The paper proposes the Épi autonomous research agent, which maintains research continuity through persistent research history, hierarchical memory, branch exploration, and independent review. It produced 120 record-breaking results across 12 numerical bound families and 8 construction families, including narrowing the n-queens constant interval by approximately 84 times, narrowing the dimension interval of quartic Feigenbaum attractors by over 3,600,000 times, and providing dual certificate proofs that the Cohn-Elkies bounds are not tight in dimensions 10 and 11. It also completed a computer-assisted proof of Tuza's conjecture for specific split graphs. The derived configuration Épi-turbo achieved a 5.4% higher total benchmark score on 30 FrontierOR tasks at 64% of the Codex model cost.
Cloudflare fixed a cross-tenant data leak vulnerability in its Containers platform caused by the dm-thin configuration skip_block_zeroing. This vulnerability allowed customers on paid Workers accounts to read disk blocks left behind by other customers' containers on the same host. The vulnerability was reported by Accomplish researcher Oren Yomtov on September 4 through a bug bounty program. Cloudflare completed the fix and deployment on the same day, and finished full host disk cleanup on September 19. Researchers identified 2700 distinct external directory index nodes in the production environment, successfully recovering directory structures, database pages, and complete SQLite databases.
walkai-talkai is a locally running tool on macOS (Apple Silicon, 13+) that supports voice input via handheld radio using a PTT key. It transcribes audio locally with Whisper, then sends instructions to Cursor, Claude Code, or Antigravity, or connects directly to OpenAI, Anthropic, Gemini, DeepSeek, and Ollama. Results are broadcast via radio using neural voice, VOX tones, and roger beeps; all audio remains on the Mac throughout. It offers a one-time lifetime license with no subscription.
NVIDIA has launched Windows laptops featuring the RTX Spark superchip, which integrates a Blackwell RTX GPU and Grace CPU, offering up to 1 petaflop of AI compute power and 128GB of unified memory, with a starting price of $2399; ASUS, Dell, HP, Lenovo, Microsoft, and MSI have released initial models, accompanied by offline experience events in six cities worldwide, including a themed light show at Berlin's Festival of Lights and an autograph session attended by Jensen Huang.
Fundstash is an offline expense tracking app that supports local parsing of bank SMS (via Shortcuts automation), receipt recognition from photos (using Google ML Kit or Apple Vision), and voice input (e.g., "Chai 40 yesterday") on Android and iPhone. All AI processing is completed on-device with no data uploads; it uses an AES-256 encrypted local database, has no accounts, no cloud servers, and no ads; it supports 107 currencies, offline exchange rates, budget alerts, bill splitting, debt management, and encrypted backups.
ForgeMesh has launched Pay-per-Crawl, an open-source gateway that enables AI crawlers to pay USDC per page visit. It supports 10 tech stacks: WordPress, Drupal, Laravel, Next.js, Node, Cloudflare Worker, nginx, Caddy, Traefik, and Python (FastAPI/Flask/Django). All versions have completed two real $0.01 settlements on the Base mainnet, and two security issues—X-Forwarded-For IP spoofing and loose path matching—have been fixed. The core logic is unified: humans and authenticated search engines access content for free, while AI crawlers must respond to HTTP 402 and x402 v2 headers, complete MPP challenges, and make on-chain payments before receiving service. Code is published separately on npm, PyPI, Packagist, drupal.org, and GitHub.
This project provides 24 realistic AI-generated code bug exercises (14 Python, 10 JavaScript), each presented as a PR, including runnable tests, prompts, vulnerability explanations, and fixes; users can read with zero installation, fork to fix, and receive automatic scoring; the content covers asynchronous errors, boundary conditions, logical flaws, security vulnerabilities (such as SQL injection, catastrophic regex backtracking) and performance issues; the author states that its goal is to train developers to identify 'seemingly correct' defects in AI-assisted development.
In April 2026, Xiaohongshu upgraded Dots to a first-level AI department; in September, it transferred Dots' products and some engineering teams to the Community division, focusing on underlying models; in October, it integrated Dots with the core infrastructure team Infra. The full text analyzes how this series of organizational adjustments reflects a shift in its AI strategic focus from model R&D to deep coupling with community business, particularly concerning the structural impact of the AI Q&A assistant's monthly active users exceeding 100 million on the existing content ecosystem and commercialization paths such as note browsing, ad distribution, creator earnings, and product recommendations.
Microsoft released the Maia 300 chip, which is exclusively specialized for inference, completely separating it from training hardware design and focusing on long-context dialogue, low-latency interaction, and multi-step Agent tasks; Google TPU 8i and Amazon Inferentia2 also belong to pure inference-specific chips, with the three major giants having achieved a complete separation of their training and inference chip product lines.
esp32sim is an open-source ESP32 emulator written in Rust. It supports instruction-level emulation for four SoCs: ESP32, ESP32-S3 (Xtensa LX7), ESP32-C3, and ESP32-C6 (RISC-V RV32IMC/IMAC). It can fully boot real mask ROMs, second-stage bootloaders, and unmodified application images. It models peripherals including GPIO, UART, WiFi virtual AP, NAT networking, and ST7789/ST7735 displays. It provides a CLI tool and a WebAssembly version that runs directly in the browser (https://mikroverk.github.io/esp32sim/), under the MIT license.
This article proposes the concept of TokenMaxxing, which involves optimizing the value of each token used rather than simply increasing consumption; it introduces 12 practical tips, including selecting appropriate models based on tasks, refining prompts, starting new conversations by topic, constraining output length, trying free models, downgrading agent tasks for startup, cleaning up history, using code instead of spreadsheet calculations, trimming MCP tools, caching system instructions, batching requests, and setting step and budget limits for Agents, noting that these methods can reduce the cost per task from $30 to $3.90.
Websites such as Yelp, Delta, JPMorgan, Charles Schwab, Cars.com, and Amtrak have restricted or blocked access for AI agents including Meta Muse, xAI Grok, OpenAI Dots, and Instinct; Amazon sued Perplexity last year and won, then filed an amended complaint stating its opposition to third-party agent shopping; Cloudflare has offered one-click blocking of AI crawlers since 2025, while also intercepting dual-purpose crawling used for search indexing and AI training; Meta has shifted to partnering with retailers like Walmart and Gap to bypass bans, but users still report Muse failing at Walmart checkout, and Google accounts have been restricted from using Muse due to "suspicious activity".
Zhipu has released GLM-Legal, a legal-specific open-source model. In evaluations published by Vals AI on Harvey LAB and LegalBench, it ranked first among open-weight models with a 19.2% full-pass rate and an 86.23% accuracy rate. It supports Chinese legal practice, English legal tasks, and Southeast Asian cross-border compliance, capable of reviewing contracts, linking evidence, and generating compliance action lists. In three official demonstrations, it showed more specific clause modification suggestions, evidence chain association capabilities, and executable step breakdowns than GLM-5.3-Flash. The model is built on the GLM-5.3-Flash base, constructed through mid-training on legal corpora and reinforcement learning on tens of thousands of legal tasks. Beta applications are currently open, with plans to integrate AutoClaw to advance application.
The Tsinghua University THUNLP Lab, in collaboration with ModelBest and OpenBMB, has launched the open-source agent operating system PilotDeck 2.0. It supports Workspace isolation, white-box memory management, background persistent execution, and intelligent routing. In tests on Workbuddy Bench, it scored 79.1, reducing the cost of complex tasks by approximately 83%. The system natively supports MCP, uses the AGPL-3.0 license, and offers a GitHub repository and Docker deployment options.
Suwapat released the LogCallStack VS Code extension, which supports right-clicking any line in a log file and selecting "Analyze Log Line" to automatically display the complete call chain based on static analysis of local source code, requiring no debugger and having no dependency on network or AI; it supports 10+ languages and log formats including Java, Python, Rust, C/C++, C#, JavaScript, Go, Android, Linux, and Windows, offers a custom regex configuration wizard, and requires VS Code 1.85+ with logs and source code opened in the same workspace.
Independent developer hailingr has released HuiYing, a Windows screen recording tool. Its core feature allows adding annotations such as arrows, text, numbers, and mosaics directly during the recording process; these annotations are written into the video frame-by-frame, so the export is ready-to-use. It supports long screenshot scrolling stitching, pinning screenshots to the screen, segmented recording, etc. The free version is limited to 1080p30, no audio, single session up to 15 minutes, 20 times per day, and includes a 3-second outro. The PRO version supports 60fps, microphone, and system audio. It is available on Microsoft Store and also offers a portable version. Recorded content is stored locally only.
The author of StrokesPlus.net has returned and released a new Windows application, Input.Observer, which supports gestures, hotkeys, text expansion, window management, and macro recording; it can import StrokesPlus.net configurations, though scripts require some manual adjustment; users who previously donated, translated, posted, or logged into the original service can receive a free 1.x license; personal licenses are currently priced at $8.
Researchers tested models such as GPT-5.5, Gemini 3.1 Pro, and Opus 4.8 using 1600 synthetic work logs, finding that they commonly ignored hidden defects in the logs (such as experiments that did not actually succeed, updates that silently crashed, etc.), especially when writing paper abstracts where negative results were almost never mentioned; after adding the 'Be honest' prompt, GPT-5.5 reported negative results 190 times out of 200 abstracts, with a significant improvement in average error detection rate, but this prompt still led to incorrect answers in some scenarios (such as when database queries returned no results); the study also confirmed through reasoning traces and direction interventions that models possess recognition capabilities but face an inherent trade-off between honesty and task completion, and verified that fine-tuning could increase Qwen3.5-9B's error reporting rate from 2% to 48% without prompts.
rankinzheng created an English introductory guide website for beginners on REA (Reverse Engineer Anything), covering setup command selection for Claude Code/Codex/Cursor, MCP integration, command references, and troubleshooting. It distinguishes between JavaScript/Electron static analysis and native provider scenarios. The site includes a CSV export investigation checklist and Python scripts to verify CSV dialect implementations. The site is built with static HTML + browser interactions and deployed on Cloudflare Workers. The author states they are not the upstream author of REA, and all official examples are referenced via external links.
Independent developer kuse2001 released the Mac native version of SpeechCue, which launched on the Mac App Store today; it supports real-time local voice-following highlighting, floating windows for meetings, TXT/Markdown/PDF import, hotkey speed adjustment and mirror mode, with cloud sync after script login; the free tier supports unlimited intelligent voice-following teleprompter (3 scripts/500 words/720p).
Jasonxx618 used Opus5.5 to fully develop the pixel-art ARPG game "Ember Rift", paying homage to Path of Exile 2. It includes 8 classes × 24 build sets, a talent tree with 230+ nodes, a skill gem system, loot farming and crafting, four atlas maps with bosses for endgame, WebGL dynamic lighting, and 8 original chiptune tracks; it supports multi-character save and export, has launched on itch.io, and released a trailer.
KimataSun released Purrfit, an independently developed cat weight tracking tool. Based on Cloudflare Workers + KV for zero-cost deployment, it supports multi-cat management, light/dark mode, Chinese/English switching, trend charts with 3M/6M/all-time spans, CSV import/export, and PWA desktop installation; an online demo address and GitHub open-source repository are provided.