AI News

2026-08-26 · WednesdayPublished daily at 08:00

AI Daily

Issue 127262026-08Wednesday
19major stories12sources16firsthand announcements1new modelsAbout a 7-minute read
Today's Highlights
  1. 1
  2. 2
  3. 3
In This Issue
  1. 01Model Releases/Updates1 item
  2. 02Product Releases/Updates8 items
  3. 03Industry News2 items
  4. 04Paper research2 items
  5. 05Tips and perspectives6 items

01Model Releases/Updates

Google AI

WeatherNext forecasts cyclones: warns of Category 5 hurricanes five days in advance

Google AI released WeatherNext, a tropical cyclone forecasting model that can simultaneously predict a storm's track, intensity, and size, providing a full additional day of warning time compared with existing systems. In live testing during the 2025 hurricane season, the model predicted Hurricane Melissa's Category 5 landfall in Jamaica five days in advance, marking the first time the U.S. National Hurricane Center used an AI model in real time. The model can generate up to 1,000 simulations for a single storm, and its code and weights have been released as open source.

Original text

02Product Releases/Updates

Andrew Ng

New version of OpenWorker released with a built-in cybersecurity agent

OpenWorker, the open-source agent from Andrew Ng, releases a new version with stronger security workflows. Its harness is fully open source, allowing security teams to audit it for backdoors. The new version includes three types of cybersecurity agents for scanning code vulnerabilities, detecting dependency supply chain injections, and checking cloud security configurations. It also supports running open-weight models locally to protect sensitive code.

Original text
Claude:BlogFirsthand

Claude's memory feature now fully connects chats and Cowork, allowing users to view and edit entries individually

Starting today, Claude is unifying memories across chats and Claude Cowork, so users can draw on previously accumulated context in either setting and avoid repeating explanations. Memories update in real time during chats, and users can view, edit, or delete each memory by topic in Memory settings. Sensitive topics such as health and religion are not stored by default, but this can be enabled in settings. Sensitive identification numbers, criminal records, and similar information will never be saved.

Original text
Apple:NewsroomFirsthand

Apple introduces a new Mac Studio powered by the M5 Max and M5 Ultra

Apple released the Mac Studio with the M5 Max and all-new M5 Ultra, delivering up to 4.3 times greater AI performance, 1.8 times greater graphics performance, and twice the storage speed. The M5 Ultra version supports up to 512GB of unified memory and 1.2TB/s of memory bandwidth, enabling large LLMs to run entirely on the device; a four-unit cluster can deliver up to a 3x improvement in distributed AI inference speed. The new products are available for preorder today and go on sale September 22.

Original text
LangChain:BlogFirsthand

LangChain integrates with Airbyte: making data ingestion production-ready

The LangChain and Airbyte integration is designed to scale retrieval applications into production. It automates data ingestion through scheduling, text splitting, and more than 50 embedding models, helping developers build production-grade data pipelines that can scale.

Original text
Apple:NewsroomFirsthand

Apple releases a Mac mini powered by the new M6 and M5 Pro chips

Apple introduced a new Mac mini with the all-new M6 and M5 Pro chips. The M6 version delivers up to 4 times greater AI performance and 40% greater CPU performance, and supports Wi-Fi 7 and Bluetooth 6.

Original text
Apple:NewsroomFirsthand

Apple unveils M6 and M5 Ultra, delivering major gains in performance and AI computing power

Apple introduces its first 2nm chip, M6, with a 12-core CPU, 12-core GPU, and dual 16-core Neural Engines. It delivers up to 170GB/s of unified memory bandwidth and 1.2x the multithreaded performance of M5. M5 Ultra uses the first four-chip packaging architecture, with up to a 36-core CPU and 80-core GPU, and bandwidth of 1.2TB/s—50% higher than M3 Ultra. The two chips will power the new Mac mini and Mac Studio, respectively.

Original text
OpenAI: Website updatesFirsthand

OpenAI introduces Admin plugins for ChatGPT Work and Codex

OpenAI introduced an Admin plugin for ChatGPT Work and Codex, allowing administrators to view workspace activity, manage members and permissions, adjust usage limits, and approve spending requests in a single conversation. The plugin follows users' existing roles and permissions without expanding access, and can route pending usage requests to Slack or Microsoft Teams for approval. Workflows deployed by OpenAI's IT team have resolved approximately 45% of support tickets.

Original text
OpenAI: Website updatesFirsthand

OpenAI debuts its in-house inference chip Jalapeño: industry-leading speed and energy efficiency

OpenAI releases its in-house inference chip, Jalapeño, designed to deliver faster, more energy-efficient AI inference for modern models, with higher throughput and lower latency. Initial results show industry-leading inference speed and efficiency.

Original text

03Industry News

Anthropic:NewsroomFirsthand

Anthropic launches a $5 million grant program to support independent evaluations of AI's impact on user well-being

Anthropic launched a $5 million grant program to provide direct funding, model access, and technical support for independent research into how AI affects user well-being. Grantees will work fully independently and publish their results as open-source projects for any developer to use. Applications close on September 21, and selected applicants will be notified by October 5.

Original text
OpenAI: Website updatesFirsthand

OpenAI bans accounts linked to a Russian covert influence operation

OpenAI bans a group of Russia-originating accounts that used AI to promote a fictional Israeli think tank and a “sovereignty” index, with content praising Russia and criticizing the West.

Original text

04Paper research

Google Research:BlogFirsthand

AgentHands: Generating interactive hand gestures for XR spatial conversational agents

At CHI 2026, Google presented AgentHands, a research prototype that uses an LLM to generate expressive hand gestures synchronized with speech for XR conversational agents, providing spatially anchored physical guidance. The system registers objects through eye tracking and scene reconstruction, has the LLM generate inline GestureEvents, and coordinates TTS with the animation engine on the headset using word-level timestamps to achieve precise synchronization between gestures and speech. Applications include interactive orchid-care instruction, technical guidance for operating 3D printers, and lifestyle companionship.

Original text
Apple Machine Learning ResearchFirsthand

STARFlow2: Bridging language models with normalizing flows for unified multimodal generation

STARFlow2 proposes unifying autoregressive normalizing flows with language models for multimodal generation. The approach observes that autoregressive normalizing flows and LLMs share causal masks, KV-cache mechanisms, and a left-to-right structure, bridging the architectural gap between text and image generation while avoiding the loss of visual fidelity caused by discrete tokenization.

Original text

05Tips and perspectives

OpenRouter:AnnouncementsFirsthand

How to choose the best AI model in real time in your editor

OpenRouter proposes a framework for choosing models: define the task, shortlist candidates using real-time usage and third-party benchmarks, compare providers’ pricing and latency, then test with your own prompts. The metric is “cost per completed task,” not “cost per token.” Its MCP server can query real-time rankings, prices, and benchmarks directly in editors such as Claude Code and Cursor, and route requests using openrouter/auto-beta.

Original text
OpenRouter:AnnouncementsFirsthand

OpenRouter video generation API: A code-first integration guide

OpenRouter launched a unified asynchronous video generation API. Tasks are submitted via POST /api/v1/videos, with status polling and MP4 downloads supported. It supports models such as Seedance, Veo, and Wan; switching models only requires changing the model identifier.

Original text
Dwarkesh Patel:Podcast & Blog

Dylan Patel: Anthropic and OpenAI will control most of the world's computing power by 2028

In the latest podcast episode, SemiAnalysis founder Dylan Patel and Dwarkesh Patel discuss the economics of AI labs. They expect Anthropic and OpenAI to control most of the world’s available FLOPs by 2028 because they can monetize compute more effectively and outbid others.

Original text
LangChain:BlogFirsthand

LangChain evaluates a CSV data question-answering system: Agents, retrieval, and LLM evaluation

LangChain released a benchmark evaluation for question-answering systems over CSV data, covering agent, retrieval, and LLM evaluation methods. The article provides benchmark results, debugging insights, and open-source code to help developers build better CSV data question-answering systems. The evaluation focuses on performance in practical applications and common issues.

Original text
LangChain:BlogFirsthand

LLMs and SQL: building a reliable text-to-SQL solution with LangChain

This article explores how to use large language models to query SQL databases in natural language, focusing on technical approaches to reducing model hallucinations and building reliable text-to-SQL solutions. Based on the LangChain framework, it covers key implementation paths for converting natural language into SQL queries.

Original text
Google Blog:AIFirsthand

5 ways Google Search is upgrading home decor

Google Search introduced five practical home-decor features: AI Mode can upload a room photo and generate furniture-layout visualizations; Lens can identify vintage items and find similar products; Circle to Search lets users circle desired decor without switching apps; Search Live provides step-by-step DIY installation guidance; and product listings can compare prices and track price-drop alerts. Searches for "home decor inspo" have increased 300% in recent months.

Original text

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.