Aug 26 · AI Daily

Claude has unified memory between chat and Claude Cowork, allowing users to call upon previously accumulated context during conversations in either scenario and reducing repeated explanations. Memories update in real time during chats, and users can view, edit, or delete each memory by topic in the Memory settings. Sensitive topics such as health and beliefs are not stored by default, and sensitive identification numbers and criminal records are never saved.

Apple has released the Mac Studio powered by the M5 Max and the new M5 Ultra, delivering up to 4.3x higher AI performance, 1.8x higher graphics performance, and 2x faster storage speeds. The M5 Ultra version supports up to 512GB of unified memory and 1.2TB/s of memory bandwidth, making it capable of running large LLMs entirely on-device. A four-device cluster can deliver up to 3x faster distributed AI inference speeds. Pre-orders begin today, with availability starting September 22.

Apple has introduced its first 2nm chip, the M6, featuring a 12-core CPU, 12-core GPU, and dual 16-core Neural Engines, with unified memory bandwidth reaching up to 170GB/s and 1.2x higher multi-threaded performance than the M5. The M5 Ultra uses the first quad-die packaging architecture, featuring up to a 36-core CPU and 80-core GPU with 1.2TB/s of bandwidth, a 50% increase over the M3 Ultra. These chips are used in the new Mac mini and Mac Studio, respectively.

LangChain and Airbyte have launched an integration solution to scale retrieval applications to production environments. The solution automates data ingestion through scheduling, text splitting, and over 50 embedding models, shifting updates from data sources to RAG indexes from manual scripts to configurable pipelines to help developers build scalable, production-grade data pipelines.

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

OpenAI employee Thibault Sottiaux posted on X that OpenAI DevDay 2026 will be the company's best DevDay in history by a wide margin, hinting at major upcoming product or feature announcements.

OpenAI employee Thibault Sottiaux announced that the previously paused 5-hour usage limits for Plus accounts on ChatGPT Work and Codex will be restored the next day. The limit is intended to smooth compute loads and prevent new users from accidentally exhausting their weekly allowances. Pro ($100/$200 tiers) subscriptions remain unaffected by this limit for the coming months.

Box CEO Aaron Levie suggested that as AI agents perform volumes of work on enterprise systems that could reach a hundred times that of humans, system governance, reliability, security, and access control will become more critical than ever. Using the incident between OpenAI and Hugging Face as an example, he highlighted the risks brought by agents autonomously executing goals and noted this presents a massive opportunity and disruption window for enterprise software.

OpenRouter has launched a unified asynchronous video generation API. Users can submit tasks via POST /api/v1/videos, poll status, and download MP4 files. Supporting models like Seedance, Veo, and Wan, switching models requires only changing the model identifier. The guide also covers production details such as idempotency, webhooks, and concurrency control.

OpenRouter proposed a model selection framework: first define the task, filter candidates using real-time usage and third-party benchmarks, then compare pricing and latency across providers, and finally test with your own prompts. The evaluation criterion is "cost per completed task" rather than "cost per token." Its MCP server can query real-time rankings, prices, and benchmarks directly inside editors like Claude Code and Cursor, and supports automatic model routing per request.

In the latest episode of the Dwarkesh Patel podcast, SemiAnalysis founder Dylan Patel discussed lab economics, projecting that Anthropic and OpenAI will control the vast majority of available global FLOPs by 2028 because these two companies can better monetize compute and are therefore willing to outbid other competitors for compute resources.

Google AI released the WeatherNext cyclone prediction model, which simultaneously predicts storm tracks, intensity, and size, providing a full extra day of warning compared to existing systems. The model predicted Category 5 landfall for Hurricane Melissa in Jamaica five days in advance during real-world testing in the 2025 hurricane season, marking the first time the National Hurricane Center used an AI model in real time. A single storm can generate up to 1,000 simulations, and the code and weights have been open-sourced.