Etched in Talks to Raise at a $20 Billion Valuation

According to the WSJ, AI chip startup Etched appears to have entered talks to raise funding at a valuation of around $20 billion (roughly ¥3 trillion). The company is building the programmable ASIC "Sohu" as a challenger to NVIDIA, and speculative capital keeps pouring into inference-specific silicon. In the AI semiconductor market, Cerebras and Groq have already pulled off large raises, and Etched's round looks to be part of that same trend. The talks are still ongoing, so terms and amounts are subject to change.

President Xi Jinping Delivers Keynote at WAIC 2026

At the World Artificial Intelligence Conference (WAIC 2026) held in Shanghai, President Xi Jinping delivered the keynote address. It is unusual for China's top leader to speak directly at the conference, and the move served as a public signal of the country's commitment to placing AI at the center of its national strategy. The speech itself is available on YouTube, but the detailed remarks were not included in the data collected, so this article sticks to the basic facts. WAIC has traditionally functioned as a venue for signaling the direction of China's AI policy, and attention now turns to how this keynote will be reflected in future regulation and support measures.

BeeLlama.cpp v0.4.0 Advances KV Cache Quantization

BeeLlama.cpp, a community fork of llama.cpp, has released v0.4.0. The headline feature is the full implementation of two KV cache quantization capabilities: "KVarN" and the "KV cache precision tail."

KVarN**: KV cache quantization using distributed normalization, which reportedly maintains higher precision than prior methods at the same bit width. VRAM usage spikes have also been improved, and the project considers it production-ready.

KV cache precision tail**: A mechanism that keeps only the most recent N tokens in BF16/F16 and quantizes the rest. It is said to mitigate the problem of important information at the end of the prompt being lost to quantization, without a substantial increase in VRAM usage.

The author has published KLD benchmarks for Qwen 3.6 27B and Gemma 4 31B. SWA architectures (Gemma, GPT-OSS, etc.) are still not ready for production use, and further improvements are awaited. The release also removes the previous DFlash implementation to track upstream, and TurboQuant was dropped as well after underwhelming benchmark results. Smaller refinements such as adaptive draft-max and inference loop protection are also included.

AI Advice Cuts "I Don't Know" Responses from 44% to 3%

A preprint posted on PsyArXiv examined how AI advice affects human response behavior. When participants were shown AI advice, the share answering "I don't know" reportedly dropped from a baseline of 44% to 3%. The effect of AI "completing" answers in areas where people lack knowledge is substantial, but the study may also point to the risk of amplifying misplaced confidence. The work is at the preprint stage; see the full paper for details on participant composition and task design.

The Impact of AI Slop on OSS Contributions

A study posted to arXiv (arXiv:2607.04003) measured how AI-generated PRs (so-called AI slop) affect merge rates for first-time contributors across 294 repositories. The result was an "18.18% decrease," suggesting that the influx of AI slop increases maintainer burden and, as a result, drags down merge rates even for human contributors. The findings once again bring growing maintainer workload and the "fight against slop" into focus as challenges for OSS sustainability.

LocalLLaMA Takes Interest in Kimi K3's Security Capabilities

A discussion about the cybersecurity capabilities of Moonshot's Kimi K3 has kicked off on Reddit's LocalLLaMA. The original poster asks whether K3 carries the same "power" that got Fable judged dangerous and restricted for cybersecurity use, noting that reports so far mostly cover vibe coding and 3D scenario creation, while verification around guardrails remains thin. K3 has drawn attention mainly for coding and benchmark-related topics so far, but security-focused evaluations should start coming in.

Wrap-Up

This window brought a wide range of news spanning hardware to societal impact: large-scale investment in AI chips, nation-level AI messaging, efficiency gains for local LLMs, and measurements of how AI affects human behavior. In particular, Etched's round symbolizes the expectations building around inference-specialized silicon, while BeeLlama.cpp's KV quantization features are a practical improvement that translates directly into cost savings for local deployments.