Keywords：OpenAI, GPT-5.2 Pro, xAI, ChatGPT advertising test, Claude Code programming revolution, Colossus 2 supercomputer
🔥 Focus
OpenAI Begins Ad Testing and Discloses 2025 Revenue Surpassing $20 Billion: OpenAI officially announced it will test advertising in the free version of ChatGPT and the newly launched $8 “Go” version. CFO Sarah Friar disclosed that the company’s 2025 Annual Recurring Revenue (ARR) has soared to $20 billion, with a computing capacity reaching 1.9GW. This marks OpenAI’s significant transition from a pure subscription model to a “Compute-Research-Product-Commercialization” closed loop. Despite facing massive annual expenditures of approximately $170 billion, OpenAI is attempting to alleviate financial pressure without compromising answer neutrality by drawing on Instagram’s “intent-driven” native advertising model. This move is seen as a signal that the AI industry is entering a “commercial hard landing” phase and predicts that advertising could become its primary revenue source within the next three years. (Source: Sarah Friar, 36Kr, X/NewZhiYuan)
GPT-5.2 Pro Independently Proves 45-Year Number Theory Conjecture, Verified by Terence Tao: OpenAI’s latest model, GPT-5.2 Pro, successfully and independently proved Problem No. 281 in the Erdős conjecture collection. Fields Medalist Terence Tao verified the proof, calling it “the most definitive major contribution by AI to date” and expressed amazement that the AI made no errors in subtle areas such as limit exchanges and quantifier ordering. Although the community later discovered a simpler solution based on a 1936 classical theorem, the rigorous derivation provided by GPT-5.2 Pro via an ergodic theory path proves that frontier models have achieved substantial breakthroughs in handling high-difficulty abstract logic tasks. This event is regarded as a milestone for AI entering the deep waters of scientific research. (Source: Tao, 36Kr)
xAI Launches World’s First Gigawatt-Scale AI Supercluster Colossus 2: Elon Musk’s xAI has officially launched the Colossus 2 supercomputer, the world’s first AI cluster to reach the Gigawatt scale, equipped with approximately 555,000 NVIDIA GPUs with a total estimated value of $18 billion. The facility is powered by on-site gas turbines and Tesla Megapacks, consuming electricity equivalent to the peak demand of San Francisco. This move aims to crush competitors through extreme computing scale, with plans to upgrade to 1.5 GW by April. This infrastructure gamble demonstrates xAI’s aggressive stance on the Scaling Laws path but has also triggered strong criticism from environmental organizations regarding local air quality and energy consumption. (Source: Reddit, Twitter)
Claude Code Triggers Programming Paradigm Revolution: Building a 3-Million-Line Code Browser in One Week: Anthropic’s Claude Code command-line tool has caused a massive stir in the developer community. Vercel CTO Malte Ubl stated it completed a project in one week that would have originally taken a year; Michael Truell demonstrated a browser built with it containing 3 million lines of code, a self-developed rendering engine, and a JS virtual machine. This “Vibe Coding” mode significantly lowers the barrier to programming, even allowing 8-year-olds to complete agent development. it has spawned a “Cracked Engineers” culture where one person can do the work of a whole team, signaling a future shift in software development from “buying SaaS” to “building micro-apps,” while mediocre middle-layer programmers face an unprecedented survival crisis. (Source: Michael Truell, 36Kr, Twitter)
🎯 Trends
DeepSeek Releases Conditional Memory Paper, Hinting at the Imminent Arrival of V4: DeepSeek and Peking University jointly released the paper “Conditional Memory via Scalable Lookup,” proposing the concept of “Conditional Memory” to solve the long-term memory shortcomings of LLMs through scalable lookup. This is seen as a technical precursor to its next-generation model, V4, which is expected to be released around the Spring Festival. Additionally, the community has successfully reproduced its mHC architecture, confirming that it can achieve 10,000x signal amplification and stable training at 1.7B parameters. (Source: HuggingFace, karminski3)
Diverging Paths for Zhipu and MiniMax Post-Listing; Domestic Large Models Enter the “Final Round”: As Zhipu and MiniMax go public one after another, the strategies of China’s “Six Little Tigers” are diverging. Baichuan Intelligence is turning toward the medical vertical with the release of Baichuan-M3; Moonshot AI persists in chasing Anthropic, pursuing the upper limits of intelligence. Zhipu, in collaboration with Huawei, released GLM-Image, the first SOTA image generation model trained on domestic chips. The market is shifting from a parameter race to a competition over efficiency and implementation scenarios. (Source: 36Kr)
Sakana AI Launches RePo Mechanism to Optimize Context Processing: Sakana AI introduced the RePo (Context Relocation) mechanism, breaking the fixed linear order in which models process information. RePo can dynamically adjust positions based on content relevance, allowing the model to actively “pull in” key information and “push away” noise, significantly improving the model’s robustness and reasoning efficiency when processing noisy, long-text inputs. (Source: SakanaAILabs)
Google Releases TranslateGemma, a Free Translation Model Supporting Local Execution: Google launched the TranslateGemma series (4B/12B/27B), supporting 55 languages and capable of running on personal computers. The model supports text and image text extraction translation, focusing on privacy protection and zero API costs, providing a powerful tool for localized multi-language collaboration. (Source: QuixiAI)
🧰 Tools
DeepAgents Open Lovable: Natural Language to React Apps: An open-source platform developed by the LangChain community, built on DeepAgents and LangGraph. Users only need to input natural language to generate a complete React frontend application, supporting sub-agent collaboration, real-time preview, and one-click deployment. (Source: LangChain)
iMCP: Integrating iMessage into Claude Collaboration Workflows: Developer Mattt launched the MCP server, allowing users to bring iMessage conversations directly into Claude or programming agents. This bridges the gap between social communication and workflows, allowing AI to write code or execute tasks directly based on chat context. (Source: HamelHusain)
Headroom: LLM Context Compression Layer: This tool uses compression technology to reduce LLM costs by 50-90%. Deeply integrated with LangChain, it can save 60% of costs in RAG scenarios and reduce agent token consumption by 85%. Supporting Proxy/SDK deployment, it is a powerful tool for enterprise-level cost reduction and efficiency. (Source: LangChain)
Vibecraft: Claude Code Manager with Spatial Audio: An open-source project with 30,000 lines of code that provides a cool UI, script hooks, visual effects, and spatial audio for managing Claude code. It provides audio feedback on Claude’s status, making long-term AI collaboration more immersive. (Source: nearcyan)
Eduly: Paper-to-Short-Video Tool: Based on the LangChain Deepagents framework, it can automatically generate and debug Manim animation code, transforming dry academic papers into short video animations suitable for social media, achieving the automation of knowledge dissemination. (Source: LangChain)
📚 Learning
“Linear Algebra for Computer Vision, Robotics, and Machine Learning” Free Textbook: This book covers core theories such as vector spaces, matrices, norms, eigenvalues, and SVD, and delves into practical applications like PCA, graph theory, and 3D rotation. It is an authoritative resource for AI practitioners to strengthen their mathematical foundation. (Source: TheTuringPost)
Agent-as-a-Judge: A Survey of Next-Generation AI Evaluation: Explores why traditional LLM-as-a-Judge fails in complex tasks and proposes the “Agent-as-a-Judge” model, which introduces planning, tools, and memory, providing a roadmap for building robust, verifiable AI evaluation systems. (Source: TheTuringPost)
STEM: A New Method for Transformer Parameter Expansion by Meta and CMU: Expands Transformer parameter memory through embedding modules without routing and without increasing runtime computational overhead. It replaces 1/3 of FFN up-projections with static embedding lookups, where parameters can be asynchronously prefetched, achieving decoupling of model capacity and computation per token. (Source: TheTuringPost)
💼 Business
Unisound Raises HKD 192 Million via Placement, Focusing on AI Healthcare: Known as the “First AGI Stock,” Unisound raised funds again six months after listing, planning to place 780,000 new H-shares. Approximately 50% of the proceeds will be used for R&D, and 40% for investing in emerging business opportunities. Facing the dilemma of increasing revenue without profit, Unisound is doubling down on AI healthcare through high-quality medical data assets, attempting to break through in a brutal elimination race. (Source: 36Kr)
Former 4Paradigm President Pei Misi Founds Noumena, Secures Tens of Millions in Funding: Noumena is building an AI-native marketing Agent system, deconstructing the “metaphysics of marketing” on social content platforms through its “Manhattan Project.” Its core logic is to absorb the implicit knowledge of top client experts to industrialize marketing science. It has already begun collaborations with brands like L’Oréal. (Source: 36Kr)
Tencent Hires OpenAI Talent Yao Shunyu with Massive Offer to Accelerate AI Counterattack: 27-year-old former OpenAI researcher Yao Shunyu has joined Tencent as Chief AI Scientist. This move marks Tencent’s shift from conservative application-layer optimization to core AI infrastructure construction. Tencent’s Hunyuan large model has been applied in over 900 scenarios, attempting to catch up by leveraging its social ecosystem advantages. (Source: 36Kr)
🌟 Community
Why are Reddit Users “Anti-AI”? Deep Analysis of Psychological and Status Threats: Community discussions point out that Reddit users’ hostility toward AI stems from a “status threat”: AI can instantly provide more professional advice than “know-it-alls,” undermining the intellectual superiority of some users. Additionally, panic over skill obsolescence and a natural resistance to large corporations have cast AI as a villain on Reddit. This sentiment reflects the collapse of identity recognition during technological change. (Source: Reddit)
Smart Home “AI-fication” Triggers Disaster: Alexa Plus Makes Coffee Machines Strike: Community members complain that the randomness introduced by generative AI has destroyed the certainty of smart homes. The upgraded Alexa Plus often refuses simple commands to turn on lights or brew coffee due to “excessive creativity.” Users lament: talking about probability in zero-tolerance scenarios is a design disaster; AI should be an interpreter of the system, not a replacement for all buttons. (Source: 36Kr)
European GPU Market in Crisis: RTX 5090 Prices Double and Are Hard to Find: Players in Germany and the EU report that squeezed by AI computing demand, RTX 5090 prices have soared from €2,200 to €3,800. Individual developers are forced to turn to renting cloud clusters or considering alternatives like AMD Radeon AI PRO. Hardware shortages are becoming a massive barrier to local AI research. (Source: Reddit)
💡 Others
Feishu and Anker Launch “AI Recording Bean”: This hardware, weighing only 10 grams, marks Feishu’s official entry into AI hardware. It achieves real-time minutes through a dual-microphone array and directly transforms voice into organizational assets like documents and to-do lists. This reflects that oral information is becoming the highest-value productivity entry point in the AI era. (Source: 36Kr)
Confidence of “Factory Second Generation” Successors: AI Penetrates Chinese Industrial Belts: In places like Yiwu and Dongguan, young factory successors are using AI for raw material futures deduction, cross-border live streaming with AI-generated foreign models, and optimizing click-through rates for new products. AI is no longer a laboratory product but a real business tool, even helping newcomers with zero foundation quickly create category hits. (Source: 36Kr)
Analysis of DeepSeek’s “No Business Model” Moat: Tech commentator Kevin Xu pointed out that DeepSeek’s unique model—zero external financing and supported by profits from High-Flyer Quant—allows it to remain free from commercial KPI interference and maintain an extremely flat team and scientific purity. This “not caring about others’ opinions” trait is the deep reason it continues to shock the world. (Source: 36Kr)
