Aug 25 · AI Daily

Claude has unified memory between chats and Claude Cowork, letting people pull in past context in either setting. Memories update in real time as you talk, and you can view, edit, or delete individual memories by topic in the Memory settings. Sensitive topics like health and faith aren't stored by default but can be turned on manually, while sensitive ID numbers and criminal records are never saved.

Apple has released the Mac Studio featuring the M5 Max and the new M5 Ultra, delivering up to 4.3 times higher AI performance, 1.8 times better graphics performance, and 2 times faster storage speeds. The M5 Ultra supports up to 512GB of unified memory with 1.2TB/s of memory bandwidth, making it capable of running large LLMs entirely on device. A four-unit cluster can achieve up to 3 times faster distributed AI inference speeds. Pre-orders start today, and it goes on sale on September 22.

Apple has introduced its first 2nm chip, the M6, featuring a 12-core CPU, a 12-core GPU, and dual 16-core neural engines, with a unified memory bandwidth of up to 170GB/s and 1.2 times higher multi-threaded performance than the M5. The M5 Ultra uses the first quad-die packaging architecture with up to 36 CPU cores, 80 GPU cores, and 1.2TB/s of bandwidth—a 50% increase over the M3 Ultra. They power the new Mac mini and Mac Studio respectively.

OpenWorker, the open-source agent project led by Andrew Ng, has released a new version featuring a fully open-source harness that security teams can audit for backdoors. The update includes three types of cybersecurity agents for code vulnerability scanning, dependency supply chain injection detection, and cloud security configuration checks, while supporting local execution of open-weight models to keep sensitive code on-premise.

Replit CEO Amjad Masad shared that Replit Agent has completely replaced Claude Cowork as his daily work tool, citing its greater persistence and rigor, and its ability to more effectively leverage code and software capabilities to finish tasks.

Box CEO Aaron Levie pointed out that AI agents will perform 100 times more work than humans on systems of record—querying data, handling tasks, and running workflows—making governance, reliability, security, and access control more important than ever. He noted that the recent OpenAI and Hugging Face incidents are just a glimpse of the risks in the agent era, and systems that can support agents are facing massive disruption opportunities.

An article analyzes how long-running AI agents suffer from attention decay due to accumulated context, turn temporary instructions into permanent rules, and introduce security risks. It suggests giving daily assistants a 24-hour lifespan with midnight session resets, delegating specific tasks to single-use sub-agents that live for only 30 seconds, and writing persistent preferences to disk before resets to avoid long-term memory degradation across sessions while narrowing the window of exposed permissions.

OpenRouter has proposed a model selection framework: first define the task, filter candidate models using real-world usage and third-party benchmarks, compare pricing and latency across providers, and finally run tests with your own prompts. The core idea is to replace 'cost per token' with 'cost per task completed' as the judging standard. It also offers an MCP server to check real-time rankings, prices, and benchmarks directly in editors like Claude Code and Cursor, alongside openrouter/auto-beta for automatic routing per request.

OpenRouter has launched a unified asynchronous video generation API. Users can submit tasks via POST /api/v1/videos, poll for status, and download MP4 files, with support for multiple models including Seedance, Veo, and Wan. Switching models only requires changing the model identifier. The guide also covers production essentials like idempotency, webhook notifications, and concurrency control.

LangChain and Airbyte have rolled out an integration designed to scale retrieval-augmented applications from prototypes to production. It automates data ingestion through scheduling, text splitting, and support for over 50 embedding models, helping developers build scalable production data pipelines and cut down on manual script maintenance costs.

SemiAnalysis founder Dylan Patel predicted on the Dwarkesh Patel podcast that Anthropic and OpenAI will control most available FLOPs globally by 2028, because they are better at monetizing compute, allowing them to outbid other competitors in hardware procurement and reflecting how lab economics are driving compute centralization.

OpenAI employee Thibault Sottiaux announced that starting tomorrow, the 5-hour usage limit will be re-enabled for ChatGPT Work and Codex Plus accounts. The limit helps smooth out compute loads, maintain generous weekly allowances, and prevent new users from accidentally burning through their weekly quota and getting confused. The Pro $100 and $200 tiers will not have this restriction applied in the coming months.

Thibault Sottiaux, head of Codex and ChatGPT at OpenAI, stated that DevDay 2026 will be the best DevDay in company history by a wide margin, hinting at major product announcements ahead.

Every CEO Dan Shipper reported that his company's YouTube channel was disabled by Google without notice or explanation, and he is publicly seeking help to restore the account, highlighting the risks of unilateral platform actions against creator accounts.

Google AI has released WeatherNext, a cyclone prediction model that forecasts storm paths, intensity, and size simultaneously, giving a full extra day of warning over existing systems. During the 2025 hurricane season, it predicted Hurricane Melissa's Category 5 landfall in Jamaica five days in advance, marking the first time the National Hurricane Center used an AI model in real time. A single storm can generate thousands of simulations, and the code and weights have been open-sourced.

The new FORGE benchmark tested 12 commercial and open-source LLMs across 225 real products, 15 categories, and 5 consumption scenarios, finding that all models are vulnerable to GEO content poisoning attacks. A single poisoned page caused up to a 27% deception rate, which jumped to 73.8% when replacing all top 3 search results, exposing the widespread fragility of retrieval-augmented recommendation systems.