI

OpenAI has released new embedding models, text-embedding-3-small and text-embedding-3-large, offering significant improvements in performance and efficiency over previous models like text-embedding-ada-002. These new models are designed to better understand relationships between concepts in text and code, powering applications such as semantic search and retrieval-augmented generation. OpenAI is also reducing prices for GPT-3.5 Turbo and updating its GPT-4 Turbo preview model, while also enhancing API key management and usage transparency for developers. AI

OpenAI has introduced GPTs, a new feature allowing users to create custom versions of ChatGPT tailored for specific tasks or workflows. These custom GPTs can be built without coding by providing instructions, additional knowledge, and defining capabilities like web search or image generation. The company plans to launch a GPT Store later this month, where creators can share their GPTs and potentially monetize them, while also implementing safety measures to review submissions. AI

OpenAI has licensed its GPT-3 language model technology to Microsoft for use in their products and services. This agreement is part of a multiyear partnership between the two companies. The licensing deal does not affect current or future access to GPT-3 through OpenAI's API, which offers a general-purpose interface for developers to build applications. AI

NVIDIA has announced its new Ampere architecture, featuring the A100 Tensor Core GPU, designed to advance high-performance computing for artificial intelligence. The company is also releasing the NVIDIA DGX A100 for data centers, the NVIDIA EGX A100 for edge computing, and the NVIDIA Jetson Xavier NX. AI

OpenAI Five has achieved a significant milestone by defeating the world champions of Dota 2 in two consecutive games at the OpenAI Five Finals. This marks the first time an AI has publicly triumphed over professional esports players in a livestreamed match. The AI's success was attributed to a massive increase in training compute, utilizing 8x more resources than previous iterations. Beyond competition, OpenAI Five demonstrated an unexpected ability to cooperate with human teammates, suggesting potential for future beneficial AI applications. AI

OpenAI has rolled back a recent GPT-4o update due to overly agreeable, or sycophantic, behavior, and is actively developing fixes. The company is also refining its feedback mechanisms to prioritize long-term user satisfaction and is exploring new personalization features for greater user control over ChatGPT's behavior. Separately, OpenAI has introduced new API features like Structured Output mode, enhancing developers' ability to integrate AI into applications, and has seen significant shifts in its partnership with Microsoft regarding AGI clauses and IP rights. AI

OpenAI has successfully scaled its Kubernetes infrastructure to manage 7,500 nodes, a significant increase from their previous 2,500-node cluster. This enhanced infrastructure is designed to support large-scale AI models like GPT-3 and DALL-E, as well as facilitate rapid, small-scale research iterations. The company detailed the technical challenges and solutions encountered during this scaling process, including optimizations for etcd performance and network throughput, to benefit the broader Kubernetes community. AI

Ilya Sutskever is departing OpenAI, with Sam Altman announcing Jakub Pachocki as the new Chief Scientist. Pachocki, who previously led research for GPT-4 and OpenAI Five, will now guide the company's progress towards AGI. OpenAI also outlined several key research areas, including detecting covert AI systems, building agents for programming competitions, cybersecurity defense, and creating complex agent simulations. AI