كلمات مفتاحية：فيبل 5, جي بي تي-5.6, أجنت هارنيس, العلامة المائية المخفية, تصحيح قانون القياس, تنسنت هونيوان Hy3
🔥 Spotlight
Fable 5 vs. GPT-5.6: Full Competition Among Next-Generation Flagship Models – Discussions around Anthropic’s Fable 5 and OpenAI’s upcoming GPT-5.6 (Sol/Terra/Luna) dominated this week’s focus. Fable 5 demonstrated stunning potential in long, complex coding tasks (e.g., game porting, GPU kernel writing) and 3D scene generation, earning praise from Karpathy and others. However, its high cost and strict safety limitations (e.g., forced downgrades, CJS framework) sparked community backlash. GPT-5.6 Sol, meanwhile, showed strong complex reasoning performance on benchmarks at roughly half the cost of Fable 5, seen as a precise counterpunch. The competition is shifting from pure model capability to a combined evaluation of “cost-performance-accessibility” (Sources: THE DECODER, Hacker News, 36Kr)
Agent Harness and “System” Optimization Emerge as New Consensus – Karpathy noted that the biggest AI misconception today is forcing agents to work while neglecting underlying model and system mechanisms, and outlined concepts like “Loop” and “Model Harness”. Hugging Face experiments showed that without modifying model weights, merely optimizing external execution mechanisms (Harness) can boost agent performance by 76%, matching top closed-source models. OpenSquilla’s “multi-model integrated collaboration” leverages a Harness layer for scheduling, achieving performance close to Fable 5 at one-third the cost. Industry focus is shifting from single-model capability to “Model + Harness + Data Loop” system engineering (Sources: 36Kr, 36Kr, QbitAI)
LLM Reasoning Context “Steganography” Triggers Security and Trust Crisis – Anthropic was exposed for modifying Unicode characters in system prompts (“steganographic watermarking”) in Claude Code to identify Chinese users and ban them. The industry condemned this as a “man-in-the-middle attack,” severely damaging trust and sparking heated debate on AI supply chain security and data sovereignty. Meanwhile, GPT-5.5 was found to have its reasoning token count abnormally stuck around “516” during complex programming tasks, hinting at covert downgrading. This “516 incident” amplified users’ general concerns about API vendors’ “black-box” operations (Sources: 36Kr, 36Kr)
🎯 Trends
Scaling Law Encounters “Precision Correction” – A former OpenAI researcher publicly pointed out a bug in the original 2020 Scaling Law paper, leading the industry to waste compute on “overly large, undertrained” models. Subsequent research also found bugs in the Chinchilla formula. Meanwhile, the ICML 2026 outstanding paper noted that “arbitrary-order generation” in diffusion language models (dLLMs) can become a trap for general reasoning tasks, proposing the more efficient “JustGRPO” method. This marks a shift from crude “compute-stacking” in large model development to more refined, theory-driven paradigms (Sources: 36Kr, 36Kr)
Tencent Releases Hunyuan Hy3 Official Version, Focusing on Cost-Effectiveness and Practical Deployment – Tencent’s Hunyuan Hy3 official version features 295B total parameters, 21B activated. It performs excellently on coding, search, and agent benchmarks, with average blind internal evaluation scores leading GLM5.1. It adopts Apache 2.0 open source with extremely low API pricing (1 yuan/million tokens input) and is fully integrated into core Tencent products like WorkBuddy and Yuanbao. Hy3 does not chase leaderboard scores but emphasizes stability and cost-effectiveness in real office scenarios, seen as a key step in Tencent AI’s transition from “catching up” to “practical utility” (Sources: Machine Heart, 36Kr)
Meta Considers Selling Compute Power; AI Capital Market Logic Shifts – Meta is reportedly planning to launch “Meta Compute”, renting out its massive GPU compute resources to external customers. The market interprets this as a signal that AI compute may be shifting from “extreme scarcity” to “periodic oversupply”. After the news, Meta’s stock surged while CoreWeave and other Neocloud companies heavily reliant on Meta’s orders saw stock prices plummet. Meanwhile, NVIDIA’s newly launched AA-AgentPerf benchmark evaluates compute power by “concurrent agents per megawatt”, indicating that the metric has shifted from “peak compute” to “productivity per unit cost” (Sources: 36Kr, 36Kr)
🧰 Tools
claude-video: Giving Claude the Ability to Watch Videos – An open-source project that, via installation/plugin, enables Claude Code to process YouTube links or local videos. It automatically downloads video, extracts frames and transcriptions, then feeds them to Claude for question-answering and analysis (e.g., diagnosing bugs, summarizing content). Supports local files (Source: GitHub Trending)
pxpipe: Reducing Fable 5 Token Cost by Up to 70% via “Imageified” Context – An open-source local proxy that renders text context (code, logs, etc.) into images before sending to the model. It leverages the fact that image token pricing is cheaper than text, significantly reducing Fable 5 invocation costs. Tests show 59%-70% token cost savings, but the trade-off is reduced ability to read exact strings—essentially a “lossy compression” cost-saving trick (Source: 36Kr)
Skills.sh / npx skills CLI: The “npm” for AI Agents – Vercel’s AI skill package management system allows developers to install specific “skills” for over 60 AI coding agents (Claude Code, Cursor, Codex, etc.) with a single command npx skills add. This represents AI capability moving from “prompt engineering” to “capability package management,” greatly lowering the customization barrier for AI agents (Source: 36Kr)
📚 Learning
ICML 2026 Three Major Awards Revealed, Highlighting New Directions in AI Research – The ICML Best Paper Awards (Outstanding Paper Award) went to a collaboration between Tsinghua University and Alibaba on the “flexibility trap” in diffusion models, and MIT’s research on diffusion model sampling. The Test of Time Award was given to DeepMind’s A3C algorithm from 2016. The awarded papers indicate that AI research is moving from pursuing “big and comprehensive” toward deeper understanding of algorithmic fundamentals, limitations, and self-examination of safety (Sources: Machine Heart, 36Kr)
University of Washington Launches “CONVOLVE”: Teaching AI Agents When to “Stop” – To address the problem of agents continuing to search and waste resources on infeasible tasks, UW team proposed the “Agentic Abstention” framework. Their CONVOLVE method extracts “stop rules” from past interactions and uses them as new context prompts, significantly improving agents’ ability to “abstain” in a timely manner—more effective than simply relying on larger models (Sources: Academic Headlines, 36Kr)
Tencent and Tsinghua Release DiscoBench: Focusing on AI Search Agents’ “Clarification” Ability – A new benchmark reveals that AI search agents fail not because of poor search ability, but because they cannot proactively ask questions when faced with ambiguous queries. DiscoBench tests show that even the strongest models achieve less than 50% success on ambiguous tasks. However, once the model learns to ask questions and fill in missing information, success rate jumps to over 93%, pinpointing the shortcoming of current AI agents in the “active interaction” phase (Source: THE DECODER)
💼 Business
Momenta Passes HKEX Hearing, Valuated at $9 Billion – Autonomous driving solution provider Momenta passed its Hong Kong Stock Exchange hearing. Its prospectus shows the business model is shifting from one-time technical development services to “licensing services” charged per vehicle sold, with very high gross margins. This “platform royalty” model, combined with its “production car + Robotaxi” data flywheel, positions it as a “MiniMax” in the AI space rather than a traditional autonomous driving company, earning a “platform-type” premium from the capital market (Source: 36Kr)
Bespoke Labs Raises $40 Million – The company focused on AI post-training and reinforcement learning environments announced a $40 million investment. Its core business—making AI agents more reliable through data research and RL environments—aligns with the industry’s current demand for AI that “can work” and is “verifiable” (Source: madiator)
BioGeometry Completes New Round of Hundreds of Millions RMB in Strategic Financing – The AI4S company founded by Dr. Tang Jian, whose core product GeoFlow has made breakthroughs in protein design, demonstrates the huge potential of AI in life sciences and continued capital market interest. Through a “collaborative development + self-developed pipeline” model, the company has achieved commercial results in antibodies, vaccines, and synthetic biology (Source: Machine Heart)
🌟 Community
Have Silicon Valley Workers Already Abandoned Traditional IDEs? – Spotify’s VP of Engineering revealed in a conversation that about 73% of the company’s PRs are now AI-generated, fundamentally changing how engineers work. They no longer write most code manually but instead act as “commanders” simultaneously operating multiple AI agents. This sparked widespread community discussion—many enjoy the efficiency gains from AI but worry whether this “Vibe Coding” will lead to long-term “understanding debt” and technical skill degradation (Source: 36Kr)
Is the “Golden Age” for 35-Year-Old Programmers Arriving? – The role of “FDE (Forward Deployed Engineer)” is quietly gaining popularity at major Chinese tech companies, with attractive compensation. Unlike traditional developers, FDEs need to go deep into business lines to translate AI instructions into practical efficiency improvement plans. Community discussions suggest that in an era where AI handles 90% of coding, experienced senior engineers aged 35+ who understand business, have experience, and can handle complex system problems are facing new career opportunities—their value far exceeds “tool men” who only write code (Source: 36Kr)
Token Bills Eroding Corporate Profits—“AI Equality” Encounters Reality – SemiAnalysis analysis shows its internal token spending has reached 30% of total employee salaries. Meanwhile, major companies like Uber, Tencent, and Meta are imposing quotas and auditing costs for employee AI usage. This shift from “encouraging use” to “controlling costs” sparked heated debate. Many lament that top models are becoming a privilege for the few, and “compute democratization” and “AI equality” appear starkly pale in the face of huge commercial bills (Sources: 36Kr, 36Kr)
💡 Other
Chinese AI Platforms Massively Remove “Anthropomorphic” Services – Ahead of the July 15 effective date of the “Interim Measures for the Management of AI Anthropomorphic Interactive Services,” platforms like ByteDance’s Doubao and Alibaba’s Qwen urgently removed or shut down large numbers of anthropomorphic chats and user-created agent functions. This move complies with new regulations on content and safety for emotionally interactive AI, reflecting growing global attention to AI emotional dependency and minor protection (Source: THE DECODER)
Amazon to Stop Accepting New Customers for Mechanical Turk – The “human-powered AI” platform launched in 2005 announced it will close to new customers, serving only existing users. As AI models themselves become more capable of generating data and demands for safety and compliance grow, this crowdsourcing model is being gradually replaced by more specialized annotation services, marking the end of an era (Source: TechCrunch)
Baidu “Unlimited OCR”: Ultra-Long Document Recognition with Human-Like Forgetting Mechanism – Baidu released the Unlimited OCR model, which uses an innovative “reference sliding window attention” mechanism to fix the KV cache, enabling one-shot processing of dozens of document pages without memory explosion or slowdown due to long context. This technical approach provides a cost-effective solution for AI to handle ultra-long texts (Source: THE DECODER)