Thursday, August 13, 2026

1. DeepSeek V4 Pro 0813 Released, API-Only
   T2Simon Willison 博客（Atom）

   DeepSeek has quietly released a new model, V4 Pro 0813, available via API only. There is no official announcement page, so the link goes through OpenRouter. Simon Willison tested it and noticed strikingly different outputs across low, medium, and high reasoning levels—something he hasn't seen in other models. Benchmark figures appear to have been shared in DeepSeek's official WeChat group, then copied to Reddit where moderators deleted the post as 'low-effort', and finally reposted as an ASCII-art table on Hacker News. Open-weight plans are unconfirmed, but since both the April V4 Pro and July V4 Flash models were released with weights, another open release seems likely.

2. Show HN: Experiments with Long-Term Memory Architecture for LLMs
   T2Hacker News 热门（Algolia · 中文化）

   A Hacker News submission introduces MindCache, an open-source experimental project that explores long-term memory for large language models using a collapsed tree structure. The approach aims to create a scalable memory organization method, enabling models to retain and retrieve key information across sessions. Still in early exploration with minimal discussion, the project offers a fresh direction for LLM persistent memory design.

1. Claude Code v2.1.229 Released: Remote Control Resume, SSE Keepalive, and Stability Fixes
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code v2.1.229 introduces `claude remote-control --continue` for resuming the latest Remote Control session, and adds server-supplied hook support for self-hosted runners to match managed environments. Gateway streams now emit SSE keepalive pings during long thinking pauses to prevent idle disconnects from Vertex and Bedrock. The release also adds plugin marketplace command sources, marks disconnected remote sessions as offline, and fixes issues including truncated/duplicated streaming output, narrow-terminal rendering crashes, Windows extended-length path crashes, MCP OAuth redirects, and auto-mode failures when attribution headers are disabled. Security-conscious changes also stop `/commit-push-pr` from auto-approving dangerous git commands.

2. Claude Code v2.1.231 Fixes MCP OAuth Sign-in Compatibility Issue
   T1.5Claude Code：GitHub Releases（RSS）

   Claude Code, Anthropic's agentic coding tool, has shipped version 2.1.231. The release fixes an issue where MCP OAuth sign-in would fail with a redirect URI mismatch for servers that rely on a pre-registered OAuth client, such as Slack. This fix addresses compatibility when connecting to external services that have strict OAuth configuration, ensuring that the authentication flow now correctly matches the registered URI. Developers using MCP-based integrations with Slack or similar services should upgrade to this version to avoid login failures. The change is targeted but meaningful, reinforcing the stability of Claude Code's ecosystem and its support for standard authentication mechanisms.

3. Amazon will train on Twitch streamers' content by default, unless they opt out
   T2TechCrunch：AI（RSS）

   Amazon has announced that it will train its AI models on Twitch streamers' content by default, meaning that unless streamers actively opt out, their broadcasts and interactions will be used as training data. In a livestream response to user feedback, Twitch CPO Mike Minton stated, 'If this was opt-in, nobody would opt in. That's honestly the answer.' The remarks have sparked strong backlash from creators, raising concerns about data ownership, consent, and the ethics of AI training. This decision highlights the growing tension between platform policies and creator rights.

4. Write Your First Prompt with the GitHub Copilot App
   T1GitHub Blog（网页）

   GitHub's official blog publishes a guide to help users write their first prompt with the GitHub Copilot app. The article covers choosing the right context and model, crafting effective prompts, and confidently starting your first task. It aims to lower the barrier to using AI coding assistants and improve developer productivity in collaborating with Copilot.

5. Anthropic's New Watermarks Anger Some Claude Users Who Fear Being Caught at Work or in Class
   T2TechCrunch：AI（RSS）

   Anthropic recently added a watermarking feature to its AI assistant Claude, intended to identify AI-generated content and improve transparency, but it has sparked backlash from some users. Complaints are circulating on social media that the watermarks will expose their use of AI at work or in classes, potentially being seen as cheating. Some have even called the system a "travesty." The tension between privacy, trust, and real-world usage of AI tools lies at the heart of this controversy.

6. Mesh, Automattic's AI-Powered CRM for Everyone, Comes to Android
   T2TechCrunch：AI（RSS）

   Automattic has announced the launch of Mesh, its AI-powered contacts app and relationship manager, on Android. Mesh is described as a 'CRM for everyone,' leveraging artificial intelligence to automatically organize contacts, surface context about relationships, and send smart reminders to help users stay on top of personal and professional follow-ups. Until now, Mesh was available only on iOS; the Android version brings the experience to a significantly wider user base. This move strengthens Automattic's growing portfolio of consumer tools and reflects the increasing trend of AI-assisted relationship management. The app aims to reduce the manual effort of maintaining an address book, turning it into a proactive assistant that helps users nurture meaningful connections.

1. Amid AI Safety Concerns, Three Pioneers Make the Case for Staying Open
   T2TechCrunch：AI（RSS）

   At the Ai4 conference, three AI pioneers—Geoffrey Hinton, Fei-Fei Li, and Andrew Ng—debated rising AI safety concerns while making a unified case for staying open. They discussed regulatory boundaries, the value of open-source models, and how America can stay competitive as China advances across Asia. The rare joint appearance highlights the tension between openness and safety, offering a key reference for global AI governance and policy-making.

2. Sandbar CEO: The Future of AI Wearables Is Voice
   T2TechCrunch：AI（RSS）

   In a recent episode of TechCrunch's Equity podcast, Sandbar CEO Mina Fahmi shares her perspective on the future of AI wearables, arguing that voice is the primary interface that will make these devices truly indispensable. She explains that many AI hardware companies have gone wrong by prioritizing flashy features, complicated gestures, and hardware specifications over the most natural form of human communication — voice. Fahmi believes that a voice-first approach not only makes interactions smoother but also enables wearables to fade into the background of daily life. Sandbar's Stream ring, a smart ring designed with this philosophy, aims to offer a more subtle and intuitive AI companion. The discussion provides a candid look at what has failed in the AI hardware space so far and what the next generation of products should be focusing on. It is an important conversation for startup founders, product designers, and anyone interested in the evolution of wearable technology.

1. AVA-Encoder: Towards Agent-Native Video Representation Learning
   T2arXiv AI/CL/LG（论文）

   This paper proposes AVA-Encoder, an agent-native video representation learning framework that transforms videos into knowledge graph representations and reconstructs them back, enabling hierarchical structured text storage and an editable asset layer. Its novel textual-gradient optimization mechanism allows agents to iteratively improve encoding policies via natural language instructions. Experiments show AVA-Encoder outperforms the strongest external baseline by 20.7 percentage points, and its pseudo-trained policy surpasses carefully human-tuned baselines while using 74.3% fewer system-prompt tokens. The framework also releases the first high-quality film KG representation dataset.

2. Earth Observation Embeddings as Effective Sub-grid Descriptors for Probabilistic Weather Downscaling
   T2arXiv AI/CL/LG（论文）

   This study proposes using embeddings from the Earth observation foundation model TESSERA as local surface descriptors to augment a convolutional conditional neural process for probabilistic downscaling of coarse ERA5 reanalysis fields at ~25 km resolution. Across five climatically diverse regions, the method significantly improves predictions of 2 m temperature and 10 m wind speed, with overall CRPS skill gains of 11.5% and 6.2%, respectively. Improvements persist when the coarse input is changed to forecasts from the Aurora AI model and when predicting at newly deployed stations with no regional history. To our knowledge, this is the first evidence that long-timescale Earth observation embeddings can support short-timescale weather downscaling.

3. One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL
   T2arXiv AI/CL/LG（论文）

   This paper shows that using a single large language model to simulate user behavior in multi-agent reinforcement learning for human-AI interaction systematically fails to generalize due to a phenomenon called simulator collapse. Because the simulator LLM is mode-collapsed, the policy overfits to narrow strategies that exploit the dominant mode, transferring poorly to unseen simulators and real users. The authors formalize this collapse theoretically and propose two complementary solutions: Verbalized Sampling at inference time, which broadens behavior by sampling from a verbalized response distribution, and Co-Training at training time, which jointly optimizes the policy against a population of trainable simulators. Experiments on Persuasion for Good, τ²-bench, and CooperBench show that Verbalized Sampling improves held-out success by up to 9% over single-simulator RL, and Co-Training pushes gains to 14%, with a human study confirming similar gains. The paper also releases SCOPE, an open-source framework for population co-training, and argues that training-environment diversity, not just policy diversity, is critical for real-world deployment.

4. Redistribution-based Cost Inference Improves Sparse Safe Offline RL
   T2arXiv AI/CL/LG（论文）

   Safe offline reinforcement learning typically assumes access to dense per-step cost annotations, but in practice supervisors only provide trajectory-level stop-feedback: a binary signal at the first unsafe transition with no per-step attribution. The authors frame this as a temporal credit assignment problem and propose the Redistribution-based Cost Inference (RCI) framework, which converts sparse stop-feedback into dense per-step costs via return decomposition, then trains a constrained offline policy on the augmented dataset. They show that return-equivalent redistribution preserves the feasible policy set and the optimal Lagrangian in a CMDP, establishing that the transformation is lossless in theory while yielding better-conditioned cost critic learning in practice. Experiments on highway driving and robotic manipulation demonstrate substantially lower violation rates than sparse and classifier-based baselines, with robustness to heterogeneous dataset compositions and label noise.

5. AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
   T2arXiv AI/CL/LG（论文）

   This paper proposes a test-time capability transfer method where a stronger builder model constructs inference-time harnesses that help a weaker target model solve tasks more reliably, without any parameter updates. On four representative Theory-of-Mind benchmarks, average target-model performance nearly doubles from 0.49 to 0.91. Analysis shows that gains come primarily from offloading unstable reasoning into deterministic code, benchmark-specific routing, and strict answer-format enforcement, rather than from encouraging deeper reasoning or broader sampling. The authors further find that builder-model reasoning effort improves harness quality monotonically, platform effects are modest relative to builder capability, and weaker targets benefit most. These results suggest inference-time harness design is an important complement to conventional training-time distillation.

6. Beyond Trial-and-Error: Agentic Optimization for Image-to-Video Adherence
   T2arXiv AI/CL/LG（论文）

   Modern black-box Image-to-Video (I2V) models offer powerful automated content creation, yet lack fine-grained control and reliability due to stochastic outputs. This paper introduces an 'Agentic Self-Improvement' framework that reframes video synthesis as closed-loop, goal-directed optimization. Stage one iteratively refines prompts using a multimodal LLM with Davidsonian Scene Graph (DSG) queries for semantic adherence and Common Mistake Questions (CMQ) for artifact detection. Stage two applies Bayesian optimization to co-tune stochastic seeds and CFG scales, guided by quality metrics including a novel Video-Text Adherence (VTA) score. In human preference studies, agentic videos achieved win rates up to 69%, significantly outperforming unguided search. This work offers a practical method for enhancing predictability and control, moving video generation toward reliable production-ready tools.

7. How Organizations Use AI: Evidence from ChatGPT
   T2arXiv AI/CL/LG（论文）

   This study links ChatGPT Enterprise account records to worker roles, task classifications, and public-company financial data through March 2026, enabling a privacy-preserving analysis of over 1,500 organizations and 17 million messages. It documents four facts: rapid growth in enterprise adoption, concentration among larger and more R&D- and SG&A-intensive firms, active use across job functions with high intensity among early-career workers, and a broad range of knowledge work tasks including writing, technical work, communication, and information synthesis. The results suggest that firms differ widely in the speed, breadth, and purpose of AI adoption and are still learning to integrate AI into workflows.

8. SCOUT: Unlocking Enhanced Spatial Reasoning via Structured Chain-of-Thought and Multi-Objective Process Reward
   T2arXiv AI/CL/LG（论文）

   Existing Vision-Language Models exhibit a critical bottleneck in robust spatial reasoning. Recent RL methods close the gap with verifiable outcomes, but suffer from poor credit assignment across intermediate steps, and structured reasoning approaches overlook depth perception. SCOUT introduces a structured Chain-of-Thought framework that explicitly models 3D environmental perception, plus a novel RL algorithm with multi-objective process rewards and tailored advantage estimation for fine-grained credit assignment. The authors also contribute SCOUT-24k, a structured spatial reasoning CoT dataset. Experiments show SCOUT-3B improves by 16.85% and 6.3% on general spatial benchmarks and complex spatial reasoning tasks, while SCOUT-7B surpasses GPT-4o by 4.28%, and generalizes to multi-image and video scenarios.

—