China-Born Giant Models Scale Fast Without Nvidia GPUs

Several related stories landed on July 5. One is a report that China's Meituan **trained a 1.6-trillion (1.6T) parameter AI model without using Nvidia GPUs**. The other is the news that **LongCat-2.0, also from China, has become the largest AI model on a path independent of Nvidia chips**. With US export controls still in place, China is increasingly demonstrating in practice that it can scale on alternative chips.

Read together, the two stories point to a clear direction: the race toward ever-bigger models continues, but the procurement route is shifting into a different shape than before. That said, details on training methods and benchmark scores are still limited at this point, so objective performance comparisons remain material for the road ahead.

A California Study: "The Most Affected Are Highly Educated Workers"

A California state study reported by SFGate finds that "**the more highly educated workers are, the more strongly they are affected by AI**." The likely backdrop is that AI automation no longer stops at simple tasks — it now extends into white-collar work such as analysis and document processing.

Meanwhile, a hands-on record from Japan posted on Qiita lines up in the same context. In a case report titled "**I made e-Stat (data from Japan's Statistics Bureau) analysis 100x faster with AI**," the author reflects: "I believed data analysis was a safe zone, but the industry as a whole had reached the same conclusion." As a raw, first-person voice showing that analytical work is precisely where AI is becoming a complement — and a substitute — it makes for an interesting read.

Mouse, an "Editing" Tool for AI Coding Agents

**Mouse**, released by hic-ai.com, is a tool that provides "Precision Editing" operations for AI coding agents. It drew discussion on Hacker News with 17 points and 19 comments, where the comment section surfaced demand for fine-grained human control over agent-generated diffs.

In a neighboring niche, HN also featured **Agentsnap**, a tool that identifies "which config change broke your AI agent." Observation and control tooling for running agents in production seems to be gradually forming its own ecosystem.

A Japanese Article That Maps Out Claude Code's "Memory Design"

An article on Qiita that condenses **Claude Code's memory design** into a quick-reference chart is gathering attention. It organizes, scope by scope, where CLAUDE.md, rules, and automatic memory should live, and it is structured for both "beginners who just want to create a single CLAUDE.md for now" and "developers who want to design the whole system, including imports and automatic memory." As operational know-how for weaving AI agents into real development workflows, it is a practical summary.

Bonus: A "Delete the Sun" Test for Local Models

A playful test was thrown out on Reddit's LocalLLaMA: a thought-experiment prompt asking "**if the Sun suddenly vanished, would there be a day within the next 1,000 years when two planets collide?**" As an easy way to probe a local LLM's spatial reasoning and long-horizon computation, the thread has become a place for people to bring their favorite models. On an axis apart from serious benchmarks, it is fun to see each model's "thinking quirks" come into view.

---