Anahtar Kelimeler：Fable 5, GPT-5.6, Agent Harness, Steganografik Filigran, Scaling Law Düzeltmesi, Tencent Hunyuan Hy3
🔥 Spotlight
Fable 5 vs. GPT-5.6: Full-scale competition among next-gen flagship models: Discussions surrounding Anthropic’s Fable 5 and OpenAI’s upcoming GPT-5.6 (Sol/Terra/Luna) dominated the week. Fable 5 demonstrated astonishing 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) sparked community backlash. GPT-5.6 Sol, meanwhile, exhibited strong complex reasoning capabilities in benchmarks and costs roughly half as much as Fable 5, positioning it as a precise counterpunch. The competition is shifting from pure model capability to a holistic “cost-performance-accessibility” battle (Sources: THE DECODER, Hacker News, 36Kr)
Agent Harness and “system” optimization become new consensus: Karpathy pointed out that the biggest misconception in AI today is forcing agents to work while neglecting underlying model and system mechanisms, and he outlined concepts like “Loop” and “Model Harness.” Hugging Face experiments showed that without modifying model weights, simply optimizing the external execution mechanism (Harness) can boost agent performance by 76%, matching top closed-source models. OpenSquilla’s “multi-model integrated collaboration” uses a Harness layer to schedule models, achieving performance close to Fable 5 at one-third the cost. Industry focus is shifting from single-model capability to system engineering combining “model + Harness + data loop” (Sources: 36Kr, 36Kr, QuantumBit)
LLM reasoning context “steganography” triggers security and trust crisis: Anthropic was exposed for modifying Unicode characters in the system prompt (“steganographic watermark”) to identify and ban users in China within Claude Code. The industry condemned this as a “man-in-the-middle attack,” severely damaging trust and sparking fierce debate over AI supply chain security and data sovereignty. Meanwhile, GPT-5.5 showed anomalous reasoning token counts clustering around “516” during complex programming tasks, suspected of being quietly de-intelligenced—this “516 incident” exacerbated 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 a bug in the original 2020 Scaling Law paper, causing the industry to waste compute on “too-large, under-trained” 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 become a trap for general reasoning tasks, and proposed a more efficient “JustGRPO” method. This signals a shift from crude “compute-stacking” to more refined, theory-driven paradigms (Sources: 36Kr, 36Kr)
Tencent releases Hunyuan Hy3 official version, emphasizing cost-effectiveness and scenario deployment: Tencent’s Hunyuan Hy3 official version has total parameters of 295B with 21B activated. It performs well in coding, search, and agent benchmarks, with internal blind tests showing it leading GLM5.1. It is open-sourced under Apache 2.0 and has 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 chase leaderboard scores but emphasizes stability and cost-effectiveness in real office scenarios, marking a key step in Tencent AI’s shift from “chasing” to “practical application” (Sources: Machine Heart, 36Kr)
Meta considers selling compute power, AI capital market logic changes: 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 shift from “extreme scarcity” to “temporary oversupply.” Meta’s stock rose sharply after the news, while Neocloud companies like CoreWeave, which previously relied on Meta orders, saw their stocks plunge. Meanwhile, NVIDIA’s new AA-AgentPerf benchmark evaluates compute by “concurrent agents per megawatt,” indicating the metric has shifted from “peak compute” to “productivity per unit cost” (Sources: 36Kr, 36Kr)
🧰 Tools
claude-video: Gives Claude the ability to watch videos: An open-source project that, via installation/plugin, allows Claude Code to process YouTube links or local videos. It automatically downloads videos, extracts frames, and transcribes audio, then passes 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: Reduces Fable 5 token cost by 70% via “image-based context”: An open-source local agent that renders text context (code, logs) into images before sending 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” cost-saving technique (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 npx skills add command. 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 revealed, indicating new AI research directions: The ICML Best Paper Award (Outstanding Paper) went to Tsinghua University and Alibaba’s study on the “flexibility trap” in diffusion models, and MIT’s research on diffusion model sampling. The Test of Time Award went to DeepMind’s A3C algorithm from 2016. The award-winning papers reveal AI research moving from pursuing “big and comprehensive” to deeply understanding fundamental principles and limitations, and examining its own safety (Sources: Machine Heart, 36Kr)
University of Washington releases “CONVOLVE”: Teaching AI agents when to “stop”: To address agents wasting resources by continuing to search when tasks are infeasible, the UW team proposed the “Agentic Abstention” framework. Their method, CONVOLVE, extracts “stop rules” from past interactions and uses them as new context prompts, significantly improving an agent’s ability to “refuse to answer” in time—more effective than simply relying on larger models (Source: Academic Headlines, 36Kr)
Tencent and Tsinghua release DiscoBench: Focusing on AI search agents’ “clarification” ability: A new benchmark points out that the failure point of AI search agents is not their search ability but their failure to proactively ask questions when faced with ambiguous queries. DiscoBench tests show that even the strongest models have a success rate below 50% when faced with ambiguity. However, once the model learns to ask questions and obtain missing information, the success rate jumps to over 93%, precisely targeting the current shortfall in “active interaction” for AI agents (Source: THE DECODER)
💼 Business
Momenta passes Hong Kong Stock Exchange hearing, valuation $9 billion: The autonomous driving solution provider Momenta passed its hearing. Its prospectus shows that its business model is shifting from one-time technology development services to “licensing services” charged per vehicle sold, with very high gross margins. This “platform rent” model, combined with its “production vehicle + Robotaxi” data flywheel, positions it as a “MiniMax” in AI rather than a traditional autonomous driving company, earning a “platform premium” from the capital market (Source: 36Kr)
Bespoke Labs secures $40 million funding: The company focused on AI post-training and reinforcement learning environments announced $40 million in 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 is “verifiable” (Source: madiator)
BioGeometry completes hundreds of millions RMB strategic financing round: Founded by Dr. Tang Jian, the AI4S company’s core product GeoFlow has achieved breakthroughs in protein design, showing the enormous potential of AI in life sciences and continued investor confidence. Through a “co-development + self-developed pipeline” model, it has achieved commercial results in antibodies, vaccines, and synthetic biology (Source: Machine Heart)
🌟 Community
Have Silicon Valley workers already ditched traditional IDEs?: Spotify’s VP of Engineering revealed in a conversation that about 73% of internal PRs are now AI-generated, completely changing engineers’ work patterns. They no longer manually write most code but act as “conductors,” simultaneously operating multiple AI agents. This has sparked widespread community discussion—many enjoy the efficiency gains but worry 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 arriving?: The role of “FDE (Forward Deployed Engineer)” is quietly gaining traction at major Chinese tech companies, with attractive salaries. Unlike traditional development, FDEs need to go deep into business lines, translating AI instructions into practical efficiency improvement plans. Community discussions suggest that in an era where AI can handle 90% of coding work, experienced senior engineers (around 35) who understand business and can handle complex systems are facing new career opportunities, far more valuable than mere “tool users” who only write code (Source: 36Kr)
Token bills eroding corporate profits, “AI equality” faces reality check: SemiAnalysis analysis shows that its internal token spending has reached 30% of total employee salaries. Meanwhile, large companies like Uber, Tencent, and Meta are imposing limits on AI usage and scrutinizing costs. This shift from “encouraging use” to “controlling costs” has sparked heated debate. Many lament that top-tier models are becoming privileges for a few, and “compute democratization” and “AI equality” appear hollow in the face of massive commercial bills (Sources: 36Kr, 36Kr)
💡 Other
China’s AI platforms massively take down “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 suspended large numbers of anthropomorphic chat and user-created agent functions. This move aims to comply with new regulations on the safety and content of emotionally interactive AIs, 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, online since 2005, announced it will close to new customers, serving only existing users. As AI models improve at generating their own data and as security and compliance requirements grow, this crowdsourcing model is being replaced by more specialized 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 due to long context. This technical approach provides a cost-effective solution for AI handling of ultra-long texts (Source: THE DECODER)