Unmanned Lab Begins Autonomous Operations with Robots and AI

A laboratory with no researchers in it — that has now become reality. According to Japan Today, an unmanned lab has opened, and a system in which robots autonomously carry out experiments is now up and running. The fusion of AI and automation has reached the realm of "unmanning" the lab.

Traditionally, lab automation has been limited to specific steps, but this initiative is notable for entrusting robots with the entire pipeline — from experiment planning to execution and data collection. Researchers' roles would be consolidated into designing experiments and interpreting results, potentially freeing them from repetitive work.

That said, details on the types and scale of experiments covered remain limited at this point, so how far autonomous experimentation can actually go will have to wait for future developments.

AI Is Tilting Bargaining Power in the Job Market Toward "Experienced Workers"

According to a Bloomberg analysis, the spread of AI is increasingly tipping bargaining power in the labor market in favor of experienced (older) workers. The argument is that the use of AI tools is shifting the value balance between younger workers' "new skills" and older workers' "experience-based judgment."

Specifically, as AI takes over efficiency improvements and information gathering, the relative value of "adaptability to new technologies" — once considered a strength of younger workers — is declining, while abilities backed by experience, such as complex decision-making and relationship building, are standing out more.

Note that this analysis may be limited to certain industries and does not necessarily apply to every occupation.

Benchmarking Qwen3.6's MTP on RTX 4090/5080 — Is 2x Generation Speed for Real?

The MTP (Multi-Token Prediction) merge into llama.cpp that we touched on in the previous roundup has now been benchmarked for real, with results published on Zenn. The tests measure generation speed with MTP enabled on the Qwen3.6 model, using two GPUs: an RTX 4090 and an RTX 5080.

Coming right after the MTP support PR was merged into llama.cpp, community interest is high. See the original article for the actual measurements and environment details — as a practical indicator of under what conditions the "2x generation speed" claim holds, it is an extremely informative piece of verification work.

An Attempt to Run the LLM8850 NPU on Raspberry Pi 5

An attempt to connect an LLM8850 (an AI accelerator with an onboard NPU) to a Raspberry Pi 5 via an M.2 HAT and run LLMs on an edge device has been reported on Zenn.

Bottom line: the LLM8850 was recognized as a PCIe device, but the DKMS build of the `axcl_host.ko` driver failed, so it never actually operated as an NPU. The cause, however, has been pinned down quite clearly, and the whole thing is documented in detail as a technical issue with the kernel module build environment.

In edge AI, it is precisely these "records of hitting a wall" that help the next challenger. The Raspberry Pi 5 + NPU combination holds real promise as a low-cost edge AI environment, and hopes are high for future driver support.

The Design Choice of Writing an MCP Server in C++20 Instead of Rust

While Rust-built tools have become mainstream in the Model Context Protocol (MCP) space, an interesting design philosophy — "why I wrote it in 400 lines of pure C++20" — was shared on Zenn.

The developer built `mcp-cst-core`, an ultra-fast parsing foundation for having AI agents analyze LLVM-scale codebases. In the process, they concluded that generic abstraction layers and the "bloat" of official SDKs were becoming performance barriers, which led to the choice of C++20.

Readable as a counterargument to "Rust can do anything," the article is full of insight for anyone thinking about "what to optimize" when designing AI development tools. It shows, with concrete examples, that deliberately choosing tight coupling can be a rational decision for certain use cases.

---