AWS Bedrock AgentCore: Run Coding Agents in the Cloud and Never Leave Your Laptop Open Again
AWS has announced the Amazon Bedrock AgentCore Runtime, which lets you run coding agents (Claude Code, Codex, Kiro, Cursor, and more) in isolated microVMs in the cloud. Each agent session gets a dedicated persistent workspace, and the design ensures that secrets, ports, and filesystems are never shared between sessions.
Until now, developers have suffered from 'agent syndrome' — keeping the laptop open while agents churn away. With AgentCore, you can run multiple agents in parallel, close your laptop, head home, and pick up right where you left off the next morning. It also offers tool access via a Gateway and built-in observability, dramatically reducing the operational burden of agent development.
WWDC 2026 Approaches: A Major Siri Overhaul and the Evolution of Apple Intelligence
Apple's annual developer conference, WWDC 2026, is just around the corner, and a complete overhaul of Siri is widely expected. According to TechCrunch's forecast, Apple Intelligence features will be significantly expanded alongside iOS 27. Siri has long been hampered by criticism of its lack of comprehension, and this refresh could fundamentally strengthen its LLM-based reasoning capabilities.
Specific feature details will be revealed at the keynote, but expectations include tighter coordination between on-device AI and cloud AI, deeper integration with third-party apps, and advances in privacy-preserving technologies.
Xiaomi Hits 1000+ tps with a 1T-Parameter MoE on a Standard 8-GPU Server
Xiaomi's MiMo team has announced 'MiMo-V2.5-Pro UltraSpeed,' claiming an output speed of **more than 1,000 tokens per second** with a 1-trillion-parameter Mixture of Experts (MoE) model. What makes this notable is that it was achieved on **a single standard 8-GPU server** — not custom wafer-scale hardware like Cerebras or SRAM-heavy chips like Groq.
The news has also made big waves in Reddit's LocalLLaMA community, with commenters saying it would be 'astonishing if real.' It stands as a significant milestone showing just how far inference optimization for MoE architectures has come.
AutoMegaKernel: Compiling an Entire LLM into a Single Verified CUDA Kernel
RightNow-AI has released an open-source project called 'AutoMegaKernel,' an ambitious effort to compile the entire LLM inference process into a single CUDA kernel with mathematically proven correctness.
Normally, LLM inference executes many small kernels in sequence (matrix multiplication, activation functions, attention computation, and so on), and the overhead between these kernels degrades performance. AutoMegaKernel consolidates them into one and minimizes GPU memory reads and writes, aiming for substantial throughput gains. It's a project with the potential to change the design paradigm of inference engines.
Luce Spark: A 35B MoE on a 16GB GPU, Without the Offload Penalty
'Luce Spark' is drawing attention in the local LLM community. It can reportedly run a 35B-parameter MoE model **on a GPU with just 16GB of VRAM** — models of this size have traditionally required 20GB or more.
The mechanism is simple. In an MoE model, only about 8 of the 256 experts are active for each token. Luce Spark keeps only the most frequently used experts resident on the GPU and leaves the rest in system RAM, swapping them in as needed. It also features self-tuning that optimizes the cache to match traffic patterns.
Qwen3.6 35B-A3B runs in about 13.3 GiB, and Laguna XS.2 33B-A3B in about 14.6 GiB. This demonstrates that large-scale MoE models could become practical on affordable GPUs like the RTX 3090 or the 4060 Ti 16GB.
Signs of an AI Industry 'Slowdown' — Ed Zitron Weighs In
Journalist Ed Zitron has published an analysis titled 'AI Is Slowing Down,' which earned 62 points on Hacker News and sparked debate. Zitron presents data and observations suggesting that the pace of generative AI progress is slowing, pointing out that the magnitude of model improvements is shrinking and the returns on the enormous investments being poured in are becoming harder to see.
Similarly, Mashable reports on the growing AI backlash in 'The AI vibe shift is real.' Gary Marcus has also flagged problems with 'slop' (low-quality AI-generated content) and productivity, reinforcing the view that the AI boom is running into a dead end.
While these skeptical voices grow louder, technical breakthroughs like Xiaomi's ultra-fast inference and AutoMegaKernel continue apace — giving the impression that the industry's direction is beginning to diverge.
Concerns Over AI Data Centers Being Built in Clusters in Drought-Hit Regions
According to an investigation by The Guardian, most of the AI data centers planned for construction in the US are slated to be built in **regions struggling with water shortages**. Data center cooling consumes enormous amounts of water, putting additional pressure on water resources in areas already facing severe drought.
As demand for AI computing surges, the impact on water resources — not just energy consumption — is emerging as a critical issue. How to strike a balance between environmental considerations and technological progress looks set to become a focal point of future policy debates.
---