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 00:42 (UTC+8) · refreshed hourly

10/15 · Thursday6 items

Databricks:BlogTutorialsAI Score61

Databricks explains how to cut Qwen 3.8 27B fine-tuning time by 6x on AI Runtime.

Databricks demonstrates how systematic performance optimization reduced Qwen 3.8 27B fine-tuning time on AI Runtime from 7 hours and 44 minutes to 1 hour and 17 minutes, a 6x speedup.

Why it's recommended:The article details the full optimization process, from identifying bottlenecks to using fused kernels and data packing, along with quantitative results that readers can apply to their own LLM fine-tuning tasks.

OpenAI Developers:BlogTutorialsAI Score75

OpenAI Engineer Shares Practices and Prompts for Cleaning Up Codebases with GPT-6 Models

An engineer on OpenAI's Core Experimentation team shares practical experience using GPT-6 Astra and GPT-6.1 Sol in Codex for code maintenance. The team has merged more than 1,400 maintenance PRs and deleted over 800,000 lines of code.

Why it's recommended:Drawing on firsthand experience, the author provides ready-to-use Codex prompts and model cost comparisons, allowing readers to start with small tasks and adapt the approach to maintaining their own codebases.

Lydia Hallie ✨@lydiahallieTutorialsAI Score69

Claude Code's idle compaction resumes from a summary; idleCompaction lets you disable it and customize the threshold

Claude Code's idle compaction makes Claude resume from a summary rather than the full context. The author explains that it can be disabled by setting "idleCompaction": false in settings., and that CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS can customize the threshold, which defaults to 200k.

Why it's recommended:The original article provides specific parameters for disabling Claude Code's idle compaction and instructions for setting its threshold, which readers can apply directly to their own settings. configuration.

LangChain:BlogTutorialsAI Score63

LangChain explains the context engineering of Deep Agents in detail: how to remember information without filling the context window

LangChain published a blog post explaining how its open-source agent framework Deep Agents uses middleware for context engineering to address context rot in long-running tasks.

Why it's recommended:Using a competitive analysis agent as an example, LangChain explains how middleware implements five context management techniques: offloading, summarization, skills, subagents, and long-term memory.

10/14 · Wednesday2 items

Google Developers BlogTutorialsAI Score70

Google uses Gemini 4 Argon to optimize ADK's 200,000-line Python codebase without supervision

A Google engineer ran Gemini 4 Argon unattended on google/adk-python for 11 hours. It proposed 6 optimizations, of which 2 were merged after review, reducing in-memory session preparation time for long conversations by about 40% and speeding up package imports by 21%.

Why it's recommended:The author reviews the complete workflow for letting a model work unattended on a real, large repository, including rejected changes and reusable .agents/skills setup instructions.

Anthropic:EngineeringTutorialsAI Score72

Anthropic develops its own version-control system, Scion, and completes the migration of its monorepo from Git in five months

Anthropic published an engineering retrospective stating that, because Claude increased the average daily number of merged code changes from about 2,500 in May 2026 to over 30,000 in September, it developed its own version-control system, Scion, in five months and completely replaced git, while also launching the code-review tool Antique.

Why it's recommended:A comprehensive retrospective on Anthropic's development of its own version control system, Scion, covering design principles for agent-scale workloads, architectural details, and measured benefits that can be applied to other large repositories.

Also reported by 2 other sources

10/13 · Tuesday2 items

LangChain:BlogTutorialsAI Score61

LangChain released the langchain-scaledown integration, using ScaleDown to constrain decoding and manage long-running agent contexts

LangChain released the langchain-scaledown integration and added support for ScaleDown's constrained decoding models (CDM) in LangSmith LLM Gateway to reduce the cost and latency of context management for long-running agents.

Why it's recommended:The article provides specific instructions for configuring ScaleDown's constrained decoding models with LangChain and Deep Agents middleware, which readers can adapt to compress agent context for their own use cases.

Vapi Products and Agent MethodsTutorialsAI Score67

How Vapi's in-house code review bot cuts PR merge times in half

Vapi engineer Lat Ware developed Pedantic Review Bot, reducing the average time from PR opening to merging from more than 130 hours to 65 hours. The bot runs tests and small reproductions to validate findings; analysis and comment posting are handled by two separate jobs, and it can comment but cannot approve or push. From September 2, 2026 to October 2, 2026, it reviewed 1,694 PRs and raised 363 blocker findings in 276 of them.

Why it's recommended:The author experimented with prompt versions, comparisons across multiple runs, and scope instructions on their own codebase, offering practical findings on code review agents that readers can apply directly.

10/12 · Monday2 items

Tessl: Product and Engineering BlogTutorialsAI Score65

How to constrain coding agents using only files you own

A Tessl blog post explains how four controllable file-based levers constrain coding agents: CLAUDE.md instructions, path-level rules in .claude/rules/, capability configurations under .claude/, and hooks in .claude/hooks/.

Why it's recommended:The author tested the same ticket three times to compare the effects of the control mechanisms in each charter file. The method can be applied to readers' own repositories and identifies what machines can check and what requires human review.

MeshyAI@MeshyAITutorialsAI Score70

A seven-step tutorial for building a game with Claude Opus 5.5, Blender, Unreal Engine, and Meshy

A tutorial introduces a method for building games with Claude Opus 5.5, Claude Code, Blender, Unreal Engine 5, and Meshy. The core idea is to assign each tool a single responsibility: connect Blender and Meshy through MCP, have Meshy generate props and characters, Blender clean them up and place them, Unreal handle lighting, physics, and the game itself, and Opus coordinate all three in a single conversation. The workflow begins by asking Opus to brainstorm 5 game ideas and then develop one, while providing the GPU specifications so it can push the visual performance to the limit. The first complete build took about 4 hours and approximately 19% of the weekly Max plan usage; subsequent builds will be much faster because assets can be reused.

Why it's recommended:The article lays out a seven-step workflow showing how to divide up the tools, along with actual time and cost figures, so readers can directly replicate the game-building workflow.

10/09 · Friday5 items

Hugging Face Official Team BlogTutorialsAI Score72

Hugging Face engineer uses ML Intern to build 7 small models for about $103

Hugging Face engineer yuvraj sharma used HuggingChat's ML Intern mode to build 7 models in just a few days, including a 0.8B prompt rewriter that runs on CPUs and produces 99.7% valid outputs, and a fine-tuned Qwen3.5-2B model that improved citrus disease identification accuracy from 14.9% to 52.8%. Total compute costs were approximately $103.

Why it's recommended:The author summarizes prompt-writing techniques for budget controls, baseline evaluations, and smoke tests, enabling an agent to independently handle data construction, training, and release. The methods can be readily adapted and reused.

Comfy BlogTutorialsAI Score66

How to Generate Real-Time Video with MiniMax H3 on a Single RTX 5090

The ComfyUI blog author used an agent to gather various optimizations and get MiniMax H3 real-time video generation running on a single RTX 5090 (32 GB of VRAM). Generating a 30-second video took 23.51 seconds, achieving 1.28x real-time throughput.

Why it's recommended:The article provides a complete optimization checklist and open-source nodes for real-time livestream video generation on a single RTX 5090, so readers can reproduce this cost-cutting approach.

NVIDIA Technical Blog:Agentic AI / Generative AITutorialsAI Score62

NVIDIA KGMON Team Shares Lessons from Building a Data Analysis Agent for KDD Cup 2026

The NVIDIA KGMON team placed second in the KDD Cup 2026 Data Agents competition and published a retrospective on methods for building reliable data analysis agents.

Why it's recommended:The NVIDIA KGMON team reviews nine practices from the KDD Cup 2026 data agent competition and shares transferable methods for building a verifiable agent harness around a fixed small model.

LangChain:BlogTutorialsAI Score61

LangChain Builds Restock, a Payment-Capable Agent, with Stripe Link and Managed Deep Agents

LangChain released Restock, a sample project featuring an office-supplies purchasing agent that runs in Slack using Managed Deep Agents. It demonstrates how an agent can complete payments securely.

Why it's recommended:The article provides a complete architecture for agent payments and credential isolation practices that readers can reuse for any agent scenario involving spending money.

Anthropic: Claude.dev Developer BlogTutorialsAI Score68

Anthropic’s practical guide to building scheduled agent automations with Claude Managed Agents

Anthropic released a reference implementation of a daily briefing based on Claude Managed Agents (beta). It reads Slack and GitHub sources on a schedule and posts briefings to Slack.

Why it's recommended:The article breaks scheduled agent automation into six reusable components and provides practical safeguards for bookmarks, ledgers, and permissions that can be directly adapted.

Also reported by 1 other source

10/08 · Thursday4 items

Tessl: Product and Engineering BlogTutorialsAI Score68

Tessl Engineering Blog: AI Isn’t Stupid, It’s Blind—Three Key Design Decisions for Enterprise Agent Memory

Drawing on a year of experience building a memory system for agents working across companies, the author argues that AI often fails not because its reasoning is poor, but because it cannot see critical decision-making context. Most agent memory systems just give a blind person a bigger filing cabinet.

Why it's recommended:Using firsthand measurements from a system they built, the author explains why memory should be treated as a governed record rather than storage. The three design decisions and cost figures can be directly adapted.

vLLM Official BlogTutorialsAI Score63

vLLM explains DeepSeek-V4.1-Flash optimizations in detail: 5x higher throughput for agent scenarios

Inferact and the vLLM community completed optimizations within three weeks of the DeepSeek-V4.1-Flash release, increasing speed by 1.9x at low concurrency and throughput by 5.3x under a 150 TPS constraint.

Why it's recommended:The vLLM team breaks down SWA bounded replay and a series of kernel optimizations, explaining exactly where the fivefold throughput increase comes from. The methods can be reused in other agent-serving scenarios.

Lydia Hallie ✨@lydiahallieTutorialsAI Score68

In Claude Code, you can set Haiku 5.5's autocompact window to 100K to stay in a lower pricing tier

Lydia Hallie explains that API-billed users can set Haiku 5.5's autocompact window to 100K to stay in the cheaper token pricing tier. The setting is saved per model and applies only to Haiku, including subagents. The commands are /model haiku and /autocompact 100k. Pricing cited from @ClaudeDevs shows that below 100K tokens, rates are $0.10/$0.50 per million (cache reads $0.01); above that, they are $0.50/$2.50 (cache reads $0.05).

Why it's recommended:The article provides the exact commands and threshold so API-billed users can stay in a lower token pricing tier by configuring the autocompact window.

Unsloth AI@UnslothAITutorialsAI Score66

Unsloth Open-Source Tutorial: Train a Qwen3.5 0.8B Decision Model Locally, Improving Accuracy from 20.7% to 74.3%

Unsloth released a tutorial and open-source repository that can fine-tune LLMs such as Qwen3.8 and Gemma 4 into decision models that output probabilities for options. Qwen3.5 0.8B achieved a combined accuracy increase from 20.7% to 74.3% on 3 decision-making benchmarks, using only 4GB of VRAM.

Why it's recommended:The original article provides a complete fine-tuning method and an accuracy comparison before and after. Readers can reproduce the process of turning an LLM into a decision model on a GPU with 4GB of VRAM.

10/07 · Wednesday3 items

Google@GoogleTutorialsAI Score68

Google opens the SynthID Detector portal, which can detect whether images, videos, and audio were generated by AI

Google describes how to check whether media was generated by AI: visit https://synthid.com and upload an image, video, or audio file. The portal scans it for SynthID watermarks from Google or its partners.

Why it's recommended:The official source provides a specific tool and steps for detecting AI-generated media, so readers can follow the process to upload a file and check whether it contains a SynthID watermark.

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.