Keywords：Fable 5, GPT-5.6, Agent Harness, Tencent Hunyuan Hy3, Steganographic Watermark, Scaling Law Correction
🔥 Spotlight
Fable 5 vs. GPT-5.6: Full-Scale Competition for Next-Gen Flagship Models: Discussions around Anthropic’s Fable 5 and OpenAI’s upcoming GPT-5.6 (Sol/Terra/Luna) have dominated the week. Fable 5 has shown remarkable 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 restrictions (e.g., forced downgrades, CJS framework) have sparked community backlash. GPT-5.6 Sol, meanwhile, demonstrates strong complex reasoning capabilities in benchmarks at roughly half the cost of Fable 5, positioning it as a precise counter. The competition is shifting from pure model capability to a comprehensive “cost-performance-accessibility” battle (Source: THE DECODER, Hacker News, 36Kr)
Agent Harness and “System” Optimization Become New Consensus: Karpathy points out that the biggest current AI misconception is forcing agents to work while neglecting the underlying model and system mechanisms, and outlines concepts like “Loop” and “Model Harness”. Hugging Face experiments show 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” leverages the Harness layer for scheduling, achieving performance close to Fable 5 at one-third the cost. Industry focus is shifting from singular model capability to “model + Harness + data loop” systems engineering (Source: 36Kr, 36Kr, QbitAI)
LLM Reasoning Context “Steganography” Sparks Security and Trust Crisis: Anthropic was exposed for using Unicode character modifications ( “steganographic watermark”) in Claude Code’s system prompt to identify and ban users from China. The industry condemned this as a “man-in-the-middle attack” that severely damages trust, sparking fierce debate about AI supply chain security and data sovereignty. Meanwhile, GPT-5.5 was found to have its reasoning token count oddly stuck around “516” in complex programming tasks, suggesting silent intelligence downgrades. This “516 incident” has intensified user concerns about API vendors’ “black box” operations (Source: 36Kr, 36Kr)
🎯 Trends
Scaling Law Faces “Precision Correction”: A former OpenAI researcher publicly revealed a bug in the 2020 Scaling Law paper, causing 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 the “arbitrary-order generation” of diffusion language models (dLLM) can actually be a pitfall for general reasoning tasks, proposing the more efficient “JustGRPO” method. This marks a shift from brute-force compute scaling toward more refined, theory-driven development (Source: 36Kr, 36Kr)
Tencent Releases Hunyuan Hy3 Official Version, Emphasizing Cost-Effectiveness and Real-World Deployment: Tencent’s Hunyuan Hy3 official version has been released with 295B total parameters and 21B activated. It performs excellently in coding, search, and agent benchmarks, and leads GLM5.1 in internal blind tests. It adopts Apache 2.0 open source, has extremely low API pricing (input 1 yuan/million tokens), and is fully integrated into core Tencent products like WorkBuddy and Yuanbao. Hy3 does not chase leaderboard scores but focuses on stability and cost-effectiveness in real office scenarios, marking a key step in Tencent AI’s shift from “catching up” to “practicality” (Source: Machine Heart, 36Kr)
Meta Considers Selling Compute, AI Capital Market Logic Shifts: Meta is reportedly planning “Meta Compute,” leasing its massive GPU resources to external clients. This is interpreted by the market as a signal that AI compute may shift from “extreme scarcity” to “temporary oversupply.” After the news, Meta’s stock rose, while Neocloud companies like CoreWeave, previously dependent on Meta orders, saw their stocks plummet. Meanwhile, NVIDIA’s new AA-AgentPerf benchmark evaluates compute via “concurrent agents per megawatt,” signaling a shift from “peak compute” to “productivity per unit cost” (Source: 36Kr, 36Kr)
🧰 Tools
claude-video: Gives Claude the Ability to Watch Videos: An open-source project that uses an install/plugin to allow Claude Code to process YouTube links or local videos. It automatically downloads the video, extracts frames, and transcribes audio, then sends everything to Claude for Q&A and analysis, such as diagnosing bugs or summarizing content. Supports local files (Source: GitHub Trending)
pxpipe: Reduces Fable 5 Token Costs by 70% Through “Image-Based” Context: An open-source local agent that renders text context (code, logs, etc.) into images before sending them to the model, leveraging the fact that image token pricing is cheaper than text, significantly reducing Fable 5 call costs. Tests show 59%-70% token cost savings, but at the cost of reduced accuracy in reading precise 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 (Claude Code, Cursor, Codex, etc.) with a single command npx skills add. This represents a shift from “prompt engineering” to “capability package management,” significantly lowering the barrier to customizing AI agents (Source: 36Kr)
📚 Learning
ICML 2026 Three Major Awards Reveal New Directions in AI Research: The ICML Best Paper Awards (Outstanding Paper Awards) went to a study on the “flexibility trap” in diffusion models by Tsinghua University and Alibaba, and a study on diffusion model sampling by MIT. The Test of Time Award was given to DeepMind’s A3C algorithm from 2016. These award-winning papers reveal that AI research is moving from pursuing “big and comprehensive” to deeply understanding fundamental algorithm principles, limitations, and examining their own safety (Source: Machine Heart, 36Kr)
University of Washington Releases “CONVOLVE”: Teaching AI Agents to “Stop”: Addressing the problem of agents continually searching and wasting resources on infeasible tasks, the UW team proposes the “Agentic Abstention” framework. Their method, CONVOLVE, distills “stopping rules” from past interactions and provides them as new context prompts, significantly improving the agent’s ability to “abstain” from answering, outperforming simply relying on larger models (Source: Academic Headlines, 36Kr)
Tencent and Tsinghua Release DiscoBench: Focusing on AI Search Agents’ “Clarification” Ability: A new benchmark finds that AI search agents fail not because of search ability but because they don’t proactively ask questions in the face of ambiguous queries. DiscoBench testing shows that even the strongest models succeed less than 50% of the time with ambiguity. However, once the model learns to ask questions and fill in missing information, success rates jump to over 93%, precisely pointing to a shortcoming in AI agents’ “active interaction” (Source: THE DECODER)
💼 Business
Momenta Passes Hong Kong Stock Exchange Hearing, Valued at $9 Billion: The autonomous driving solution provider Momenta has passed the HKEX hearing. Its prospectus shows its business model is shifting from one-time technology development services to “license services” charging per vehicle sold, which have very high gross margins. This “platform rent-collecting” model, combined with its “mass-production car + Robotaxi” data flywheel, positions it as a “MiniMax in AI” rather than a traditional autonomous driving company, earning a “platform-type” premium from capital markets (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 uses data research and RL environments to make AI agents more reliable, aligning with the industry’s current demand for AI that is both “productive” and “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 immense potential in life sciences and continued capital market interest. Through a model of “co-development + self-research pipeline,” the company has achieved commercial results in antibodies, vaccines, and synthetic biology (Source: Machine Heart)
🌟 Community
Are Silicon Valley Engineers Already Ditching Traditional IDEs?: Spotify’s VP of Engineering revealed in a conversation that approximately 73% of internal PRs are now AI-generated, completely transforming engineers’ workflows. They no longer manually write most code but act as “directors,” operating multiple AI agents simultaneously. This has sparked widespread community discussion, with many enjoying the efficiency gains while worrying that this “Vibe Coding” could lead to long-term “understanding debt” and technical skill degradation (Source: 36Kr)
The “Golden Age” for 35-Year-Old Programmers?: The “FDE (Field Deployment Engineer)” role is quietly becoming popular at major Chinese tech companies, offering generous salaries. Unlike traditional development, FDEs need to go deep into business lines and translate AI instructions into practical efficiency solutions. Community热议 believes that in an era where AI handles 90% of coding work, experienced senior 35-year-old engineers who understand business and can handle complex system problems are facing new career opportunities, their value far exceeding “tool people” who only write code (Source: 36Kr)
Token Bills Erode Corporate Profits, “AI Equality” Faces Reality Check: SemiAnalysis shows that its internal token spending has reached 30% of total employee salaries. Meanwhile, big companies like Uber, Tencent, and Meta have started imposing quotas and reviewing costs on employee AI usage. This shift from “encouraging use” to “controlling costs” has sparked heated discussion. Many lament that top-tier models are becoming a privilege for the few, and “compute democratization” and “AI equality” ring hollow in the face of massive commercial bills (Source: 36Kr, 36Kr)
💡 Other
Chinese AI Platforms Massively Take Down “Anthropomorphic” Services: Ahead of the July 15 effective date of the “Measures for the Management of AI Anthropomorphic Interactive Services,” platforms like ByteDance’s Doubao and Alibaba’s Qwen have urgently taken down or closed many anthropomorphic chat and user-built agent features. This move aims to comply with new regulations on content and safety of emotional AI interaction, 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 AI” platform launched in 2005 announced it will close to new customers, serving only existing users. As AI models themselves become better at generating data, and as security and compliance requirements increase, this crowdsourcing model is being gradually replaced by more professional annotation services, marking the end of an era (Source: TechCrunch)
Baidu’s “Unlimited OCR”: Ultra-Long Document Recognition via 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-pass processing of dozens of document pages without memory explosion or slowdown from long contexts. This technical approach provides an economical and efficient solution for AI handling of ultra-long texts (Source: THE DECODER)