AI News

2026-08-25 · TuesdayPublished daily at 08:00

AI Daily

Issue 126252026-08Tuesday
14major stories11sources13firsthand announcements1new modelsApproximately 6 minutes to read
Today's Highlights
  1. 1
  2. 2
  3. 3
In This Issue
  1. 01Model Releases/Updates1 item
  2. 02Product Releases/Updates5 items
  3. 03Industry News2 items
  4. 04Paper research1 item
  5. 05Tips and perspectives5 items

01Model Releases/Updates

OpenAI: Website updatesFirsthand

GPT-5.6 comes to Kiro, offering developers better value for money

The GPT-5.6 model family is now available in the software development agent Kiro, with three models: Sol, Terra, and Luna. In Terminal-Bench 2.1 testing, GPT-5.6 Terra reduced task completion costs in Kiro by approximately 82%. Optimized through a collaboration between OpenAI and AWS, the update aims to help developers produce higher-quality code with fewer iterations and more value per token.

Original text

02Product Releases/Updates

NVIDIA BlogFirsthand

NVIDIA Vera Rubin NVL72 Sets a New Standard for AI Agent Efficiency: Up to 30 Times More Work per Watt

NVIDIA's measurements show that, for agent workloads, Vera Rubin NVL72 delivers up to 30 times the throughput per megawatt of GB300 NVL72 and reduces the cost per million tokens by a factor of 35.

Original text
Meta Engineering BlogFirsthand

MetaRoCE: A New RDMA Transport Protocol Built for Ethernet at AI Scale

Meta designed and open-sourced MetaRoCE, an RDMA transport protocol built specifically for AI workloads on standard Ethernet, releasing its specification, reference software implementation, and compliance test suite through the Open Compute Project (OCP). The protocol moves intelligence to the endpoints and natively supports out-of-order delivery, multipath transmission, tolerance for lossless network conditions, and bidirectional congestion control. It requires no PFC and can deliver high throughput and low tail latency at a scale of a million GPUs. Existing RDMA Verbs APIs and software stacks can run without modification.

Original text
NVIDIA BlogFirsthand

How NVIDIA Uses NVLink Fusion to Integrate Custom XPUs into World-Class AI Factories

NVIDIA introduced NVLink Fusion to connect custom XPUs to its NVLink scale-up domain, delivering 3 times lower end-to-end latency between XPUs and 10 times higher packet rates than solutions based on off-the-shelf Ethernet.

Original text
Meta Engineering BlogFirsthand

MTIA 300: Meta's first training chip with an integrated NIC and communication offload engine

Meta releases MTIA 300, the first chip in its in-house accelerator family optimized for training recommendation and ranking models. The chip package integrates 12 800 Gbps RDMA NICs, provides 1.2 TB/s of total I/O bandwidth, and offloads communication through 16 dedicated message engines, keeping computational throughput loss below 0.5% when large-scale GEMM and collective communication run concurrently.

Original text
NVIDIA BlogFirsthand

NVIDIA Expands Vera Rubin Inference Capabilities as Groq 3 LPX Enters Full Production to Support Agent Systems

NVIDIA announced expanded fast token generation capabilities for Vera Rubin NVL72 to support agent systems. Its rack-scale NVIDIA Groq 3 LPX system is now in full production, aiming to define the next generation of AI inference through coordination across every layer of the AI factory, rather than relying on breakthroughs in a single chip, network, or system.

Original text

03Industry News

Mistral AI:NewsFirsthand

Mistral and HUMAIN Form Strategic Partnership to Advance Sovereign AI in Saudi Arabia and the Middle East

Mistral and HUMAIN announced a strategic partnership spanning AI infrastructure, advanced model development, and AI solution deployment, focused on Saudi Arabia and the wider region. The two companies will jointly develop localized AI models, initially focusing on cybersecurity and voice, with plans to develop frontier models with strong Arabic capabilities. The partnership is worth hundreds of millions of euros. Mistral will explore using HUMAIN's data center infrastructure, and the companies plan to develop a joint go-to-market strategy for regulated industries in Saudi Arabia.

Original text
LangChain:BlogFirsthand

How Toyota North America Scales Enterprise AI with Deep Agents and LangSmith

Toyota North America uses Deep Agents and LangSmith to run more than 50 AI agents in production, cutting delivery cycles from 6 months to 4 days. The article shows how it uses LangSmith to track AI return on investment (ROI) and support enterprise AI implementation at scale.

Original text

04Paper research

Apple Machine Learning ResearchFirsthand

Beyond Visual CoT: Internalized Visual Thinking Enables Active Video Reasoning

Multimodal large language models often use visual chain-of-thought (Visual CoT) for reasoning about space, time, and embodied environments, but generating intermediate reasoning images adds substantial inference overhead. The newly proposed post-training framework Internalized Visual Thinking (IVT) internalizes visual thinking during training and directly performs text prediction and optimization at inference time, enabling active video reasoning without increasing inference costs.

Original text

05Tips and perspectives

TechCrunch:AI

OpenAI Is Building AI Agents for Everything, but Will Users Be Willing to Hand Over Control?

OpenAI introduced ChatGPT Work, transforming Codex into an agent product for non-engineers, available with subscriptions starting at $20 per month. It aims to let white-collar workers use LLMs to complete multistep tasks autonomously. In June, 98% of OpenAI employees used Codex, compared with just 17% of organizational subscribers and fewer than 1% of individual subscribers. The company is simplifying the interface to expand adoption and support its substantial investment in training.

Original text
Google Developers BlogFirsthand

How ADK Evaluates Real-Time Voice Agents

Google has brought native real-time evaluation capabilities to ADK. It can use simulated users to drive real-time voice agents with audio, score their voice responses, and complete the process in the same evaluation loop used for text agents. The example uses a workflow composed of three agents based on gemini-live-2.5-flash-native-audio, supports both conversational scenarios and fixed-dialogue test cases, includes user personas such as NOVICE, and uses max_allowed_invocations to limit the total number of turns.

Original text
Databricks:BlogFirsthand

How Databricks runs, debugs, and scales workloads from a local IDE

Databricks workspaces are built specifically for data analytics and data engineering and now support running, debugging, and scaling workloads directly from a local IDE. The feature is designed to bridge the gap between local development environments and the cloud-based Databricks platform, allowing developers to code, test, and deploy without switching contexts.

Original text
GitHub BlogFirsthand

Passing an automated check for your alt text does not mean it is actually acceptable

The WebAIM Million report shows that 16.2% of images on mainstream webpages lack alt text, while another 10.8% have vague or repetitive alt text. GitHub built an alt-text plugin for Accessibility Scanner that uses five deterministic rules to detect missing alt text, filenames, placeholders, generic terms, and adjacent repetition. It also introduced layout-based duplicate detection and optional visual-model review to balance the false-positive rate with practical usefulness.

Original text
Claude:BlogFirsthand

How an Anthropic Marketer Uses Claude Code to Send Personalized Weekly Updates to Every Sales Rep

Anthropic marketing team member Adam Ward shares how he uses Claude Code to turn weekly sales reports into personalized Monday briefings for each account executive. He connects BigQuery and CRM data through MCP and iterates on prompts based on feedback from sales reps and managers, adding nine content rules, including “never fabricate URLs.” The briefing has now been rolled out to all sales teams, automatically sending three priority action items and account-related activities every Monday.

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.