Tech news picked on 2026-10-02
AIHOT Open Source to Easily Create Customized Industry News Summary Sites with AI Integration
- AIHOT is a site operation framework that collects materials from various news sources, has them evaluated twice by a large language model to select important content, groups similar news into one incident, and then delivers the latest issues on a daily basis.
- This repository publicly shares all prompts including the entire AIHOT website, backend, selection process, clustering and popularity algorithms, evaluation criteria, and thresholds so that anyone can customize signal sources and criteria to create a personalized issue site tailored to their industry.
- Built with universal IT technologies such as Node.js 24 version, PostgreSQL 17, and Docker Compose, it supports 6 types of signal sources including RSS, web, JSON, SNS, WeChat, and scripts. All processes including signal source classification, selection, summarization, clustering, and popularity aggregation can be tuned.
- Evaluation scores are independently assigned twice per signal source document, then selection is made based on document grade and signal source stage thresholds. Clustering determines whether news covers the same incident through recent 2-week vector similarity of similar titles and summaries along with model re-verification.
- The Shimano Bicycle Museum is located in Sakai, about 15 km south of Osaka, Japan, showcasing the evolution of bicycles through various bikes and components.
- The museum is housed in a modern building focused on bicycles themselves, displaying multiple types from early hobby horses to Olympic competition models.
- Visitors can directly handle and compare the weight of bicycle tubing made from steel, aluminum, and titanium, and see the parts configuration through a disassembled high-performance racing bike display.
- The museum exhibits year-by-year models of Shimano's signature groupset, Dura-Ace, but covers little about the company's history or other business areas like fishing gear. Admission is approximately 500 yen.
Open-source 'Strata' for running 125 billion parameter AI models locally on gaming PCs
- Strata is a free open-source project that enables large AI models with 125 billion parameters (Qwen3.8-Flash-Next) to run on typical gaming PCs running Windows or Linux with NVIDIA or AMD graphics cards of at least 12GB.
- It requires at least 32GB RAM and about 80GB of storage space. During installation, it automatically detects the graphics card and recommends an appropriate model size and settings, supporting download and execution.
- This model overcomes memory limits by utilizing GPU, RAM, CPU, and SSD in a distributed manner, reads long texts at over 1,000 tokens per second, and generates answers at speeds ranging from 53 to 93 tokens per second depending on GPU performance (benchmarks measured by the development team on two gaming PCs).
- Compatible with OpenAI APIs for integration with other apps and coding agents, and since it runs locally, data is not transmitted to external clouds.
Emergence and Related Questions of AMD Gainsborough Chip Mentioned in Steam Deck 2 Development
- The newly developed semi-custom chip 'Gainsborough' by AMD has been leaked, raising the possibility of its use in Steam Deck 2.
- This chip is manufactured with TSMC's N3P process and measures 25mm x 25mm, similar to the portable Ryzen Z3 series chips AMD is preparing.
- The name Gainsborough is a codename released by AMD, not Valve, and the Steam Deck 1 generation chip 'Aerith' was also named by AMD.
- Valve has not yet decided the development timing or method for Steam Deck 2, and there is no official confirmation that this chip will necessarily be equipped in Steam Deck 2.
DeepSeek Harness Open Source App Supporting Document Work, Development, and Task Automation
- DeepSeek Harness is an open-source desktop and web-based application that supports various everyday tasks including document organization, data analysis, document writing, and slide creation.
- Development tasks such as code exploration, bug fixing, feature development, and test execution can be performed within DeepSeek Harness.
- Users can install plugins as needed or extend functions by creating plugins themselves through the interactive 'Creator Mode.'
- Automation features allow scheduling tasks and running scripts at set times, and developer diagnostic tools are included to track execution steps and tool calls.
- 'FirstDate', a pilot program operated by the Singapore government, matches government employees who have completed a survey consisting of 30 questions.
- This service uses the Gale-Shapley stable marriage algorithm, developed by Nobel laureates in economics, to attempt a stable matching based on each participant's preferences.
- The survey questions cover a variety of topics such as hobbies, lifestyle, and communication styles, combining multiple-choice and open-ended responses, with matching occurring only once per cycle.
- However, this algorithm essentially guarantees only a 'stable matching' between two groups, without ensuring emotional compatibility or actual satisfaction, and due to limitations in self-understanding and expressing preferences, it does not solve dating issues comprehensively.
Performance and Usage of the Open-Type Decision AI Model Laya-MLX Running Locally on Apple Silicon
- Laya-MLX is a local AI model running on Apple Silicon that instantly produces decision results by choosing answers or scoring among options for input sentences.
- This model operates on a proprietary MLX runtime without PyTorch or cloud support and can be easily installed and used in Python via code 공개 on GitHub.
- Based on the M3 Max chip, it records median latencies of 13.4 ms for single question processing and 7.4 ms for multilingual models, with a max throughput of 146.8~395 per second when processing batches of 64 (Presented by mizorewww, Test environment: M3 Max, macOS 27.2, Python 3.12.13).
- The original checkpoints and FP16 converted weights are 공개 on GitHub and Hugging Face, offering three main models suitable for English and multilingual input.
Android Central Continues Operating With Freelance Writers, No Staff
- Android Central has laid off all its full-time employees but the website continues to publish posts through freelance contributors.
- Out of 6 employees, 4 have publicly disclosed their layoffs, and the site operator Future stated it plans to continue running the site.
- All former managing editors and key staff were laid off, and content is now produced entirely by freelancers.
- A similar Android-focused site, Android Police, recently laid off some staff as well, and its operator Valnet has not commented further.
- Clef is a decision model that accepts text and image inputs, selects one category among predefined options, and provides probabilities for each choice, helping automate complex decision-making tasks.
- In Cloudflare's internal evaluations, Clef and Clef-flash models demonstrated equal or superior accuracy compared to the competing Jev model across multiple quality benchmarks, with processing latencies of an average of 209ms for Clef and 39ms for Clef-flash (according to Cloudflare evaluation results).
- The Clef model is deployed in the Cloudflare Workers AI environment to reduce network latency and offers full API compatibility with the existing Jev model, making it easy to adopt in existing systems.
- Cloudflare announced a reinforcement learning-based fine-tuning service for task-specific customization of Clef and is preparing a platform that will allow customers to manage everything from data collection to fine-tuning and redeployment themselves in the future.
- Jev operates on Android (ARM 64 and above with 11 or higher), Windows (version 10 1903 or above), and macOS (Apple Silicon 13 or newer). It reads on-screen conversations to determine the other party's intent and risk level, suggesting 3 possible replies.
- The app does not modify or hack chat software; instead, it uses system Accessibility services to read displayed conversation content. Replies are not sent automatically but are reflected in the user's input box.
- It locally stores contacts, notes, and optional conversation history for analysis and allows custom configuration by registering separate AI API keys for judgement, replying, and screen recognition (visual) functions.
- Supported apps include QQ (Android 9.3.50 version), X (Twitter DM) with Korean interface, and 飞书 (Lark). Unsupported apps can be analyzed manually through screen OCR. The Android WeChat version is not supported.
Multiple Security Vulnerabilities Found in Linux Kernel, Debian Advises Security Update
- Hundreds of security vulnerabilities discovered in the Linux kernel between 2024 and 2026 have been disclosed.
- The Debian project released a kernel security update (DSA-6528-1) and recommends users apply it promptly.
- The vulnerabilities include various security risks such as privilege escalation and information disclosure, making updates crucial.
- The CVE list included in this security notice pertains to the Debian distribution kernel, and separate advisories may be necessary for other Linux distributions.
LocalSend: Transfer Files Between Phones and Computers Without Logging In
- LocalSend is a free open-source file transfer app created by the LocalSend project. It works on Windows, macOS, Linux, Android, and iOS.
- By installing the app on both sending and receiving devices and connecting them to the same local network, you can share photos and files without creating accounts or logging in.
- Files move only within the local network. Remote transfers to devices in other locations over the internet are not supported.
Excalidraw for organizing travel routes or work sequences visually
- A web whiteboard created by the Excalidraw team and open-source contributors. You can place shapes, arrows, and images to create drawings that look hand-drawn.
- Multiple people can draw together in real-time on the web app, and the work is automatically saved in the browser in use.
- Drawings can be exported as PNG or SVG images, or saved as .excalidraw files for further editing.
TechCrunch Disrupt 2026, held in San Francisco from October 13 to 15
- TechCrunch Disrupt 2026 is an event for founders covering practical topics on sustainably growing businesses in the AI era, held from October 13 to 15 at Moscone West, San Francisco.
- The event is designed to help startups raise capital and expand business, featuring customized tracks per founding stage, over 200 expert sessions, and investor networking opportunities.
- The Startup Battlefield competition selects 200 startups that prepare beforehand and are judged before investors, allowing participants to engage with an extensive investor network.
- 6 main stages (Disrupt, Builders, AI, Real World AI, Smart Money, Smart Systems) provide practical advice in business operations, AI technology application, finance, and infrastructure.
Apple Developing Camera That Sends Text Alerts Instead of Video for Home Monitoring
- Apple is developing a smart home security camera that does not record video but uses artificial intelligence (AI) to detect the home environment and provide alerts in text form.
- This camera features a metal cylindrical design with a low-frame image sensor, where AI processes the video within the device without storing or transmitting it.
- It includes facial recognition to detect who has entered a room, but users receive text alerts rather than video footage.
- The product is reported to be part of a new smart home ecosystem linked with Apple's smart home hub and display currently under development.
- California will enforce a law starting 2028 July that fines robo-taxi operators if their vehicles block police or fire department emergency responses for more than 30 minutes.
- The law requires that robo-taxis have support personnel on site when the vehicles break down or obstruct emergency response, and notify local authorities of the vehicle's location and status.
- A new condition mandates that remote operators must be U.S. residents holding a valid U.S. driver's license; this aims to clarify the qualifications of remote drivers.
- Waymo, the largest robo-taxi operator in the U.S., operates approximately 4000 vehicles, with 1200 running in the San Francisco Bay Area, and has committed to compliance with this law.
Armadin, a Service That Simulates Hacking Attacks to Identify Company Security Weaknesses
- Armadin provides a service that uses multiple AI-driven agents to identify corporate security vulnerabilities in real-time.
- Founder Kevin Mandia established the cybersecurity firm Mandiant, which was acquired by Google in 2022 for $5.4 billion.
- As of October 2026, Armadin secured a Series B funding round worth $255.5 million, and the company was valued at $2.5 billion.
- This funding round included participation from major venture capital firms such as Andreessen Horowitz, Accel, Bain Capital Ventures, and Google Ventures.
Source Code Release of ZCode, an AI Coding Tool for PC, Web, and Command Line
- ZCode is an AI programming work environment that includes a desktop app, a web client, and an agent for the terminal, and it provides client, server, UI, and CLI source code together.
- Development requires specific versions of Git and Node.js 24.14.0 and pnpm 10.33.2, and tasks such as installing dependencies, running the development server, and building are carried out with pnpm commands.
- The desktop version is based on Electron and supports preparing local and remote resources and setting up development and test environments. The web development server runs at the default address http://localhost:5173에서.
- For deployment, builds are specified by OS (macOS, Windows, Linux) and CPU architecture (x64, arm64), and the command-line version can run in a Node.js environment.
Google Guided Vision Describes Small Text and Objects Seen by Camera via Voice
- The Guided Vision feature is available in Google's Gemini Live app and verbally describes objects that the user's phone camera captures.
- This feature is designed for visually impaired and low-vision users, enabling reading small text, describing nearby objects, and providing details of specific items.
- Users can ask additional questions by voice, such as requesting detailed information like a product's expiration date.
- Google clearly warns not to use Guided Vision for navigation or obstacle detection purposes.
AI Work Tool Pi 1.0 Released, Pi Durable for Long-Term Work is Experimental
- Pi 1.0 is an AI agent software used by hundreds of thousands worldwide, with the development team improving and refining it for months to ensure stable use for users and businesses.
- This version natively supports various AI and virtual models including Codemode, adds features like delayed tool loading, AI model cache warming, and mid-conversation system message adjustments.
- Earendil released a separate experimental package called Pi Durable, which maintains Pi's minimalist approach but is suitable for long-term conversations and tasks, enabling development of more complex agent applications.
- Both Pi and Pi Durable are under the MIT license, with installation scripts and documentation available on the official website (pi.dev) and GitHub (earendil-works/pi).
Lyft Settles Driver Worker Classification Lawsuit for $272.5 Million
- Lyft agreed to pay $272.5 million to settle a lawsuit alleging it misclassified drivers as independent contractors instead of employees in California.
- The lawsuit claims that from April 2016 to December 2020, drivers did not receive proper employee rights such as minimum wage, overtime pay, and paid sick leave.
- California's 2019 labor law AB5 required gig workers to be classified as employees, but the 2020 Proposition 22 created an exception, currently classifying drivers as independent contractors.
- Lyft chose settlement to avoid costs and distraction from prolonged litigation, while a similar case against Uber is still ongoing.
100 Multi-language Text Reader and Decision Maker Laya
- Laya is a system that receives texts in over 100 languages and performs multi-choice format decision-making through a single neural network process.
- The main models consist of an English model, a multilingual model for 100 languages based on mmBERT, and task-specific typed-decision models, with a router that automatically selects the appropriate model based on the request.
- It can process document inputs up to 8192 tokens, and for long documents of around 4000 tokens, it correctly answers 16 to 18 out of 20 test requests (measured based on publicly available research scripts).
- It can be installed in Python environments 3.10 or higher, supporting reinforcement learning, ONNX conversion, and batch processing. Separate libraries for TypeScript and Node.js are also provided.