Articles

This article delves into OpenAI's current predicaments, tracing its journey from being hailed as AI's 'savior' to confronting substantial funding gaps and a crisis of trust. It highlights the diminishing impact of GPT-5 demonstrations and Google's forceful counterattack with its Gemini model, both technologically and across the industry chain. This creates an ironic parallel: OpenAI's current 'red alert' state mirrors Google's own challenges just two years prior. The article underscores OpenAI's projected financial deficit of up to $207 billion and the 'interdependent funding' dilemma exposed by its management's anxious responses to scrutiny. In contrast, Google's deep cash reserves, proprietary TPUs, and cloud infrastructure provide unparalleled advantages, suggesting the AI competition is now less about models and more about infrastructure and capital strength. Ultimately, the article posits that this emotional shift is a 'stress test' for AI expectations, indicating that AI technology will move from blind enthusiasm to an 'industrialization stage' focused on profitability and product quality, where those with robust infrastructure will hold a decisive edge.

This article provides an in-depth interpretation of Google DeepMind's latest research, "Evo-Memory," which introduces a new paradigm for continual learning without modifying model parameters. The article begins by identifying the limitations of traditional RAG systems' memory capabilities, specifically their static retrieval mechanism, which lacks self-reflection and filtering, making them easily overwhelmed by noise and unable to genuinely learn from experience. To address this, "Evo-Memory" introduces the core concept of "Context as Weights." Through the meta-reasoning (Refine), pruning, and reorganizing mechanisms within the ReMem framework, the model dynamically adjusts and optimizes context during inference, thereby endowing fixed-parameter LLMs with learning capabilities akin to weight adjustment. The article showcases the significant improvements in task success rate and problem-solving efficiency achieved by the ReMem framework through specific case studies, emphasizing the pivotal role of meta-reasoning. Ultimately, by integrating "Evo-Memory" with DeepSeek V3.2's planning capability, the article outlines a complete closed loop for next-generation AI's "System 2 Thinking," foreshadowing an AGI prototype where parameters remain constant but wisdom continually grows.

The article introduces DeepAgents CLI, a model-agnostic, Python-based coding agent designed for interactive terminal use, featuring shell execution, filesystem tools, and memory. It addresses the critical challenge of evaluating coding agents in isolated environments due to their interactive nature and potential for side effects. To solve this, the authors leverage Harbor, a framework enabling scalable, containerized agent evaluations with automatic test execution and reward scoring across various sandbox providers. They specifically developed `deepagents-harbor` to integrate DeepAgents with Harbor. The evaluation was conducted on Terminal Bench 2.0, a benchmark comprising 89 tasks across domains like software engineering, biology, and security, designed to measure agent proficiency in terminal environments. DeepAgents CLI, powered by Claude Sonnet 4.5, achieved an average score of 42.65%, establishing it as a competitive baseline. The article concludes by highlighting future work on systematic trace analysis for performance optimization.

This paper introduces the Self-Referenced Policy Optimization (SRPO) framework, proposed by teams from Fudan University, Tongji University, and Shanghai Institute for Intelligent Computing (SiiRL), aiming to address the high dependency on expert demonstration data and reward sparsity issues prevalent in VLA model reinforcement learning. SRPO incorporates an internally built self-reference evaluation mechanism that utilizes the model's own successfully generated trajectories to provide progressive rewards for failed attempts. This approach facilitates adaptive policy optimization without the need for external expert data or task-specific reward engineering. The framework achieved a State-of-the-Art (SOTA) success rate of 99.2% on the LIBERO benchmark and significantly boosted performance by 167% on the LIBERO-Plus generalization task. SRPO's core technical solutions encompass on-policy trajectory collection, trajectory clustering based on latent world representation, and the calculation of progressive rewards using trajectory distance, all integrated into a Proximal Policy Optimization (PPO)-style policy optimization objective. Experiments demonstrate that SRPO not only substantially improves training efficiency and generalization capabilities but also encourages robots to explore new strategies beyond pre-defined expert trajectories.

The article details the significant updates in NVIDIA CUDA Toolkit 13.1. Hailed as the largest in 20 years, this update introduces the core CUDA Tile programming model, which simplifies GPU programming through higher-level abstraction and better leverages dedicated hardware like Tensor Core. In addition, the runtime API opens up Green Context, allowing developers to perform more granular GPU resource partitioning and management. The article also covers enhancements to CUDA Multi-Process Service (MPS), performance optimizations for mathematical libraries such as cuBLAS, cuSPARSE, cuFFT, cuSOLVER, and CCCL, as well as new features for developer tools like Nsight Compute, Compute Sanitizer, and Nsight Systems. These updates aim to improve GPU programming efficiency, performance, and debugging capabilities, especially for AI workloads.

Under the leadership of Jared Palmer, GitHub is undergoing a significant strategic transformation in the AI era. This includes two key aspects: the launch of Agent HQ, integrating third-party AI Coding Agents into a unified 'Task Console' for developer and agent collaboration; and the complete restructuring of the GitHub homepage based on developer feedback, evolving it from a 'decorative page' into a practical hub for tasks, PRs, and activities. Palmer leverages his experience from Vercel and v0 (an AI-powered UI Generation Tool) to broaden GitHub's focus from specific languages or frameworks to building activity spaces for all, achieving broader platform-level AI integration. The article also explores Agent Memory Mechanisms, AI SDK development, the potential of Dev Containers, and future features like StackTips. It emphasizes that the maturity of agent technology hinges on the synchronous upgrade of Model Intelligence and Infrastructure Reliability, and highlights the limitations of current data analysis tools in the AI era. Overall, GitHub is proactively adapting to the AI era, aiming to create a new ecosystem for seamless collaboration between developers and AI Agents.

The article thoroughly analyzes a significant observation from the Y Combinator F25 Demo Day: 53% (83 out of 156) of the companies are AI-driven. This indicates that AI has transitioned from a single sector to an underlying capability and default premise for all industries. The article categorizes the 83 AI companies into nine major groups, covering aspects such as AI Agent infrastructure, vertical applications, developer tools, finance, marketing, content creation, healthcare, and underlying infrastructure. Through specific company examples, the article demonstrates the maturity of AI Agents at the infrastructure level, including memory layers, security authentication, and observability, as well as their deep application and innovative models in vertical domains such as IT support, legal compliance, medical billing, sales customer service, and coding. The article concludes by pointing out that future innovation in startups will increasingly focus on "industry depth," "value chain position," and "degree of automation," and emphasizes that Product-Led Growth (PLG) has become the mainstream market expansion strategy for AI companies.

This article discusses Andrej Karpathy's latest insights on Large Language Model (LLM) interaction strategies. Karpathy emphasizes that LLMs are essentially simulators, not autonomous, conscious entities. Therefore, when posing questions, one should avoid personified questions like 'What do you think?'. He suggests a more effective prompt strategy: ask, 'What kind of group of people would be suitable for exploring a certain topic? What would they say?', to guide the LLM in simulating multi-perspective answers. The article also explains the engineered nature of AI persona, clarifying that an LLM's 'personality' is deliberately constructed through techniques like SFT and RLHF, rather than emerging naturally. Finally, the article differentiates the impact of this new prompt strategy in factual and subjective domains. Through practical validation, by simulating the perspectives of Deep Learning Theorists, RLHF Engineers, Cognitive Scientists, and Advanced Prompt Engineers, the article further elaborates on and verifies the effectiveness of Karpathy's viewpoint, demonstrating that treating LLMs as simulators can significantly enhance their performance and answer quality. This piece is an interpretation, validation, and application of Karpathy's core ideas.

This article details NVIDIA CEO Jensen Huang's high-profile discussion in Washington, D.C., where he introduced a five-tier framework for understanding the AI industry: energy, chips, infrastructure, models, and applications. Huang analyzed China-U.S. competition across these layers, highlighting China's advantages in energy reserves, infrastructure development speed, and open-source models, while the U.S. excels in chip technology and advanced models. He stressed that energy is the primary bottleneck for AI development and urged the U.S. to pursue reindustrialization via industrial policies to secure AI leadership. Furthermore, Huang envisioned the future of Embodied AI (Physical AI), explaining that AI will transform tasks rather than eliminate jobs, offering an optimistic and realistic perspective on AI's societal impact.

The article details the collaboration between Memory Tensor and SenseTime Megascale Cluster, successfully implementing a commercial Prefill-Decode Separation inference cluster with integrated 'memory-compute-scheduling' on domestic GPGPUs. This solution deeply couples Memory Tensor's core product, MemOS's activation memory system, with Prefill-Decode Separation, reconstructing the large model inference paradigm. Real-world test results show that the cluster improves single-card concurrency efficiency by 20% and throughput by 75% under real Consumer-facing loads, with a comprehensive inference price-performance ratio 50% higher than the same generation NVIDIA A100.