On August 2, Google greatly expanded the availability of its personal AI agent, Gemini Spark, making it usable in Japan. Meanwhile, community efforts to run the open-weight DeepSeek V4-Flash on home GPUs have been advancing, and the first real-world benchmark numbers are now coming in. On the research side, a new approach to shrinking attention is drawing interest, and on the policy side, an analysis depicting China's AI influence is making waves.

Gemini Spark Lands in Japan, with Chrome Integration to Handle Web Tasks

Google expanded the availability of its personal AI agent, Gemini Spark, to more than 160 countries including Japan. It was previously offered only in a limited set of regions, so its new availability in Japan is the notable point here.

Its distinguishing feature is that the agent keeps running on Google's cloud infrastructure even when your PC is shut down or your smartphone is locked, processing tasks in response to preconfigured triggers. In addition, Google announced a Chrome integration that uses login information and saved passwords to carry out various web tasks on your behalf.

This is a step forward for autonomous agents in the direction of 'completing operations on the other side of the screen on a human's behalf.' That said, a design that lets an agent use saved passwords trades convenience for new trust and security concerns, so we'll want to carefully assess its actual behavior and controls once it rolls out.

DeepSeek V4-Flash-0731: Local Benchmarks of Quantized Builds Arrive

Efforts to run the open-weight DeepSeek V4-Flash-0731 on consumer-grade GPUs are moving along. It was recently reported that fixes for the model arrived in llama.cpp, and now things have moved into the 'real-world measurement' stage — how well does it actually run on top of that?

One user ran the UD-IQ2_M quantization (~90.9GB) entirely in VRAM on a 3x MI50 (32GB) configuration totaling 96GB of VRAM. They consistently recorded roughly 15-16 tokens per second for text generation and about 105-110 tokens per second for prompt processing, reportedly never dropping below 14 tokens per second even on long outputs. The model also handled a simple coding test (generating an HTML animation of a Rubik's Cube), though it did make one mistake, conflating a small fact about memory bandwidth — an interesting example of the limits that come with quantized builds.

Another user took a quality-first approach, publishing a configuration (~111GB) that requantizes only the 129 routing experts to IQ3_XXS while keeping the rest at higher precision. On a mixed-GPU setup with CPU overflow, they report sustaining roughly 1.4x the speed of the standard MXFP4-based build while also improving quality metrics (KLD). With smaller 2-bit quantizations favored when speed is the top priority, the three-way trade-off between size, quality, and speed is becoming visible across the community.

However, a known issue causing stalls on long contexts (around SWA/rollback) is addressed only in some patched builds, and it's pointed out that upstream llama.cpp itself is not yet fully fixed. For serious use, the choice of environment looks likely to matter.

A Proposal to Shrink Attention with INT4 'Memory Cells'

One study that drew interest on Hacker News (7 points at the time of posting) is 'Persistent State Machines: LLM Attention with INT4 In-Memory Cells.' Published on Zenodo, it appears to explore a direction for cutting computation and memory by treating LLM attention as a persistent state machine with INT4 (4-bit integer) 'memory cells.'

Based only on what can be inferred from the title and the reactions, it is an approach aligned with a central theme of current LLM research: lowering inference costs while preserving accuracy. Since the detailed method and its actual effect require reading the paper itself, we'll simply introduce it here as 'a direction worth watching.'

'Token Diplomacy': Depicting China's AI Influence

A Semafor analysis, 'Token diplomacy: How China is shaping the AI future,' is generating buzz. Under the 'token diplomacy' headline, it depicts the reality of China engaging in shaping the direction of the world's AI through the spread of LLMs and the expansion of its influence.

The details of its specific arguments require consulting the original, but the perspective that open-weight models and low-cost API offerings are functioning not merely as technological competition but as a source of geopolitical influence overlaps with the series of developments around DeepSeek. Amid ongoing debates over regulation and sovereign models, it is a piece worth reading closely.

Claude Code in Practice: Plugins and settings.json

On Japanese engineering blogs, insights from real-world Claude Code operation are coming out one after another.

One is the plugin 'i-have-adhd,' which gained traction on Reddit. Once installed, responses take on the structure of 'one-line conclusion → numbered decision criteria → next step.' A comparison asking the same question with and without the plugin reportedly showed that having the conclusion come first makes the answers easier to digest.

The other is a practical settings.json setup that reduces confirmation prompts while still blocking dangerous operations. It combines sandbox and auto modes with a PreToolUse hook that rejects dangerous commands; after about a month of operation, the author reports fewer confirmation burdens and fewer mistakes. The full configuration is published ready for copy-and-paste, making it a useful reference for anyone getting serious about Claude Code.

---

Lately, the impression is that the practical rollout of autonomous agents (Gemini Spark) and the local optimization of open models (DeepSeek V4-Flash) are advancing simultaneously. These past few days, it seems fair to say that 'AI cost and influence' is taking the lead on both the research and geopolitical fronts.