New Orleans Plans to Have AI Answer 911 Emergency Calls

New Orleans, Louisiana has reportedly been considering having AI handle 911 emergency calls in place of human operators. According to reports, the system would have AI respond instead of human dispatchers, and its impact on emergency response and crime handling has become a point of debate. Many details remain unclear at this point — including the specifics of the rollout and whether it will be a phased transition — and how safety and reliability will be ensured looks set to become the central issue.

Bringing AI into public emergency services pairs the benefit of easing staffing shortages with the heavy responsibility of misresponse risk in life-critical situations. A proof of concept at 911 — the highest-stakes entry point there is — could become a case that forces a rethink of the assumptions behind AI operations in other public services.

Atlassian Dismisses 'Swallowed by AI' Fears with Revenue Growth

Atlassian, the software collaboration company, posted strong earnings and its stock surged in after-hours trading. Fourth-quarter (June fiscal period) revenue reached $1.77 billion, up 28% year over year, easing market concerns that its business would be taken over by AI apps. Its revenue guidance for the next quarter of $1.71–$1.72 billion also topped market expectations of $1.67 billion.

The market appears to have rewarded the company for sustaining growth by building AI into its own products, countering the view that incumbent SaaS companies will be culled by the rise of AI. It stands as a reference case for whether B2B software companies can turn the AI threat into a revenue engine rather than let it remain a threat.

Concerns Raised About the 'Real-World' Power Consumption of Agent AI

An article analyzing the power consumption of agentic AI — which autonomously works through multiple steps — has been drawing attention. It points out that agentic workloads tend to consume far more power than conventional chat-style usage, because a single task invokes the model many times. With data center power demand already surging, the analysis highlights once again how the shift toward agents piles additional load on top.

The trade-off — that pursuing accuracy and autonomy sends cost and environmental impact soaring — will become an unavoidable issue as agent AI adoption picks up in earnest. Efficiency measures and how that power is sourced look set to become evaluation criteria just as important as the technology itself.

The Front Lines of Local LLM Inference: The llama.cpp 'Split Mode' Trap and Running 27B on Two GPUs

Among enthusiasts running large models locally, two practical pieces of know-how for rethinking inference speed have been shared back to back.

The first concerns a trap lurking in llama.cpp's settings. A user running Qwen 3.6 27B on a dual RTX 3090 setup discovered that after sticking with `--split-mode tensor`, prompt processing (reading in the input) had effectively fallen to the CPU. For dense models, tensor splitting is not supported by the backend processing on the GPU, so the workload silently fell back to the CPU. As a result, prompt processing was stuck at roughly 400 tokens per second.

Switching to `--split-mode layer` and placing all layers on the GPU sent prompt processing soaring roughly fourfold to 1,500–1,700 tokens per second. Generation speed drops somewhat, from 60–70 tokens per second down to the 50s, but loading long contexts becomes overwhelmingly faster, which is a major practical win. The user also cautions that settings can easily cost you the GPU's benefits — for example, raising `ubatch` too high pushes a layer onto the CPU and wipes out the gains.

The second is a 27B setup on a more affordable configuration: a case where someone ran Qwen 27B on two RTX 5070 Ti (16GB) cards and published a stable, practical configuration. By using a recent inference engine with the KV cache fix incorporated, they kept generation-speed degradation in check even with two parallel requests, recording 94–87 tokens per second of generation across a 0–120K token context, with prefill at 4,600–2,400 tokens per second. They secured 170K tokens of KV cache on the GPUs and added another 246K tokens using 8GB of main memory, arguing the setup holds up well enough for long-context agent workloads.

We're starting to see the threshold where a 27B-class model can be put to practical local agent use on two consumer GPUs. Because the fine details of configuration dramatically swing performance, the spread of community know-how matters a great deal.

Trump Administration's AI Framework Kept Under Wraps

The Trump administration's AI framework is drawing criticism because its contents have not been made public. Coverage in opinion media argues that the very fact that the substance of the framework is not being shown is itself the problem. Transparency is once again being called into question over the direction of AI policy.

Note that the source article comes from an opinion outlet written from an explicitly critical standpoint, and there is limited ground at this point for stating anything definitive about the framework's actual contents or intent.

AI Agent Dev Environment 'Orca' Gains Plugin Support and a Marketplace

Orca, a development environment for AI agents, added plugin support and a marketplace in v1.4.162. What's notable is how the mechanism differs: rather than storing plugin packages in one central place like the VS Code extension store, it reportedly organizes and manages where each plugin is fetched from.

As the ecosystem of agent development tools expands, this reflects a design choice about how to safely incorporate third-party functionality. With competition among coding agents and dev environments intensifying, the design philosophy behind extensibility itself is becoming an axis of differentiation.

---

All of these stories are moves characteristic of a transitional period, as AI permeates everyday work and the gateways of public life. Across multiple axes — convenience versus safety, cost versus environmental impact, and transparency — field-level know-how and institutional debates are advancing at the same time.