☕️ Major AI models go down at once

• Claude, Grok, ChatGPT, and Gemini all hit outages simultaneously today, leaving developers who rely on frontier models scrambling as Anthropic, X, OpenAI, and Google services all faltered at roughly the same time.
• OpenAI reported problems across 15 of its services, while Anthropic's tracker showed Opus 4.8 and Opus 5 still degraded after Mythos, Fable, and other models recovered to baseline error rates.
• Cursor flagged degradation across all Grok models, automations, cloud agents, and review bots; Google's status page claimed Gemini was fine, though Down Detector logged hundreds of failure reports over several hours.

• OpenAI released GPT-6 Astra, a frontier model it pitches as "the world's best computer use model" that navigates browsers, spreadsheets, desktop apps and websites by acting like a human user rather than through purpose-built API integrations.
• On an offline subset of OSWorld 2.0, Astra scored ~73% while taking about 40 minutes per task, versus GPT-5.6 Sol's ~66% at roughly 75 minutes, cutting time per task by about 47%.
• Astra began rolling out yesterday to enterprise customers via the gated Daybreak program, with wider availability coming to ChatGPT Plus, Pro, Business, Enterprise, the OpenAI API, AWS Bedrock and Microsoft Azure over the following days.

• Nvidia unveiled PAIR (Personal AI Router), a free open-source tool announced at IFA 2026 that spreads AI agent workloads across idle PCs on a local network so subagents no longer compete for a single GPU.
• PAIR scans the network, routes task requests to whichever system is idle and has spare capacity, letting the main machine keep gaming or working while subagents split jobs like inbox triage across separate PCs.
• Currently in beta with graphical and terminal interfaces on Windows, macOS, and Linux, it supports GeForce RTX 20 Series and newer, RTX PRO Turing-and-later workstation cards, DGX Spark, and Apple M4 or newer.

• AWS cut GPU model loading for a 203 GB Llama-4-Scout to 25 seconds on EKS, down from 423 seconds, using configuration-only changes to how vLLM and SGLang stream weights from S3 rather than any code rewrite.
• Aligning download chunks to SafeTensors shard boundaries (~4 GB), matching concurrency to shard count, and killing stalled S3 connections after a few slow seconds fixed a calling pattern that had left 98% of bandwidth idle.
• Caching torch.compile output on local NVMe instead of ephemeral storage drops recompilation from 34-53 seconds to a 4-6 second cache hit, cutting warm-node pod restarts on scale-ups, rolling updates, and OOM recoveries from up to eight minutes to under 30 seconds.

• Meta has rolled out ZGateway, a stateless proxy tier sitting in front of ZippyDB, its most heavily used key-value store, to unify client traffic and add admission control, load balancing, and cross-region resilience.
• ZGateway now carries about 40% of ZippyDB traffic at over 1 billion operations per second, adding ~6% compute overhead while cutting total persistent connections roughly 19x by multiplexing many clients onto shared backend connections.
• A shared batcher coalesces requests across unrelated callers headed for the same shard, folding a hot key into one backend read; under a >90% CPU overload across ~1,350 tenants, only the 6 noisy neighbors were shed while goodput held near 97%.

📚 Tutorials & Advice

🔗 Other News & Articles

• Shopify Introduces gisting: compressing LLM system prompts into learned tokens LINK

• Audacity 4.0 audio editor released with Qt6 Based UI LINK

• AI makes pull requests less central in open-source contribution LINK

• I turned my $15 ESP32 into a mini PC with this tiny OS LINK

• Bun v1.4.1 LINK

• Kubernetes v1.37: DRA Updates LINK

📄 Papers & Resources

🧰 Latest Tools

💬 How did you find today's edition?

We read every reply — just reply to this email and let us know how we can improve!