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 11/17 00:23 (UTC+8) · refreshed hourly

10/24 · Saturday1 item

Figma AI / Agent / MakeTutorialsAI Score70

Figma code layers enters open beta: explore designs collaboratively with code layers and Figma agent on the canvas

Figma announced that code layers has entered open beta for 100% of users, allowing teams to create and edit code-based interactive prototypes directly on the canvas.

Why it's recommended:The original article uses a complete example to demonstrate how to collaborate after the open testing of code layers, enabling designers to assess whether to move prototype exploration back to the canvas.

10/23 · Friday1 item

Claude:BlogTutorialsAI Score62

Anthropic guide: When to merge subagents into the main Agent loop

Anthropic's Applied AI team published a guide explaining when to merge the router, planner, and domain-specific sub-agents under an orchestrator into a single main agent loop. The article's comparison shows that a single agent with skills often outperforms an all-in-one prompt and sub-agent designs in quality, while also costing less and providing lower latency. One customer saw runtime become approximately 2x faster after switching to a single loop.

Why it's recommended:Based on comparisons across multiple enterprise deployments, Anthropic summarizes when to merge subagents into the main loop and provides a migration kit, evaluation set, and phased merging method that can be directly applied to decisions about one's own agent architecture.

Also reported by 1 other source

10/21 · Wednesday1 item

Tessl: Product and Engineering BlogTutorialsAI Score61

The Tessl team shares: Stop reviewing AI-generated code line by line; validate intent instead.

Tessl engineer Ankit Jain shared how the team responded after its PR volume quadrupled and nearly all code was AI-generated: it abandoned line-by-line diff reviews and instead began reviewing intent before code generation, then verifying acceptance criteria one by one. Key practices include maintaining offer-skill to help understand code changes and turning the most frequently recurring review comments into invariant rules executed automatically by machines. The author expects the review burden to shrink significantly starting in the second month.

Why it's recommended:After the author's team's PR volume quadrupled, they abandoned line-by-line code reviews and switched to reviewing intent in advance plus verifiable criteria, providing practices and time expectations that can be directly adopted.

10/20 · Tuesday3 items

Google Developers BlogTutorialsAI Score70

Google Releases TorchTitan TPU Fork and Provides a Model Optimization Playbook

Google released a TorchTitan TPU fork (github.com/google-pytorch/torchtitan), which, together with the native TorchTPU backend, trains PyTorch models on Cloud TPU and provides SimpleFSDP, Graph Trainer, Pallas kernels, and an optimization Playbook.

Why it's recommended:The original article provides a complete TPU optimization sequence covering benchmarks, compilation, kernels, and memory, along with measured figures for FLUX.1, Qwen3 235B, and others. The approach can be transferred to one's own training workflow.

Official account: Mianbi IntelligenceTutorialsAI Score68

ModelBest breaks down how the forward-looking expert scheduler LES enables 35B MoE to run in real time on mobile edge devices without loss

The MiniCPM-V team at ModelBest introduced the Lookahead Expert Scheduling (LES) technology supporting on-device operation of MiniCPM-V 4.7, claiming to be the first to enable a 35B-A3B-scale MoE to run on-device in real time without loss. Tests on an iPhone 18 Pro Max achieved decoding speeds above 20 tokens/s.

Why it's recommended:The original article breaks down LES's specific design across scheduling, storage, and compute, and provides ablation data. The method can be transferred to edge deployment of other MoE models.

Google Developers BlogTutorialsAI Score65

Google ADK Tutorial: Building Voice Agents That Never Run Out of Conversation with Streaming Tools and Response Scheduling

Google Developers Blog published a tutorial explaining how to use the open-source ADK and gemini-3.8-live to build voice agents with no awkward silences, with non-blocking tool calls at the core.

Why it's recommended:The official Google tutorial presents two modes and measured figures for non-blocking ADK voice tools, which can be adapted for developing your own Gemini Live voice agents.

10/19 · Monday5 items

LangChain:BlogTutorialsAI Score61

How Numan built the evaluation-first AI health coach Nu with LangGraph and LangSmith

The UK digital health company Numan reviews the development of its AI health coach Nu's evaluation system. The core conclusion is that the scarce asset is not prompts but evaluation standards. Single-turn static datasets and scripted multi-turn conversations both struggle to cover real-world conversations, so Nu switched to a topic-based dataset scored by category, a simulated conversation layer in which a less expensive model plays the patient, and automatic safety regression checks that block releases.

Why it's recommended:Numan provides a firsthand retrospective on building an evaluation system for an AI health coach. Practices such as using a fixed dataset, simulated conversations, and clinical-doctor sign-off criteria can be transferred to other regulated scenarios.

Meta Engineering BlogTutorialsAI Score70

Meta explains the Thin Orchestrator pattern in detail: Make routing do less to improve AI agent reliability.

Meta's engineering team describes the thin orchestrator pattern used by its data-pipeline agents: an orchestrator of approximately 1,300 lines is responsible only for routing and reliability checks, dispatching more than 20 workflows to 14 specialized subagents with isolated contexts, while performing no domain reasoning itself.

Why it's recommended:Meta's engineering team presents a directly transferable architecture for a thin orchestrator, using fixed routing protocols and isolated sub-agents to reduce the hallucination risk of production agents.

n8n BlogTutorialsAI Score66

n8n Tutorial: Run a Complete AI Workflow Offline on NVIDIA DGX Spark

The n8n blog published a tutorial explaining how to deploy seven local models on a single NVIDIA DGX Spark (128GB unified memory), use LiteLLM’s single OpenAI-compatible endpoint to drive n8n workflows, and avoid calling cloud APIs throughout.

Why it's recommended:The author benchmarked the entire local model stack on their own DGX Spark, providing data on memory allocation, license screening, and concurrent performance that can be transferred to private deployments for small teams.

Meta Engineering BlogTutorialsAI Score66

Meta Explains Its Gigawatt-Scale Prometheus Supercluster in Detail: Phase One Completed and Used to Train the Muse Model

Meta introduces Prometheus, its gigawatt-scale supercluster in New Albany, Ohio. Phase one is complete and being used to train the Muse series of foundation models, with a target expansion to approximately 1 million GPUs.

Why it's recommended:Meta officially breaks down the networking, thermal management, and reliability engineering of its Ohio supercluster, offering readers insight into the operational details of gigawatt-scale training clusters.

Browser UseTutorialsAI Score61

Browser Use explains how browser agents work and when to use them.

Browser Use published an educational article explaining that browser agents consist of three parts: a model, a harness, and a browser. They complete tasks through a loop of observing the page, selecting an action, executing it, and checking the result. Pages can be read through the DOM, accessibility tree, or screenshots.

Why it's recommended:The original article systematically breaks down the components, execution loop, and selection boundaries of browser agents, and uses its own cloud usage data to explain typical scenarios, cost risks, and how to get started.

10/18 · Sunday1 item

Addy Osmani@addyosmaniTutorialsAI Score69

Addy Osmani discusses making Claude Code self-validate its output: Code review must evolve into a multilayer verification system.

Addy Osmani shared the Anthropic team's code-review experience, noting that providing agents with verification methods can significantly improve output quality, and that Claude Code can now write more than 90% of its own code. He believes that when generation speed exceeds human review capacity, line-by-line manual review is no longer feasible. Teams should instead adopt a multilayer verification system consisting of end-to-end tests, property-based tests, and adversarial multi-agent review. Humans should focus on product intent, risk assessment, and evidence evaluation, with approvals tiered by risk.

Why it's recommended:Based on practices from the Anthropic team, the author presents concrete approaches for having agents prove their own work and replacing manual review with a multilayer verification system.

10/17 · Saturday1 item

WeChat Official Account: Digital Life KazikeTutorialsAI Score67

Kazike puts the 37999-yuan RTX Spark thin-and-light laptop to the test: a week to achieve local token freedom

Digital Life Kazike shares a week of in-depth hands-on experience with the ASUS ProArt Creator 16 (RTX Spark N1X chip, 128G unified memory, 37999 yuan). When it went on sale globally, the 128G version sold out on JD.com within minutes.

Why it's recommended:In a week of hands-on testing, the author explains how to deploy local models on an RTX Spark laptop with 128GB unified memory, and provides a reusable WorkBuddy expert integration workflow.

10/16 · Friday5 items

vLLM Official BlogTutorialsAI Score60

vLLM analyzes inference optimizations for Kimi K3 on AMD Instinct MI355X, with single-GPU throughput 24.4% higher than NVIDIA HGX B300

AMD and Embedded LLM explain in detail on the official vLLM blog how they optimized inference for Kimi K3 on eight AMD Instinct MI355X GPUs.

Why it's recommended:The original article breaks down the cache, MoE, kernel-fusion, and speculative-decoding optimization approaches behind more than 50 PRs, which readers can adapt to troubleshoot bottlenecks in their own inference stacks.

Anthropic: Claude.dev Developer BlogTutorialsAI Score76

How the Claude Code Team Conducts Code Reviews: Six Layers of Checks and Risk-Based Approval

The Anthropic Claude Code team explains how to organize code review when Claude writes about 90% of the code: Review shifts from a final gate to a continuous process. Every PR goes through six layers of checks: verification, required CI, AI review, security review, risk-based human approval, and a merge queue, supplemented by rapid rollback.

Why it's recommended:Anthropic's own team reviewed its Claude code-review process and provided directly transferable practices for implementation, such as layered checks and risk-tiered approvals, along with real convergence data.

ChatGPT@ChatGPTTutorialsAI Score73

ChatGPT demonstrates how to compare tabs, fill out forms, and draft emails in Chrome

The official ChatGPT account introduces various examples of using ChatGPT in Chrome. Examples include comparing the sizes, materials, delivery times, and return policies of three open sofas; interpreting an article's arguments and evidence in plain language; organizing supplier quotes and features from multiple tabs into a comparison table; filling out a registration form based on an event brief and stopping before submission; comparing contractor services and drafting an email; and filling out a dentist registration form in @chrome and adding the appointment to work and personal calendars.

Why it's recommended:The official guide lists specific examples of having ChatGPT compare tabs, fill out forms, and draft emails in Chrome, which readers can use as a reference when adapting these tasks to their own browsers.

Meta Engineering BlogTutorialsAI Score66

How Meta developed the Forum app using AI-native engineering practices

Meta's engineering team explains how it built Forum, an independent Groups application with AI agents as participants in the development process, using a human-on-the-loop approach and measuring success by verified production-grade output. Forum reached parity with Facebook Groups' core functionality within eight weeks. Autonomous stress-test loops discovered and fixed memory issues during development, roughly halving memory usage and page faults; after launch, the agents also helped with root-cause analysis of user feedback.

Why it's recommended:The original article reviews in detail how Meta used agents throughout the development process and the metrics it used; the transferable engineering lessons are relatively solid.

Apple Machine Learning ResearchTutorialsAI Score60

Apple demonstrates running an 816B large model locally on a Mac Studio cluster using MLX and RDMA

Apple's machine learning research team demonstrated how to run frontier models with nearly one trillion parameters, such as Muse Spark Open, locally on a Mac Studio, using the newly released version of MLX (v0.30.1 and later) and the RDMA over Thunderbolt driver in macOS 26.2.

Why it's recommended:Apple provided a parallelization approach and benchmark data for running an 816B MoE model on a Mac Studio cluster, which can be applied to local large-model deployment.

10/15 · Thursday6 items

Google Developers BlogTutorialsAI Score63

Google engineers explained how to systematically evaluate AI agents and move beyond vibe checks

The Google Developers Blog published a practical guide to agent evaluation, noting that a vibe check based on running a few prompts in a terminal cannot uncover hidden failures such as agent retry loops and hallucinated tool parameters.

Why it's recommended:The article provides a complete agent evaluation workflow, from defining CUJs to CI/CD gates. It can be reproduced directly with the agent-eval toolkit, avoiding four common evaluation pitfalls.

WeChat Official Account: Digital Life KazikeTutorialsAI Score77

A creator tests open-source Skill Anatomy: generate an animated explainer and narrated video with a single prompt

The creator Digital Life Kazike shared the open-source Skill Anatomy, built around the idea that "the object itself is the explanation." It can turn abstract principles into interactive mechanical model animations, then use them to generate educational videos with explanatory voiceovers.

Why it's recommended:The author tested how well the open-source Anatomy Skill generates educational videos explaining underlying principles and provided firsthand data on credit usage, time required, and model configurations to help readers decide whether to try it.

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.