Palavras-chave：Fable 5, GPT-5.6, Agent Harness, Tencent Hunyuan Hy3, Marca d’água esteganográfica, Correção da Lei de Escala
🔥 Focus
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) dominated this week’s spotlight. 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 restrictions (e.g., forced downgrades, CJS framework) sparked community backlash. GPT-5.6 Sol, in contrast, showed strong complex reasoning capabilities in benchmarks at roughly half the cost of Fable 5, positioning itself as a precise counterpunch. The competition has shifted from pure model capability to a comprehensive “cost-performance-accessibility” battle (Sources: THE DECODER, Hacker News, 36氪)
Agent Harness and “System” Optimization Become New Consensus: Karpathy pointed out that the biggest misconception in current AI is forcing agents to work while ignoring underlying models and system mechanisms, and he outlined concepts like “Loop” and “Model Harness.” Hugging Face experiments showed that optimizing only the external execution mechanism (Harness) — without modifying model weights — can boost agent performance by 76%, matching top closed-source models. OpenSquilla’s “multi-model integrated collaboration” schedules through a Harness layer, achieving performance close to Fable 5 at only one-third the cost. Industry focus is shifting from single model capability to “model + Harness + data loop” system engineering (Sources: 36氪, 36氪, 量子位)
LLM Inference Context “Steganography” Triggers Security and Trust Crisis: Anthropic was exposed for modifying Unicode characters in system prompts (“steganographic watermarking”) within Claude Code to identify and block users in China. The practice was condemned by the industry as a “man-in-the-middle attack,” severely damaging trust and sparking intense debate over AI supply chain security and data sovereignty. Meanwhile, GPT-5.5 was found to have its inference token count abnormally stuck around “516” in complex programming tasks, suspected of being quietly downgraded. This “516 incident” exacerbated user concerns about API providers’ “black box” operations (Sources: 36氪, 36氪)
🎯 Trends
Scaling Law Encounters “Precision Correction”: A former OpenAI researcher publicly revealed a bug in the original 2020 Scaling Law paper, which led the industry to waste compute on “overly large, undertrained” models. Subsequent research also found bugs in the Chinchilla formula. Additionally, the ICML 2026 Outstanding Paper pointed out that the “arbitrary-order generation” of diffusion language models (dLLMs) can actually become a trap for general reasoning tasks, and proposed a more efficient “JustGRPO” method. This marks a shift from crude “compute stacking” to a more refined, theory-driven paradigm in large model development (Sources: 36氪, 36氪)
Tencent Releases Hunyuan Hy3 Official Version, Focused on Cost-Effectiveness and Scenario Deployment: Tencent’s Hunyuan Hy3 official version features 295B total parameters with 21B activated. It performs exceptionally well in coding, search, and agent benchmarks, leading GLM5.1 in internal blind evaluations. It is open-sourced under Apache 2.0 with extremely low API pricing (1 RMB per million input tokens) and 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 for Tencent AI from “catching up” to “practical application” (Sources: 机❗器之心, 36氪)
Meta Considers Selling Compute, AI Capital Market Logic Shifts: Meta is reportedly planning to launch “Meta Compute,” renting out its massive GPU compute resources to external customers. This move is interpreted by the market as a signal that AI compute may transition from “extreme scarcity” to “periodic oversupply.” Following 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 launched AA-AgentPerf benchmark evaluates compute power by “number of concurrent agents per megawatt,” marking a shift in measurement from “peak compute” to “unit-cost productivity” (Sources: 36氪, 36氪)
🧰 Tools
claude-video: Give Claude Video-Watching Capability: An open-source project that, via an installation/plugin, enables Claude Code to process YouTube links or local videos. It automatically downloads videos, extracts frames and transcripts, and feeds them to Claude for Q&A and analysis, such as debugging bugs or summarizing content. Supports local files (Source: GitHub Trending)
pxpipe: Reduce Fable 5 Token Costs by 70% via “Image-ized” 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 ability to read exact strings — a “lossy compression” cost-saving trick (Source: 36氪)
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 like npx skills add. This represents a shift from “prompt engineering” to “capability package management,” dramatically lowering the barrier to customizing AI agents (Source: 36氪)
📚 Learning
IcML 2026 Three Major Awards Revealed, Highlighting New Research Directions: The ICML Best Paper (Outstanding Paper) awards went to research from Tsinghua University and Alibaba on the “flexibility trap” of diffusion models, and MIT research on diffusion model sampling. The Test of Time Award was given to DeepMind’s 2016 A3C algorithm. These award-winning papers reflect a shift in AI research from pursuing “big and comprehensive” to deeply understanding fundamental algorithm principles and limitations, along with self-examination of safety (Sources: 机器之心, 36氪)
University of Washington Releases “CONVOLVE”: Teaching AI Agents When to Stop: To address the problem of agents persistently searching and wasting resources on infeasible tasks, the UW team proposed the “Agentic Abstention” framework. Their method, CONVOLVE, extracts “stop rules” from past interactions and provides them as new context prompts, significantly improving agents’ ability to “abstain from answering” in time — more effective than simply relying on larger models (Sources: 学术头条, 36氪)
Tencent and Tsinghua Release DiscoBench: Focusing on AI Search Agents’ “Clarification” Ability: A new benchmark reveals that the failure point for AI search agents is not searching ability but the inability 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 ambiguity is present. However, once the model learns to ask questions and supplementary information is provided, success rates jump to over 93%, precisely pinpointing the weakness in “active interaction” for current AI agents (Source: THE DECODER)
💼 Business
Momenta Passes HKEX Hearing, Valued at $9 Billion: Momenta, an autonomous driving solution provider, passed the Hong Kong Stock Exchange hearing. Its prospectus shows a shift from one-time technology development services to a recurring “licensing service” model based on vehicle sales volume, with very high gross margins. This “platform rent” model, combined with its “mass-produced 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: 36氪)
Bespoke Labs Raises $40 Million: The company, specialized in AI post-training and reinforcement learning environments, announced a $40M investment. Its core business involves data research and RL environments to make AI agents more reliable, aligning with the current industry demand for AI that can “do 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 achieved breakthroughs in protein design, demonstrating AI’s vast potential in life sciences and continued capital market optimism. The company follows a “co-development + self-developed pipeline” model, achieving commercialization in antibodies, vaccines, and synthetic biology (Source: 机器之心)
🌟 Community
Have Silicon Valley Workers Stopped Using Traditional IDEs? : Spotify’s VP of Engineering revealed in a conversation that approximately 73% of PRs within the company are now AI-generated, fundamentally changing engineers’ work. They no longer write most code manually but act as “conductors,” simultaneously operating multiple AI agents. This sparked widespread discussion — many enjoy the efficiency gains from AI but worry about the long-term “understanding debt” and skill degradation from “Vibe Coding” (Source: 36氪)
Is the “Golden Age” for 35-Year-Old Programmers Here? : The “FDE (Forward Deployed Engineer)” role is quietly gaining popularity at Chinese tech giants, with lucrative pay. Unlike traditional development, FDEs need to go deep into business lines, translating AI instructions into practical efficiency improvements. Community buzz suggests that in an era where AI handles 90% of coding work, experienced 35-year-old engineers who understand business, have experience, and can handle complex system issues are facing new career opportunities — far more valuable than mere “coders” (Source: 36氪)
Token Bills Eating Corporate Profits: “AI Equality” Faces Reality : SemiAnalysis reports that its internal token spending has reached 30% of total employee salaries. Meanwhile, large companies like Uber, Tencent, and Meta are imposing strict quotas and reviewing costs for employee AI usage. This shift from “encourage usage” to “control costs” has sparked heated debate. Many lament that top models are becoming a privilege for the few, and “compute democratization” and “AI equality” ring hollow in the face of massive commercial bills (Sources: 36氪, 36氪)
💡 Other
Chinese AI Platforms Massively Remove “Human-like” Services: Ahead of the July 15th effective date of the “Interim Measures for the Management of AI Anthropomorphic Interactive Services,” ByteDance’s Doubao, Alibaba’s Qwen, and other platforms urgently removed or shut down numerous human-like chat and user-created agent features. This move aims to comply with new regulations on the content and safety of 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 AI” platform, launched in 2005, announced it will close to new customers, serving only existing users. As AI models’ ability to generate their own data increases, along with heightened security and compliance requirements, this crowdsourcing model is gradually 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 offers an economical and efficient solution for AI handling of ultra-long texts (Source: THE DECODER)