Keywords：OpenAI DevDay 2025, ChatGPT Application Platform, AgentKit, AI Agent Development, GPT-5 Pro, Sora 2, CodeMender, Continuous Thought Machine, ChatGPT Apps SDK, Agent Builder Visual Constructor, GPT-Realtime-Mini Voice Model, Gemini Deep Think Technology, CTM Neurodynamics
🔥 Spotlight
OpenAI DevDay 2025 Major Announcements: ChatGPT Becomes an Application Platform, AgentKit Empowers AI Agent Development : OpenAI unveiled several significant advancements at its 2025 annual developer conference. It announced that ChatGPT now boasts 800 million weekly active users, with its API processing over 6 billion tokens per minute. Key announcements included the Apps SDK, enabling developers to build and run full-featured applications within ChatGPT, transforming it into a new operating system. Concurrently, AgentKit was introduced, comprising Agent Builder (a visual builder), ChatKit (customizable chat UI), Guardrails (a safety module), and Evals (evaluation tools, significantly simplifying AI agent development. Additionally, GPT-5 Pro, Sora 2/2 Pro video generation API, and the GPT-Realtime-Mini voice model were simultaneously launched. The Codex programming tool is now generally available, with new SDKs and enterprise features added. These updates herald a deep integration and rapid expansion of the AI application ecosystem, profoundly impacting the developer ecosystem and user experience.
(Source: dotey, jerryjliu0, gdb, Yuchenj_UW, swyx, kevinweil, scaling01, scaling01, gdb, scaling01, scaling01, swyx, scaling01, gdb, gdb, op7418, TheRundownAI, OpenAIDevs, nickaturley, reach_vb, snsf, dotey, edwin)
Google DeepMind Launches CodeMender, an AI for Automatically Fixing Software Vulnerabilities : Google DeepMind has released CodeMender, an AI agent that leverages Gemini Deep Think technology to automatically patch critical software vulnerabilities. The agent has successfully submitted 72 high-quality fixes to popular codebases, which have been accepted and adopted by maintainers. The launch of CodeMender marks a significant breakthrough for AI in software security, promising to substantially reduce the time developers spend finding and fixing vulnerabilities, thereby enhancing software supply chain security.
(Source: Google, GoogleDeepMind)
Sakana AI’s ‘Continuous Thought Machines’ Accepted as a Spotlight Paper at NeurIPS 2025 : Sakana AI announced that its “Continuous Thought Machines” (CTM) have been accepted as a spotlight paper at NeurIPS 2025. CTM is an AI that mimics biological brains, using neurodynamics and synchronization mechanisms to ‘think’ over time. It can solve complex mazes by building internal maps, classify images by ‘gazing,’ and learn algorithms. This groundbreaking design demonstrates AI’s potential in mimicking biological intelligence and solving complex problems, foreshadowing future AI systems with stronger emergent capabilities.
(Source: hardmaru, hardmaru)
ARCS V3 Achieves Abstract Reasoning Breakthrough with Minimal Parameters : ARCS V3 achieved 90-98% accuracy on the ARC-AGI-2 benchmark with only 19.9M parameters, making it 88,442 times smaller than GPT-4 and not utilizing a Transformer architecture. This achievement challenges the industry’s reliance on large-scale models, demonstrating that exceptional performance in abstract reasoning tasks can be achieved with extremely low parameter counts through innovative architectural design and methods. The research team emphasizes that this breakthrough represents true reasoning capability rather than memorization, and has provided comprehensive validation logs and demo videos.
(Source: weights_biases)
Equilibrium Matching (EqM) Simplifies and Outperforms Flow Matching, Enhancing Generative Performance : Yilun Du and colleagues shared research on Equilibrium Matching (EqM), a method that simplifies and outperforms flow matching, achieving an FID score of 1.96 on ImageNet 256×256 and demonstrating powerful generative performance. EqM achieves a simple gradient-based generation process by learning a single static EBM (Energy-Based Model) landscape for generation. This advancement offers a more efficient and higher-performing alternative for generative models.
(Source: VictorKaiWang1)
🎯 Trends
OpenAI Partners with AMD to Deploy MI450 GPUs, Accelerating AI Infrastructure Development : OpenAI announced a multi-billion dollar partnership with AMD to deploy 6 gigawatts of AMD Instinct MI450 GPUs starting next year, addressing the growing demand for AI computing. This collaboration will significantly accelerate the development of global AI infrastructure, providing more computing resources for OpenAI users while also generating substantial revenue for AMD, creating a win-win situation for both parties.
(Source: dejavucoder, jachiam0)
Google AI Pro Plan Offers Free Upgrade to University Students : Google announced that university students can receive a free one-year upgrade to the Google AI Pro plan. The plan includes Gemini, NotebookLM, and 2TB of storage, designed to help students complete assignments, understand complex concepts, create study guides, and improve writing. This initiative is expected to promote the widespread adoption of AI tools in education, empowering students in their learning and research.
(Source: Google)
Microsoft Copilot Updates Memory Feature, Supports ‘Forget’ and ‘Remember’ Commands : Microsoft Copilot has updated its memory feature, now capable of ‘remembering’ or ‘forgetting’ specific information based on user commands. Users can manage Copilot’s memory in the settings, ensuring the AI responds more precisely to personalized needs while avoiding unnecessary information retention. This update enhances the AI assistant’s flexibility in terms of privacy protection and user experience.
(Source: mustafasuleyman)
LlamaParse Now Supports Anthropic Claude Sonnet 4.5, Enhancing Document Processing Capabilities : LlamaParse announced its integration with Anthropic’s Claude Sonnet 4.5 model, providing users with more powerful document understanding and parsing capabilities. This update will enhance LlamaParse’s accuracy and efficiency in processing complex documents, and it foreshadows the release of detailed benchmark results comparing Sonnet 4.5 with existing parsing options to showcase its performance advantages.
(Source: jerryjliu0)
HuggingFace Inference Endpoints Now Support Nvidia B200 GPUs : HuggingFace announced that its Inference Endpoints now support Nvidia B200 GPUs. This upgrade provides developers with more powerful computing capabilities to run and deploy large AI models, meeting the increasing demand for computation. This move will further drive the application and innovation of AI models, lowering the barrier for high-performance AI.
(Source: scaling01)