Articles

The author showcases a peculiar scenario where AI (Claude Code) controls another AI (AutoGLM), which in turn operates a mobile phone. Initially, Claude Code was employed to deploy AutoGLM, and remarkably, Claude Code independently resolved all deployment challenges. Subsequently, the author instructed Claude Code to summarize these deployment issues, utilizing the deployed AutoGLM to dispatch these summaries to the developers. This synergy leverages Claude Code's 'thinking' and 'scheduling' prowess alongside AutoGLM's 'visual perception' and 'action' capabilities, forging a formidable mobile Agent. The article specifically highlights that this solution differs from ByteDance's 'Doubao Phone,' which is a system-level integration. In contrast, the AutoGLM-Claude Code combination represents a maverick solution, operating via debugging interfaces.

The article details entrepreneur Simon Klinkert's decision to deploy Google Gemini 3.0 as an “AI CEO” after his FinTech website, etf.capital, experienced declining traffic and revenue exhaustion. He developed a Go program that integrated Ghost CMS, Google Search Console, Sistrix, and Perplexity API, granting the AI access to the website's code and setting a daily goal of improving the website by 1%. Within three weeks, the AI CEO executed a series of “calm and ruthless” turnaround strategies, including deleting 50% of low-value content, implementing Noindex cleanup (preventing search engines from indexing low-value pages), enhancing E-E-A-T (Expertise, Experience, Authoritativeness, and Trustworthiness), and redesigning the user experience, with a strategic focus on “money-making pages.” These measures effectively halted the website's decline and demonstrated AI's immense potential in automated operations and decision-making. The author plans to further empower the AI to manage content distribution and marketing functions, ultimately aiming for fully automated operations for his startup.

The article points out that China's embodied AI and robotics technology is developing rapidly, particularly in robot mass production and practical applications, where it has emerged as a leader. The GDPS 2025 International Embodied AI Skills Competition showcased Chinese robots' real-world capabilities in emergency rescue, industrial operations, and other scenarios, drawing widespread attention and provoking strong reactions among American netizens. They believe that the US is overly fixated on meme culture and theoretical research in the embodied AI field, while China has already entered a stage of practical deployment. The article argues that China's leadership in embodied AI stems from the technical refinement and industrial implementation brought about by large-scale mass production, as well as the collaborative support from policies, real-world scenarios, and ecosystems in regions like Shanghai. The GDPS competition not only serves as a window for technology demonstration but also fosters a positive cycle between technology and business, showcasing the transformation of Chinese embodied AI from 'exhibits' to 'mass-produced products.'

At the InfoQ Dev Summit Munich 2025, Katharine Jarmul dismantled five prevalent myths in AI security and privacy, asserting that current AI safety approaches excessively rely on technical solutions while neglecting fundamental risks. The debunked myths include the belief that guardrails ensure safety, better model performance resolves security issues, risk taxonomies are sufficient, one-time red teaming is adequate, and future model versions will automatically fix present problems. Jarmul highlighted that AI automation has now surpassed augmentation, emphasizing the urgent need for addressing security and privacy. She advocated for interdisciplinary collaboration and continuous testing to develop a "risk radar muscle" to combat evolving threats and user data privacy concerns, while also encouraging users to diversify model providers and consider local models for enhanced privacy control.

This article provides an in-depth summary and analysis of recent trending topics on Hacker News. It covers multiple technological frontiers. First, it explores a fictional future Hacker News homepage created by Gemini Pro 3, showcasing the potential and limitations of AI content generation, and satirically imagining the business strategies of tech giants and future technology. Next, the article criticizes the design trend of icon overload in software interfaces, emphasizing the balance between visual information conveyance efficiency and cognitive load. Subsequently, through analogies of horses, chess, and AI, Andy Jones provides an in-depth analysis of the non-linear characteristics of AI technological advancement and its disruptive impact on human work and societal structures. Mistral AI released the open-source code model family Devstral 2, underscoring the importance of local deployment and enterprise privatization. The Jepsen report revealed serious flaws in NATS JetStream 2.12.1 regarding data persistence and consistency. In the discussion about whether AI tools can significantly reduce software building costs, the community generally believes that actual productivity improvement is limited and presents quality risks. Pebble Index 01, a miniature voice ring emphasizing privacy and ultra-long battery life, showcases innovation in edge computing for smart hardware. Bruno Simon's interactive 3D personal website, while cool, suffers from slow loading and high performance requirements, sparking discussions on the value of user experience. A paper proposed the "Universal Weight Subspace" hypothesis, exploring the potential for high compression and reuse of fine-tuning model parameters. Finally, the article deeply analyzes the complex cooperation and competition among NVIDIA, OpenAI, and Oracle, especially OpenAI's efforts to reduce reliance on NVIDIA and the future competitive landscape of AI hardware.

This article, based on an interview with AI pioneer Terrence Sejnowski, traces the development of neural networks from the Boltzmann Machine to modern Deep Learning, and contrasts the human brain with current Large Language Models (LLMs). Sejnowski believes that while LLMs perform powerfully in cognitive aspects, they lack the human brain's basal ganglia (responsible for Reinforcement Learning), hippocampus (responsible for long-term memory and information integration), and autonomous internal thought generation (for example, being unable to 'think' or 'sleep' without input, unlike humans). He emphasizes that Transformer is not the sole path to AGI and calls for AI research to not only replicate the human brain but also to focus on the diversity principles of intelligence in nature, deeply understanding the operational mechanisms of LLMs from physics and neuroscience perspectives, in order to truly achieve Artificial General Intelligence in the future.

Existing diffusion model fine-tuning methods predominantly employ static strategies, resulting in suboptimal performance when handling intricate structures and fine textures. Addressing this limitation, the NUS LV Lab team has introduced the FeRA (Frequency-Energy Constrained Routing) framework. This framework is built upon the fundamental discovery of a 'low-frequency to high-frequency' evolution rule during the diffusion denoising process. Leveraging this insight, FeRA designs a dynamic routing mechanism that utilizes a Frequency-Domain Energy Indicator (FEI) to perceive the real-time frequency-domain energy distribution of latent space features. Subsequently, a soft frequency router dynamically schedules different LoRA expert modules, while Frequency-Domain Energy Consistency Regularization (FECL) is introduced to ensure training stability. Experimental results demonstrate that FeRA achieves superior generation quality and text controllability in style transfer and subject customization tasks across models like SD 1.5, SDXL, and Flux.1. This work offers a novel perspective for the future fine-tuning of diffusion models in complex applications such as video and 3D generation.

Vivek Gupta, an AI team manager, shares insights on effectively managing machine learning engineers, focusing on their growth in both technical and interpersonal skills. His perspective is rooted in the practical, product-oriented environment of AI development, where enabling engineers directly impacts product success. He highlights the importance of providing comprehensive feedback, allocating dedicated learning time, fostering an environment where engineers feel comfortable asking for help, and encouraging cross-team collaboration to leverage shared knowledge and reduce duplicated efforts. Key aspects also include the manager's need to stay broadly informed across AI subfields, the value of mentorship (especially for senior engineers coaching juniors), and the critical role of data management and human-in-the-loop validation in production ML systems. The article concludes with an interview, offering practical examples like hackathons and learning days, and applies MLOps principles to the management of large language models, demonstrating how these practices are crucial for the robust development of AI products.

As large model parameters and context lengths increase, the GPU memory consumption of activation values has become a training bottleneck. This article details the "Fine-grained Activation Offloading" scheme implemented by Xiaohongshu's technical team within the Megatron-Core framework. This scheme effectively balances memory footprint and training efficiency by offloading activation values at the module level rather than by layer, performing offloading/reloading operations in parallel with computation, and being fully compatible with existing optimizations like Pipeline Parallelism (PP/VPP) and fine-grained recomputation. Experimental results show that this scheme achieves significant memory savings (10%-35%) and throughput performance improvements (7%-10%) on large models such as DeepSeek-V3, providing an efficient memory solution for ultra-large scale models and long sequence training, along with detailed best practice guidelines and future optimization directions. The article demonstrates excellent technical depth, but its universality and reference value could be further enhanced by providing more fundamental background knowledge and by evaluating performance under different hardware configurations.

This article is the first in the series "First Principles of Large Language Models," primarily focusing on the statistical physics perspective, belonging to the sub-field of theoretical AI research. Drawing from his team's arXiv research, Dr. Bai Bo aims to explain the mathematical principles behind large language models in an accessible manner. The article first reviews the deep historical connections between neural networks and statistical physics, especially emphasizing concepts like Hopfield Networks, Boltzmann Machines, and Gardner Capacity. Subsequently, it elaborates on how the Attention Module and Transformer Architecture can be described using statistical physics' Energy-Based Models (EBMs), revealing that the inference process involves finding the minimum of the energy function, and the training process approximates the parameter configuration that minimizes the average energy function. The article also explores the exponential growth relationship between LLMs' memory capacity and the number of parameters, the relationship between the upper bound of generalization error and the sum of absolute Logits, and ultimately suggests that the limit of LLMs' capabilities is Granger causality inference in the time series dimension, pointing out their lack of true symbolization, conceptual abstraction, and logical reasoning capabilities.