AI Beats Emergency Physicians at Triage Diagnosis in Harvard Trial

The Guardian reported that in a trial conducted by a research team at Harvard Medical School, AI outperformed physicians in triage diagnosis in the emergency department. Prioritizing patients in emergency medicine is an area that depends heavily on staff experience, so the fact that AI's objective analytical capabilities showed an edge here is highly significant. At this stage, the result is positioned not as something that "replaces physicians," but as evidence of its potential as an assistive tool.

Apple "Surprised" by AI-Driven Surge in Mac Sales — Supply Constraints Continue

In its earnings report, Apple revealed that demand for AI features has pushed Mac sales above expectations. In particular, supply for the Mac mini, Mac Studio, and Mac Neo is expected to remain constrained through the next quarter. The backdrop appears to be a growing user base seeking local AI processing as Apple Silicon performance improves. It's a data point showing that the Mac platform is steadily cementing its position as an AI development environment.

AI Discovers New Behavior of DNA Within Cells

In research at the Gladstone Institutes, AI analysis revealed that DNA behaves far more dynamically within cells than previously thought. The finding overturns the conventional understanding that "DNA is locked away inside the cell," and it is drawing attention as an example of AI in fundamental biology yielding new insights.

Multi-GPU Adoption for Local LLMs Gains Momentum — CUDA+ROCm Simultaneous Execution and a 32-Card MI50 Cluster

A new llama.cpp feature, `GGML_BACKEND_DL=ON`, makes it possible to use NVIDIA CUDA and AMD ROCm at the same time, and it's been a hot topic on Reddit. A user actually ran the Minimax 2.7 Q4 model in a CUDA+ROCm hybrid configuration and confirmed improved prefill performance. Previously you were limited to one platform or the other, so being able to run inference efficiently in mixed environments is a big deal.

Meanwhile, another user reported running Kimi K2.6 (int4 quantization) on a cluster of 32 AMD MI50 32GB cards, achieving 9.7 tok/s (generation) and 264 tok/s (prefill). Power consumption peaked at 4800W, and the user themselves concluded that "it's not worth it unless you have solar panels or free electricity," but it's a fascinating proof of concept for large-scale local inference using secondhand MI50s.

An MLX Build of Mistral Medium 3.5 Arrives — With a "Broken" Warning Attached

(An update from the previous article on the Mistral Medium 3.5 release)

A community member has converted Mistral Medium 3.5 128B to Apple MLX 4-bit quantization (about 70GB) and released it. However, the converter themselves warns that "this model is completely broken at this point. I do not recommend downloading it except for troubleshooting purposes," suggesting problems with the model's quality itself. There are reports of around 5 tok/s on a 96GB M2 Max, but looping output issues have also been flagged, so stable operation still seems a way off.

hipfire: A Tool Taking on the Long-Context Problem for Local LLMs

An AMD RX 7900 XTX user reported finding a tool called hipfire to solve long-context errors with Qwen 3.6 27B, and got it running in Docker. hipfire combines a TriAttention sidecar with DFlash drafts, and a generation speed of roughly 40 tok/s was confirmed. It's still at a stage with plenty of room for optimization, but it looks promising as a long-context inference option on AMD.

Yokohama Rubber Independently Develops an AI×Simulation Tire Mold Design System

Yokohama Rubber announced that it has independently developed a tire mold design support system that fuses simulation with AI technology. The company says that even less-experienced engineers can now design molds, delivering faster development, lower costs, and fewer rework cycles. It's a noteworthy example of practical AI adoption in manufacturing.

---