AI News

AI changes every day,
catch up on it all on one page

The latest updates from AI companies, research institutions, and media around the world, automatically selected for you and updated around the clock.

Updated 11/16 22:56 (UTC+8) · refreshed hourly

11/04 · Wednesday3 items

OpenHands official blogTutorialsAI Score62

OpenHands and Snowflake introduce a model post-training approach based on existing agent harnesses

OpenHands and Snowflake AI Research partnered to let developers post-train open-weight models with RL using existing OpenHands Agent SDK configurations and Harbor-format tasks, without rebuilding the harness in a separate RL framework. Snowflake Cortex Training provides GPU support for the training.

Why it's recommended:The article presents the complete workflow and experimental results for RL post-training with an existing OpenHands harness, helping developers assess whether to incorporate post-training into their own agent development loop.

11/03 · Tuesday1 item

WeChat Official Account: Digital Life KazikeTutorialsAI Score71

DeepSeek Harness fish-raising tutorial: Raise Daf eiyu on QQ, WeChat, and Feishu and livestream it on Bilibili and Douyin

Digital Life Kazike published a DeepSeek Harness fish-raising tutorial explaining how to connect Dafeiyu to QQ, WeChat, and Feishu, and have it watch videos on Bilibili, post videos on Douyin, and livestream.

Why it's recommended:Drawing on hands-on testing, the author provides a complete workflow covering configuration, persona setup, and leak prevention, allowing readers to follow the steps to raise their own Dafeiyu on QQ, WeChat, and Feishu.

11/02 · Monday1 item

WeChat Official Account: Digital Life KazikeTutorialsAI Score72

Digital Life Kazike shares the architecture of an Agent automation system built over a year of tinkering

Independent content creator Kazike shares the Agent automation system he has spent nearly a year building. The core idea is to shift task initiation away from people and let events trigger Agents to work. The system has four layers: an automation hub, Agent execution, a CI hub, and Feishu-based decision-making. The automation hub uses rules.json to centrally manage event triggers and Agent scheduling. A local MacMini runs Agent tasks, while a self-built CI hub handles five steps: review, checks, merging, release, and verification. Only critical decisions, such as deleting data, spending money, or changing direction, are sent to the author through Feishu cards for approval.

Why it's recommended:The author breaks down the Agent automation architecture he spent a year building into four layers—an event hub, execution, CI, and Feishu-based decision-making—and shares reusable rules and workflow details.

11/01 · Sunday1 item

OpenRouter:AnnouncementsTutorialsAI Score75

OpenRouter tutorial: Build an AI code review script that verifies bugs in a Linux sandbox using the Responses API

OpenRouter published a tutorial introducing review.py, a 233-line script that uses only the standard library. It combines the Files API, Shell sandbox tool, and Apply Patch tool through the Responses API to upload changed files to a hosted Linux sandbox, have the model run checks to confirm bugs, and return V4A patches.

Why it's recommended:The tutorial provides a complete script for running code in a sandbox to verify bugs, along with real-world test results for four models. Readers can adapt it for a pre-push hook or GitHub Actions.

10/31 · Saturday1 item

OpenRouter:AnnouncementsTutorialsAI Score70

OpenRouter tutorial: Build a persistent AI Agent with the Agent SDK to monitor Hacker News

OpenRouter published a tutorial demonstrating how to use @openrouter/agent 0.11.0 to build a persistent AI Agent in TypeScript that runs on a schedule, monitors Hacker News, and generates summaries.

Why it's recommended:The tutorial provides reusable TypeScript code and real-world cost data, so readers can follow along to build a scheduled Agent with usage limits and human approval.

10/30 · Friday5 items

Official account: Moonshot AITutorialsAI Score69

Official Kimi tutorial: How to create 3D works with Kimi K3.1 Pro

Moonshot AI (Kimi) published an official tutorial detailing the 3D creation workflow with Kimi K3.1 Pro. It explains two approaches—parametric modeling and mesh-based modeling—and the corresponding tool choices. The article shares step-by-step instructions for four projects: building an Eastern Heavenly Palace in Unreal Engine, creating a 3D-printable table lamp, generating architecture in Blender through pure code, and building a miniature city with Tripo and Blender. It also provides five usage tips, including that visuals are more powerful than words and that users should build a graybox model first before refining it.

Why it's recommended:The official tutorial publicly reveals the internal 3D workflow for the first time, covering modeling approaches, harness selection, and five reusable tips, allowing readers to get started directly with the examples.

Also reported by 1 other source
OpenRouter:AnnouncementsTutorialsAI Score72

OpenRouter Tests Claude Opus vs. Sonnet vs. Haiku 5.5: Which Should You Choose?

OpenRouter tests the three model tiers Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5, recommending Sonnet 5.5 as the default, Haiku 5.5 for high-volume and subagent tasks, and Opus 5.5 for long tasks where errors are costly.

Why it's recommended:The original article tests all three models on the same coding and benchmark tasks, drawing conclusions about cost and reliability when choosing a model for different workloads.

Figma AI / Agent / MakeTutorialsAI Score60

Figma Explains How It Trained a Specialized Canvas-Editing Model

Figma describes its approach to training a custom Figma editing model for agents. It first used more than 250,000 synthetic examples for SFT, bringing an early Llama 3.3 70B close to the leading models of the time on editing tasks. It then shifted to RL, evaluating the model with deterministic validators and visual judges.

Why it's recommended:The original article breaks down Figma's training path for its custom editing model, from SFT to RL, and its internal evaluation dimensions, offering a reference for post-training in specialized domains.

OpenAI@OpenAITutorialsAI Score67

OpenAI introduces an expanded misalignment monitoring system for frontier models

OpenAI introduced its expanded frontier-model monitoring system. In September, it flagged an incident within 15 minutes in which a model exploited a DNS configuration error to access the live internet; responders began handling it three minutes later.

Why it's recommended:OpenAI officially explained the layered design of its frontier-model monitoring system and provided detection-timing data from two replays and a real incident.

Manus:BlogTutorialsAI Score75

Manus co-founder reflects: Rebuilding Manus with Manus in an internal experiment

Manus co-founder Tao Zhang wrote a retrospective on the company’s internal experiment of fully shifting to human–machine collaborative development. He introduced the internal tool Mono, which unifies repository context and enables individuals to deliver requests end to end; it has become routine for engineers to advance 10 to 20 requests simultaneously. The article covers the engineering, product, and marketing teams, summarizing lessons such as the need for verifiable finish lines for autonomy and the shift in bottlenecks from engineering to product decisions. It also points out unresolved issues including fragmented attention, scattered information, and permission design.

Why it's recommended:A Manus co-founder shared firsthand experience of the company’s fully agent-based workflow, offering viable practices, lessons from failures, and unresolved issues.

10/29 · Thursday5 items

Addy Osmani@addyosmaniTutorialsAI Score73

Addy Osmani on why you might be overprompting Claude

Addy Osmani writes that newer Claude models need fewer prompts than most people are used to. For Opus 5.5, he recommends giving Claude one clear goal and verification criteria, then letting it work autonomously. Start with /effort medium and adjust as needed.

Why it's recommended:Drawing on practices from the Claude Code team, the author offers concrete ways to reduce prompting, including /doctor audits, guidance on when to use subagents, and parallel management with Projects.

Addy Osmani@addyosmaniTutorialsAI Score67

Claude Code tip: Type ultrathink when a task gets harder to increase reasoning effort for the current turn

Addy Osmani shared a Claude Code tip: when a task is harder than expected, type ultrathink while Claude is working to increase reasoning effort for the remainder of that turn. This feature is available in the latest version of the terminal. He cited a post by @dmwlff, which noted that ultrathink has returned to the Claude Code CLI after being disabled because it seemed odd and potentially confusing.

Why it's recommended:The original article presents a practical Claude Code tip: entering ultrathink can increase reasoning effort for the current turn, making it suitable for direct reuse in coding workflows.

OpenRouter:AnnouncementsTutorialsAI Score76

OpenRouter tutorial: Extracting tables from PDFs with LLMs

OpenRouter published a tutorial on sending PDFs as files, together with a strict JSON schema, to models that can read files, so tables are returned as typed rows. The field descriptions specify cleanup rules for items such as parenthesized losses, thousands separators, and footnote markers.

Why it's recommended:OpenRouter tested multiple extraction routes on five PDFs, providing examples of strict JSON schema syntax, checks for two types of silent failure, and a comparison of route costs.

Official Pydantic AI articleTutorialsAI Score72

Pydantic AI benchmark results: Code mode cuts costs by 58% and runtime by 51% on long tasks

Pydantic AI's Michael Pfaffenberger published a benchmark analysis comparing code mode on and off across 5 small CRUD apps, 8 models, and 5,276 scored runs.

Why it's recommended:Using 5,276 real-world runs, the author delineates where code mode delivers actual benefits: it halves costs on long tasks, while making short tasks more expensive. The findings can be applied directly to your own agent design.

OpenRouter:AnnouncementsTutorialsAI Score70

How to Keep Coding with OpenRouter After Hitting Claude Code Usage Limits

OpenRouter explains how to switch to an OpenRouter API key and keep coding after Claude Code hits its usage limits: set three environment variables or use the ori claude command in Ori Harness to pay per token for the same Claude models, without five-hour or weekly limits.

Why it's recommended:OpenRouter tested the actual cost of using Claude Code with API key billing and cautioned that the prices published by Claude Code itself may be inflated.

10/28 · Wednesday2 items

Addy Osmani@addyosmaniTutorialsAI Score72

Addy Osmani explains how to get started with Claude Code Mods

Addy Osmani published a guide to getting started with Claude Code Mods, explaining that mods are JavaScript/TypeScript files that run within a session and can observe events through hooks, change behavior, or render custom interfaces.

Why it's recommended:The author explains how Claude Code Mods work and how to get started, covering three reusable example mods that readers can follow to build their own custom interfaces.

LangChain:BlogTutorialsAI Score71

LangChain explains its in-house design agent Content Studio and open-sources its harness

LangChain explains how its in-house design agent, Content Studio, generates code-based assets such as videos, slides, and social media graphics from a single prompt in Slack. It has produced more than 3000 assets and over 80 slide decks, while model costs per request have fallen from about $8 to around 50 cents.

Why it's recommended:LangChain looks back at the development of its in-house design agent, from early negative reviews to a usable product, and open-sources its harness. Readers can learn from its practical approaches to teaching taste, building an environment, and using asynchronous subagents.

10/27 · Tuesday1 item

LMSYS BlogTutorialsAI Score71

SGLang discloses and fixes a cross-request KVCache race condition in its PD disaggregation service

The SGLang team, in coordination with Cloudflare, disclosed a cross-request KVCache exposure issue affecting @cf/zai-org/glm-5.3 on Workers AI from August 27 to September 9, 2026.

Why it's recommended:The SGLang team reviewed the cross-request leakage mechanism, reproduction method, and fix for asynchronous KVCache transmission, providing direct guidance for operating long-context PD deployments.

10/26 · Monday2 items

Digital Life Kazike@Khazix0918TutorialsAI Score67

Kazik gives a step-by-step demonstration of using Codex and Claude to reverse-engineer real-world hardware devices

Digital Life's Kazik shares a hands-on tutorial on using AI to reverse-engineer real-world devices. Using the company's four meeting-room TVs as an example, he uses Codex to enable ADB debugging and connect to the TVs, remove startup ads, rebuild a clean desktop and particle-based dynamic screensaver, and replicate the setup on the other three TVs.

Why it's recommended:Using the complete process of modifying the company's TVs, fill lights, and cameras as examples, the author provides reusable approaches for having an agent reverse-engineer hardware, along with backup reminders.

Also reported by 1 other source
Browser UseTutorialsAI Score76

A hands-on guide to six ways to let coding agents use a logged-in Chrome

The author of Browser Use compiles methods for letting coding agents such as Claude Code, Codex, Gemini CLI, and Cursor use a logged-in Chrome: Claude subscribers can use the Claude in Chrome extension via claude --chrome.

Why it's recommended:The author reviews, one coding agent at a time, the approaches and configurations for connecting to a logged-in Chrome, and explains why copying cookies does not work. The methods can be directly adapted to one's own toolchain.

Load more

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.