Developer 11ssss released the TailcatBox 1.3.0 desktop application, supporting Windows, macOS, and Linux. Based on Tailscale Tailcat for encrypted P2P connections, it requires no public IP or router configuration and offers direct file transfer, intranet penetration, SSH, audio/video calls, and local service sharing; it includes a trial download password with a Windows 11 image and a QR code.
Based on the browser-side WASM implementation of OnlyOffice, chaxus integrates WebLLM and local Qwen small models (1.7B/4B) to implement a document editing AI assistant that supports natural language instructions, featuring functions such as selection, rewriting, summarization, annotation, layout, and export, with tool calling supported via WebMCP; the project is fully open source, providing an online experience at https://edit.chaxus.com/ and source code repository at https://github.com/ranuts/document.
An independent developer released Retouchly.ai, offering an old photo restoration service that requires no registration, provides 3 free uses per day, and outputs images without watermarks. It supports photo sharpening, deblurring, and fading correction; the free version outputs images with a maximum side length of 2048px, uploaded images are automatically deleted within 24 hours and not used for training; currently, it cannot remove scratches or dust, and restoration results for group photos with small faces are limited.
Yann LeCun's team released a new world model, H-JEPA, based on the Hierarchical Joint Embedding Predictive Architecture (JEPA), which achieves state prediction across time scales through multi-layer abstract representations; this model does not rely on reward signals or pixel reconstruction. In the Visual AntMaze task, the three-layer structure increased the success rate from 18% to 73%, but performance decreased when the number of layers was increased to three or four in the Push-T task. The paper explicitly opposes the current mainstream Agent step-stacking paradigm and the self-improvement path of Recursive Superintelligence (RSI), advocating for establishing correct world representations before discussing optimization.
This paper proposes Elastic Latent Recurrence (ELR), a method for training recurrent language models that eliminates the performance loss (one-loop tax) observed in traditional implicit recurrent models during a single loop.
The paper proposes SWE-Duel, a self-expanding coding evaluation framework: two coding agents take turns acting as task authors (Red) and reviewers (Blue) in real code repositories through red-blue adversarial play. Red must implement features, plant fixable defects, and verify their solvability, while Blue must fix the defects without prior knowledge. Across 1500 rounds of testing on 15 repositories with 11 models, GPT-5.5 emerged as the strongest reviewer with a 98.3% fix rate, and Claude Opus 4.8 emerged as the strongest task author with a 27.3% win rate for the Red side.
DiffusionART is a new online reinforcement learning method for diffusion models. It constructs a shared optimization direction by aggregating prompt rewards and introduces trust-region-masked quadratic updates to achieve drift-adaptive recovery, without requiring EMA tuning or differentiable reward models. On HunyuanVideo-13B, compared to DiffusionNFT and Flow-GRPO-Fast, it saves 12.02x and 11.72x GPU hours respectively, while improving average training rewards by 28.8% and 53.0%.
AutoIndexer is a novel generative model that employs Chain of Edits training and decoding mechanisms, supporting insertion, replacement, and deletion operations on existing token sequences after generation; it simulates text editor behavior by predicting edit tokens and moving cursor positions, with training data generated by applying reverse edit perturbations to ground-truth labels for the model to learn inverse restoration; this method maintains standard LLM benchmark performance while demonstrating post-generation context editing capabilities not possessed by most autoregressive and diffusion models; code is open-sourced at https://github.com/AutodeskAILab/autoindexer.
Based on the AIHOT open-source foundation by Kazike, the author completed secondary development for outbound and indie developer scenarios: restructured the information source pool to include Product Hunt, HN Show HN, Trends.vc, V2EX Creation Node, and Stripe official updates; rewrote the scoring Prompt using 5-axis criteria such as commercial feasibility, open-source maturity, and practical implementability; added Telegram channel broadcasting with hardcoded anti-disturbance mechanisms (≥68 points, ≤4 items per day, 2.5-hour interval, silent from 23:00–08:30).
suninrain has released Calcanto, an independently developed calculator app supporting iPhone / iPad / Mac. It features precise decimal arithmetic (0.1 + 0.2 = 0.3), bilingual Chinese-English offline voice announcements (including localized reading styles), five themes with sound effects, scientific calculations, and unit & currency conversion. It is free with no ads, in-app purchases, accounts, or third-party SDKs. App Store and official website links are provided.
baipengbb released the open-source website AIGCNEWS (https://aigcnews.cn), which aggregates AI open-source projects, papers, news, product updates, and UI inspiration materials. It supports natural language search, trend analysis, and historical review. An MCP interface (https://aigcnews.cn/mcp) is provided for AI Agents to retrieve projects, papers, rankings, and UI source code. The frontend uses Vue 3 + TypeScript, the backend uses FastAPI, and the MCP service is implemented in Go. It supports Docker Compose self-hosting, and the code is licensed under AGPL-3.0.
Analysis by CoinW Research Institute indicates that compute power is evolving from GPU hardware into a financial asset capable of being financed, priced, and traded. Wall Street has built Compute Credits based on long-term customer contracts and future cash flows through cases like CoreWeave, while CME launched H100/B200 compute futures. Meanwhile, Crypto participates through multiple channels including mining farm transitions, DePIN aggregation, on-chain credit (such as GAIB, USD.AI), and Bitget perpetual contracts. Currently, the most mature infrastructure revolves around credit and price systems tied to real compute cash flows, rather than GPU tokenization.
For promoting new products on Amazon, it is recommended to first enable automatic targeting to discover relevant search terms and ASINs, then build manual keyword or product targeting based on real data for validation and control. Automatic targeting should run until objects with high relevance and signals such as clicks, spend, or orders appear, at which point they are transferred to manual testing; before transferring out, each item must be verified for relevance, behavioral signals, and testing objectives.
The article points out that blindly increasing budgets on Google Ads often fails for foreign trade enterprises due to structural account issues, such as shared budgets across multiple markets, confused conversion goals, and mismatches between landing pages and search intent. It proposes building campaigns based on business objectives (target customers, markets, qualified leads), aligning ad group themes with landing pages, setting primary and secondary conversions in tiers, verifying tracking accuracy, optimizing keywords and ad creatives based on purchase intent, and ensuring policy compliance.
Shenzhen Netvue's Birdfy smart bird feeder has sold a cumulative 1000000 units, with flagship models priced at USD 100–300+; it achieved high average order value breakthroughs through tiered AI recognition fees (lifetime free/subscription/membership), channel-exclusive software benefits (e.g., Walmart offering 30 days of AI), differentiated product selection for US and UK markets (promoting hummingbird feeders in the US, adapting to RSPB new guidelines in the UK), and endorsement by birdwatching organizations (hiring Stephen Moss as Chief Birdwatching Advisor); its core moat lies in non-hardware assets accumulated over years, such as 2200000 registered accounts, 350000 monthly active users, daily production of 8000000 videos, and trust from conservation organizations.
Amazon is conducting a grayscale test of the "XX BRAND approved" brand authorization tag on its front end. Located in the seller information module below the Buy Box, it is displayed only for sellers who have completed brand registration and been directly marked as authorized by the brand owner in the backend. Currently, some listings on the US site display "Sourced from [Brand]," while the German site displays "[Brand] approved." This tag clearly identifies to consumers that products come from the brand owner or officially authorized sellers. It has already improved conversion rates for tagged products in categories such as electronics, beauty, and maternity/baby, while weakening the competitiveness of unauthorized piggyback sellers.
Skills Hub v1.0.0 is an open-source desktop application supporting macOS, Windows, and Linux for unified management of AI agent skills; users can install once and sync to 48 AI programming tools (including Claude Code, Codex, Cursor, OpenCode), support syncing skill files and tags across devices via GitHub/GitLab/Gitee, and enable AI-driven installation, updates, and organization by calling the same skill library through coding agents; no Node.js or npm required.
Starting November 1, 2026, the EU will impose a €2 per item EU Handling Fee (UHF) on e-commerce goods entering from outside the EU, with no product value threshold. The fee is non-refundable even in cases of quality issues or consumer returns. This fee runs parallel to the already implemented €3 temporary tariff on low-value parcels, and both can be levied cumulatively.
Tencent has launched the 'Penguin Micro-Order' AI creator order platform, which supports demand posting, intelligent matching, and settlement within WeChat. The platform connects creators with capabilities in AI image generation and video editing with demand-side users such as e-commerce merchants, individuals, and teams. It provides a full-process closed-loop service within WeChat, covering demand posting, three-party communication, order payment, work delivery, and settlement, and includes an AI assistant to help clarify requirements.
From late September to early October 2026, the open-source AI penetration testing agent ARTEX was used to attack 9 banks in South Korea, resulting in the leakage of at least 65,000 customer data records; this tool integrated large models such as DeepSeek, Claude Code, and Grok, and completed automated multi-round penetration under the command of 1 non-top-tier hacker; its developer announced on October 8 that the project would be closed-source and stated they would not bear responsibility for misuse; the incident highlights core issues including the dual-use nature of AI agents, the 'bucket effect' risk of peripheral systems, ambiguous boundaries of model vendor liability, and the difficulty of tracing cross-border attack infrastructure.
Amazon beginners should prioritize Sponsored Products (SP) ads to validate single-product conversion capabilities, as they directly lead to individual ASIN detail pages, facilitating the collection of search term and product targeting signals. Sponsored Brands (SB) require brand registration, a storefront, and multi-product combinations as prerequisites, making them suitable for capturing brand keywords and series browsing. The former Sponsored Display (SD) has been integrated into the "Display Ads" system, applicable for remarketing or audience expansion after establishing stable conversions. The entire article explains the differences among the three based on objectives rather than abbreviations, pointing out common misconceptions and providing a phased practical sequence.
cobiao launched the website howmakersbuild.com, which includes 6677 products, 266 tools, and 29k verified usage evidence entries (sources include subprocessor lists, loaded scripts, DNS records, founder statements, job postings, and public repository dependencies). It offers features such as Plan my stack, Check my stack, comparison pages, and a price calculator, supporting cost estimation for domestic model APIs including DeepSeek, Kimi, Tongyi Qianwen, GLM, and Doubao. The website is deployed for free using Astro + Cloudflare Pages, with data stored in Git YAML, and AI programming uses Claude.