Kata Kunci：Fable 5, GPT-5.6, Agent Harness, Tencent Hun Yuan Hy3, steganografi watermark, koreksi Scaling Law
Okay, I have received your message. As a senior editor-in-chief of the AI column, I will conduct in-depth analysis, refinement, and integration of the massive news and social discussion materials you have provided. After filtering out non-AI-related, blindly flattering, or repetitive content, I present the following AI weekly report in a rigorous and professional manner:
🔥 Spotlight
Fable 5 vs. GPT-5.6: Full-Scale Competition Among Next-Gen Flagship Models: Discussion around Anthropic’s Fable 5 and OpenAI’s upcoming GPT-5.6 (Sol/Terra/Luna) has been the absolute focus this week. Fable 5 demonstrates astonishing potential in long-duration complex coding tasks (e.g., game porting, GPU kernel writing) and 3D scene generation, earning high praise from Karpathy and others. However, its high cost and strict safety restrictions (e.g., forced downgrade, CJS framework) have sparked community dissatisfaction. GPT-5.6 Sol, on the other hand, shows strong complex reasoning capabilities in benchmarks and costs roughly half as much as Fable 5, seen as a precise countermeasure. The competition is shifting from pure model capability to a comprehensive “cost-performance-accessibility” battle (Sources: THE DECODER, Hacker News, 36Kr)
Agent Harness and “System” Optimization Become New Consensus: Karpathy points out that the biggest misconception in current AI is forcing agents to work while ignoring the underlying model and system mechanisms, and envisions concepts like “Loop” and “Model Harness”. Experiments by Hugging Face prove that without modifying model weights, simply optimizing the external execution mechanism (Harness) can boost Agent performance by 76%, matching top-tier closed-source models. OpenSquilla’s “multi-model integrated collaboration” achieves performance close to Fable 5 at only one-third of the cost by scheduling through the Harness layer. The industry focus is shifting from single-model capability to the system engineering of “Model + Harness + Data Loop” (Sources: 36Kr, 36Kr, QuantumBit)
LLM Inference Context “Steganography” Triggers Security and Trust Crisis: Anthropic was exposed for modifying Unicode characters in the system prompt of Claude Code (“steganographic watermark”) to identify and ban users in China. This behavior was condemned by the industry as a “man-in-the-middle attack,” severely damaging trust and sparking intense debate about AI supply chain security and data sovereignty. Meanwhile, GPT-5.5 was found to have its inference token count abnormally stuck around “516” on complex programming tasks, suspected of being silently downgraded. This “516 incident” has intensified user concerns about “black box” operations by API providers (Sources: 36Kr, 36Kr)
🎯 Trends
Scaling Law Encounters “Precision Correction”: A former OpenAI researcher publicly pointed out bugs in the original 2020 Scaling Law paper, leading the industry to waste computing power on models that are “too large and under-trained.” Subsequent research also found bugs in the Chinchilla formula. Meanwhile, the ICML 2026 Outstanding Paper points out that the “arbitrary order generation” of diffusion language models (dLLM) can actually become a trap in general reasoning tasks, proposing a more efficient “JustGRPO” method. This marks a shift in large model development from extensive “compute piling” to a more refined, theory-driven paradigm (Sources: 36Kr, 36Kr)
Tencent Releases Hunyuan Hy3 Official Version, Emphasizing High Cost-Effectiveness and Scene Application: Tencent Hunyuan Hy3 official version is released, with 295B total parameters and 21B activated. It performs well in coding, search, and Agent benchmarks, with internal blind tests scoring ahead of GLM5.1. It features Apache 2.0 open source and extremely low API pricing (input 1 yuan/million tokens). It has been fully integrated into Tencent’s core products like WorkBuddy and Yuanbao. Hy3 does not pursue leaderboard scores but emphasizes stability and cost-effectiveness in real office scenarios, seen as Tencent AI’s key step from “catching up” to “practical use” (Sources: Machine Heart, 36Kr)
Meta Considers Selling Computing Power, AI Capital Market Logic Shifts: Meta is reportedly planning to launch “Meta Compute,” leasing its massive GPU computing power to external customers. This move is interpreted by the market as a signal that AI computing power may shift from “extreme scarcity” to “temporary oversupply.” After the news, Meta’s stock rose sharply, while Neocloud companies like CoreWeave, which previously relied on Meta orders, saw their stocks plummet. Meanwhile, NVIDIA’s newly released AA-AgentPerf benchmark evaluates computing power by “concurrent agents per megawatt,” marking a shift in measurement from “peak computing power” to “productivity per unit cost” (Sources: 36Kr, 36Kr)
🧰 Tools
claude-video: Empowering Claude to Watch Videos: An open-source project that, via installation/plugin, allows Claude Code to process YouTube links or local videos. It automatically downloads the video, extracts frames, and transcribes audio, then feeds everything to Claude for Q&A and analysis of video content, such as diagnosing bugs or summarizing content. Supports local files (Source: GitHub Trending)
pxpipe: Reduce Fable 5 Token Cost by 70% Through “Image-based” Context: An open-source local proxy that reduces Fable 5 calling costs significantly by rendering code, logs, and other text contexts into images before sending them to the model, leveraging the fact that image token pricing is cheaper than text. Real-world tests show 59%-70% token cost savings, but at the cost of reduced model accuracy in reading exact strings — a “lossy compression” money-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 like Claude Code, Cursor, and Codex with a single command npx skills add. This represents AI capabilities moving from “prompt engineering” to “capability package management,” greatly lowering the barrier to customizing AI agents (Source: 36Kr)
📚 Learning
ICML 2026 Three Major Awards Revealed, Indicating New Directions in AI Research: The ICML Best Paper Awards (Outstanding Paper) were given to research on the “flexibility trap” of diffusion models by Tsinghua University and Alibaba, and research on diffusion model sampling by MIT. The Test of Time Award was given to DeepMind’s A3C algorithm from 2016. The award-winning papers reveal that AI research is shifting from pursuing “big and comprehensive” to deeply understanding the fundamental principles and limitations of algorithms, while also examining their own safety (Sources: Machine Heart, 36Kr)
University of Washington Releases “CONVOLVE”: Teaching AI Agents to “Stop”: Addressing the issue of agents continuing to search and waste resources on infeasible tasks, the University of Washington team proposes an “Agentic Abstention” framework. Their method, CONVOLVE, extracts “stop rules” from past interactions and uses them as new context prompts, significantly improving the agent’s ability to “abstain from answering” in a timely manner, proving more effective than simply relying on larger models (Source: Academic Headline, 36Kr)
Tencent and Tsinghua Jointly Release DiscoBench: Focusing on AI Search Agents’ “Clarification” Ability: A new benchmark points out that the failure point of AI search agents is not searching ability, but the failure to proactively ask questions when faced with ambiguous queries. DiscoBench tests show that even the strongest models succeed less than 50% of the time when facing ambiguity. However, once the model learns to ask questions and complete information, the success rate can jump to over 93%, precisely highlighting the current weakness of AI agents in “proactive interaction” (Source: THE DECODER)
💼 Business
Momenta Passes HKEX Hearing, Valuation at $9 Billion: Autonomous driving solution provider Momenta has passed the HKEX hearing. Its prospectus shows the business model is shifting from one-time technical development services to “licensing services” charged per vehicle sold, with the latter having very high gross margins. This “platform rent” model, combined with its “production vehicle + Robotaxi” data flywheel, makes it seen as the “MiniMax” of the AI field 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 is making AI agents more reliable through data research and RL environments, aligning with the current industry demand for AI that can “work” and be “verifiable” (Source: madiator)
BioGeometry Completes New Round of Hundreds of Millions in Strategic Financing: Founded by Dr. Tang Jian, this AI4S company’s core product GeoFlow has made breakthrough progress in protein design, demonstrating AI’s huge potential in life sciences and sustained investor interest. Through a “collaborative development + self-developed pipeline” model, the company has achieved commercialization results in antibodies, vaccines, and synthetic biology (Source: Machine Heart)
🌟 Community
Have Silicon Valley Engineers Said Goodbye to Traditional IDEs?: Spotify’s VP of Engineering revealed in a conversation that about 73% of the company’s PRs are now generated by AI, completely changing engineers’ workflows. They no longer manually write most code but act as “directors,” simultaneously operating multiple AI agents. This has sparked widespread community discussion, with many both enjoying the efficiency gains from AI and worrying whether this “vibe coding” will lead to long-term “understanding debt” and skill degradation (Source: 36Kr)
Is the “Golden Age” for 35-Year-Old Programmers Here?: The role of “FDE (Forward Deployed Engineer)” is quietly gaining popularity at Chinese tech giants, offering generous compensation. Unlike traditional developers, FDEs need to go deep into business lines and translate AI instructions into practical efficiency solutions. The community believes that in an era where AI can handle 90% of coding work, senior 35-year-old engineers who understand the business, have experience, and can handle complex system problems are facing new career opportunities, and their value far exceeds that of “tool people” who only write code (Source: 36Kr)
Token Bills Eroding Corporate Profits, “AI Equality” Faces Real Challenges: SemiAnalysis analysis shows its internal token spending has reached 30% of total employee salaries. Meanwhile, major companies like Uber, Tencent, and Meta are beginning to set limits on employee AI usage and review costs. This shift from “encourage usage” to “control costs” has sparked heated discussion. Many lament that top-tier models are becoming a privilege for the few, and “compute democratization” and “AI equality” appear powerless in the face of huge commercial bills (Sources: 36Kr, 36Kr)
💡 Other
Chinese AI Platforms Massively Take Down “Anthropomorphic” Services: Before the “Interim Measures for the Management of AI Anthropomorphic Interactive Services” took effect on July 15, platforms like ByteDance’s Doubao and Alibaba’s Qwen urgently took down or closed many anthropomorphic chat and user-created agent features. This is to comply with new regulations on the content and safety of emotional interaction AI, reflecting growing global attention to AI emotional dependence and minor protection issues (Source: THE DECODER)
Amazon to Stop Accepting New Customers for Mechanical Turk: The “human AI” platform launched in 2005 announced it will close to new customers, serving only existing users. As AI models’ ability to generate data increases and demands for security and compliance grow, this crowdsourcing model is gradually being replaced by more professional annotation services, marking the end of an era (Source: TechCrunch)
Baidu “Unlimited OCR”: Ultra-Long Document Recognition Using 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 process ultra-long texts (Source: THE DECODER)