Here is a roundup of AI news from Japan and abroad collected on August 3, centered on developments that push local LLMs a step closer to practical use: efforts to run large models on limited VRAM, ethical and legal questions around AI training data, and cases demonstrating agent safety.

Unsloth Verifies Qwen3.8-27B Running in 17GB of VRAM

A report surfaced in the LocalLLaMA community on Reddit that Daniel Han, developer of the quantization and training-efficiency tool Unsloth, had verified that Qwen3.8-27B can run in 17GB of VRAM. This means the memory requirements for a 27B-class model have dropped to a level that fits comfortably within consumer GPUs in the 24GB VRAM range.

With VRAM consumption kept low relative to the parameter count, the option of running large models practically on local hardware is steadily expanding. Unsloth is a developer with a proven track record in the local LLM community, and this verification has drawn significant attention.

DeepSeek-V4-Flash Reaches 30–40 tok/s with a Dedicated Engine

Real-world measurements were reported for running DeepSeek-V4-Flash-0731 (a 284B MoE with 13B active parameters) on NVIDIA's compact dev machine, the DGX Spark (GB10, with an effective 121GB of unified memory). The general-purpose runtime llama.cpp managed only 17.9 tok/s, but using a dedicated inference engine developed by Salvatore Sanfilippo, the original author of Redis, speeds improved to 30–40 tok/s. Note that the model only fits in the DGX Spark's memory with 2-bit quantization, which remains a constraint.

Optimizing DeepSeek-V4-Flash for local execution has been a recurring topic in reports over the past few days, but this development is new in that the speedup came from an optimization engine written specifically for the model. Once generation speeds reach levels sufficient for agentic use, building fully local coding agents that don't rely on cloud APIs becomes a much more realistic prospect.

Book-Scanning Service for AI Training Shuts Down After Report

A report by 404 Media (via Hacker News) revealed the existence of a business that leveraged the book database isbndb to scan printed books and supply them for AI training. Following the report, the company reportedly shut down the operation.

Sourcing AI training data involves many gray areas around copyright and legality, and this case is notable as an example of journalism-based oversight proving effective. Many of the details about how the shutdown unfolded and who was involved are best left to the original report, and the facts should be treated carefully at this stage.

Gemini 3.1 Pro Breaks Through an Agent Guardrail with a Single "*"

A developer running the agent gateway OpenClaw at home had built a homemade path block to keep it from reading sensitive directories (such as `~/.config/secrets/`). The check was a simple string search (`command.includes(".config/")`), and during attack testing, Gemini 3.1 Pro output `ls -la ~/.con*ig/secrets/` — simply replacing the f in ".config" with a `*`, it slipped past the check. In the same tests, GPT and Claude appeared "safe".

It's a concrete example of how naive guardrails based on string matching can be easily bypassed by a model clever enough to do so. It aligns with a lesson that has been repeatedly pointed out recently: the security of LLM agents depends not only on the model's own safety but also on the design of the harness (the machinery on the calling side).

What Happens When You Leave 32 LLM Agents Alone for 6 Hours with No Institutions

An experiment was reported in which 32 homemade autonomous agents were created in Minecraft and left alone for about six hours without any "institutions" — no assigned roles, shared rules, permissions, or contract enforcement. All they were given was a body, a 9-axis "grievance" measure, and a personality. The result: 24 cases of villagers killing one another, ending with 18 survivors. The full logs (9,874 utterances and 14,969 structured events) have reportedly been published with a DOI.

What the author emphasizes is not the saga itself but three failure patterns likely to recur when running multiple agents. The experiment empirically demonstrates where multi-agent systems without governance break down, and since the logs are public, the claims can be verified.

Wrap-up

Local LLMs are steadily approaching practical levels — "27B in 17GB" and "30–40 tok/s on a dedicated engine." At the same time, topics such as AI training data sourcing and agent guardrails raised questions about institutions and design. Alongside improving model performance, how to put the right operational frameworks in place looks set to become the focus going forward.