本周 AI 回顾 — 2026年8月10日 – 8月16日

Google has introduced HEIR, an open-source compiler within its Private Computing Toolkit, designed to enable cryptographically-secure private AI inference. Since its announcement in 2023, HEIR has fostered collaborations with hardware accelerator developers like Belfort and Cornami, and academic institutions including Georgia Tech and Tsinghua University. This has led to several peer-reviewed publications and numerous citations, demonstrating its role as a productive research platform. Google aims to make homomorphic encryption easy to develop, fast to run, and ubiquitous across the industry.

A developer created a desk HUD displaying live Claude usage on a $38 Thermalright Trofeo Vision 6.86" LCD (1280×480, USB-C) from macOS. This project was inspired by a Reddit post about a similar Claude LCD display. The LCD is a USB HID device (VID:PID 0416:5302) that accepts JPEG frames via a reverse-engineered protocol. The HUD continuously streams data at 2 fps to prevent the display from blanking when idle, utilizing device classes from thermalright-trcc-linux with HidApiTransport.

The Qwen 3.8 27B model repository on Hugging Face provides FP8-quantized weights and configuration files for a post-trained model in the Transformers format. It specifies parameters for an "Instruct" mode, including temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, and repetition_penalty=1.0. The model's output structure includes appending messages with roles like "assistant" and content fields such as "answer_content" and "reasoning_content."

Gemini 3.7 Flash, released on August 13, 2026, offers enhanced reasoning and accuracy for knowledge-dense fields such as finance, law, and biosciences. It significantly outperforms 3.6 Flash on the GDP.pdf benchmark, achieving 34.0% compared to 22.0%. Additionally, 3.7 Flash surpasses 3.6 Flash in AutomationBench, demonstrating improved effectiveness in completing real-world business workflows with a score of 30.4% versus 17.0%.

DeepSeek Harness (dsh), an open-source agent harness from DeepSeek AI, is now available in developer preview. It features a plugin-based architecture powered by Cordis, a system whose design is detailed in "A Programming Paradigm for Spatiotemporal Composability." Users should anticipate rapid iteration and compatibility-breaking changes during this preview phase. A Discord community is available for engagement.

Claude, prompted by an Anthropic staff member, significantly advanced the Riemann Hypothesis by increasing the provable lower bound for the fraction of zeros of the Riemann zeta function satisfying the hypothesis from 41.6% to 67.2%. After 650 initial failed attempts, Claude, coordinating about 60 subagents, ran 2,400 shell commands and wrote hundreds of Python scripts, performing thousands of numerical checks. The staff member's encouragement helped Claude overcome initial skepticism and achieve this breakthrough.

Graft significantly enhances Claude Code, Cursor, Codex, and Gemini by improving correctness and efficiency. It achieved a 12-point increase in correctness, resolving 33 out of 50 instances compared to Cold Claude Code's 27. This improvement came with 23% fewer tokens, 25% fewer tool calls, and 32% less wall-clock time, leading to 19% cost savings. Graft also offers compiler-grade edges via `graft build --lsp` for languages like Rust, C/C++, Go, Python, and TS/JS, using language servers such as rust-analyzer and clangd.

Meta Superintelligence Labs has introduced Muse Glimmer, a 30-billion parameter model optimized for always-on local agent workflows. The model's weights are compressed to approximately 4-bit precision using quantization techniques, reducing its size to under 20 GB. This allows it to run within a 24 GB or 32 GB memory envelope, accommodating its working memory, perception encoder, and speculative decoding drafter. Muse Glimmer is open-sourced under an Apache 2.0 license, with minimal degradation on agentic tasks.

Cerebras and OpenAI have introduced Ultrafast Mode, a new service tier for the OpenAI API, powered by Cerebras. This mode, initially available to select customers, accelerates GPT-5.6 Sol to deliver up to 750 output tokens per second without compromising quality. In evaluations, GPT-5.6 Sol on Ultrafast mode answered 2,500 HLE questions in 11 hours and 11 minutes, nearly 7 times faster than Claude Fable 5, which took 78 hours and 27 minutes for the same task.

Future Claude models will incorporate text watermarking to indicate the likelihood of AI involvement in text generation, aligning with the EU AI Act. This watermark subtly influences token choices, but is less applied to exact content like code or simple sums. Additionally, Claude will attach C2PA content credentials, a cryptographically signed note in metadata, to supported file types like .png, .jpg, or .svg, to show that the file was made or processed with Claude.

Mole is a deep-research agent designed for terminals, featuring an enforced budget and a privacy boundary for local data. It boasts a 0% budget overshoot, 100% claim integrity with verified quotes and sources, and 100% citation accuracy. The grounding rate is 80%, with a precision of 70% with the confirm pass (51% without). Merge precision/recall stands at 1.000/1.000 on constructed ground truth. Mole is released under the Apache-2.0 license.