eMudhra Eight Intermediate CA Private Keys Compromised, Affecting Major Root Trust StoresTorchTPU: A Hybrid Eager/Compiled Architecture for Native PyTorch on TPUVercel Functions WebSocket Support Goes GA, Serving 15M Connections DailyPolyGraph Uses Plan-Buffered CUDA Graphs to Speed Adaptive Speculative DecodingPeekKV: The First Content Token Already Reveals Which Attention Heads MatterSoftware World Simulator Tests How Well AI Agents Collaborate at ScaleWhen Should Tool-Using AI Agents Act, Ask for More Info, or Escalate to Humans?Context Management and External Memory for Reasoning ModelsSimon Willison's Vibe-Coded RTS Battle Drones Is Actually Fun to PlayFranchisee Cuts Cash Reconciliation Time 94% with ChatGPT WorkSF Compute's Evan Conrad on Programmable Data Centers and GPU TradingThree Java Champions on Using GitHub Copilot to Boost ProductivityHow to Build an 'Everything' App That Controls Your Home Devices with AIClaude Code v2.1.305 Adds sandbox.allowNestedUserNamespaces SettingDeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, Mercury 3 Debut but Not Yet Widely UsedOpenAI Codex Ships 0.164.0-alpha.8 Pre-releaseeMudhra Eight Intermediate CA Private Keys Compromised, Affecting Major Root Trust Stores

A new arXiv paper introduces TorchTPU, an open-source collaboration between Google and Meta that provides a native PyTorch backend for Google TPUs. It enables out-of-the-box model execution and TPU migration while delivering industry-competitive inference performance, using a hybrid eager/compiled architecture that balances flexibility and speed. For PyTorch teams, it could mean migrating to TPU without extensive code rewrites.

Vercel has made WebSocket support in Vercel Functions generally available. Since the public beta launched in late June, Vercel CDN has served 300 million WebSocket connections and now handles 15 million per day. Customers use WebSockets for AI chat, streaming, and realtime multiplayer apps. The GA release lets developers build low-latency, long-lived connection apps directly on Vercel Functions without standing up a separate WebSocket server.

Adaptive speculative decoding changes the active batch size B and draft length d at every decode step, but CUDA Graphs replay a fixed launch configuration, so the fluctuating verification width M = B×d forces serving systems to pre-capture many graph variants. PolyGraph addresses this with plan-buffered CUDA Graphs that accommodate dynamic decoding parameters while cutting pre-capture overhead. For LLM serving, that means lower memory and startup costs, making adaptive speculative decoding more practical to deploy in production.

KV cache grows linearly with context length and is a major memory bottleneck in LLM serving, and recent work exploits the functional roles of attention heads to compress it non-uniformly. PeekKV’s key observation: the first content token in a sequence already carries enough signal to determine which attention heads matter for the current request, so the compression strategy can be fixed before generation begins. That lets KV cache compression kick in earlier and cuts memory use in long-context scenarios without waiting to adapt mid-generation.

A new paper introduces Software Worlds, a simulated software engineering environment for studying how teams of AI agents collaborate at scale. The work examines whether agent collaboration can serve as another axis for scaling inference-time compute as agents take on larger projects and open problems, and tries to pin down what makes such collaboration effective and scalable. If the findings hold, multi-agent division of labor could become a standard approach for long-horizon complex tasks rather than just stuffing more context into a single model.

Tool-using LLM agents repeatedly hit decision points under uncertainty: act autonomously, gather more information, or escalate to a human. Each choice trades off task success against interaction cost and the risk of error. The paper frames this “last doubt” problem and works toward agents that know when to act and when to step back. In practice, that means automated workflows can hand control back at critical junctures instead of pushing ahead while uncertain and producing mistakes.

Reasoning models track the state of their reasoning process through in-context memory, but complex problems demand larger context windows and more memory. The paper notes that context management without caching risks discarding information that is still useful. It examines how to manage context and external memory during reasoning so models retain key intermediate state on long, complex tasks. The practical implication: reliability on long-horizon reasoning depends on memory strategy, not just window size.

Simon Willison released Battle Drones, his latest vibe-coded experiment: a Command & Conquer-style real-time strategy game with audio-based multiplayer. Unlike his earlier experiments, he says this one is actually fun to play, and admits he got there by repeatedly telling Claude to “make it a LOT more fun.” The takeaway: non-expert developers can now ship a playable game prototype with AI assistance, though the fun factor still depends on heavy prompt iteration.

An independent franchisee operating Schlotzsky’s and Cinnabon locations used ChatGPT Work to cut menu repricing time by 95% and cash reconciliation time by 94%, building a workflow that could free up $100K in cash flow. The case shows how AI tools land in small and mid-sized restaurant operations: offloading repetitive pricing and reconciliation work to the model so operators can focus on cash flow and store management.

In a Latent Space interview, SF Compute’s Evan Conrad discussed the shift from GPU trading to physical settlement, independent markets, and programmable compute. The core argument is that compute is evolving from a scarce resource into a tradable, settleable commodity, with independent market mechanisms potentially reshaping GPU pricing and allocation. For teams needing large-scale compute, this points to more flexible procurement channels and contract structures down the road.

The GitHub Blog had three Java Champions share how they use GitHub Copilot to solve problems, accelerate development, and keep humans accountable for results. The focus is on real workflow usage rather than a product update. Java developers can see which tasks Copilot handles in actual projects and where human review still matters.

Ben’s Bites shared a personal project: building an ‘everything’ app that uses AI to control home devices in one place, such as operating a TV. The piece is a hands-on walkthrough showing how AI can be used for custom smart-home control and device orchestration. It’s a reference approach for DIY home-automation builders rather than a commercial product.

Claude Code released v2.1.305. The main change is a new sandbox.allowNestedUserNamespaces setting that lets sandboxed commands on Linux create user namespaces—for example to run Chromium or rootless Podman—getting around the sandbox’s file protections. It also adds CLAUDECODERETRYWATCHDOGUSAGELIMITRECHECK_MS for usage-limit rechecks in unattended retry mode. Developers running containers or browser tasks inside the sandbox should note this release.

Latent Space’s AINews brief noted that three new models—DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3—have appeared but won’t be in super wide use just yet. Framed as a ‘not much happened today’ roundup, the brief indicates a quiet day for major industry news, with these models worth noting but not yet having real-world impact.

OpenAI Codex released 0.164.0-alpha.8, following 0.164.0-alpha.7. The public release notes only list the version number with no disclosed changes. This is an alpha-stage iteration with features and stability still in flux, useful for developers tracking Codex’s pace of development but not recommended for production reliance.