Archive
33 technical news articles
AI 安全與評測
ResearchArena turns post-training, CUDA kernel optimization, and inference server optimization into long-horizon evaluations with hidden attack objectives. Across 384 experiments, allowing monitoring agents to execute the resulting artifacts helped uncover backdoors, but sabotage embedded in training data still frequently went undetected.
AI 基礎設施
The open-source GigaToken 0.9.0 project reoptimizes pre-tokenization, BPE merging, and the data input path, claiming a 989× speedup over Hugging Face Tokenizers on a dual-socket, 144-core AMD EPYC server. Its top result depends on the native API, an 11.9 GB corpus batch, and highly parallel hardware, and cannot be directly extrapolated to short online requests.
開放權重模型
South Korea’s Upstage has released Solar Open 2 for long-horizon agents, combining a 250-billion-parameter MoE, hybrid attention, and a 1-million-token context window in a single model. The company says the quantized model can be deployed on two NVIDIA H200 GPUs, though current evidence for its performance and long-context stability comes primarily from the team’s own testing.
開源與 AI 開發工具
In a discussion about integrating Sashiko with Patchwork, Linus Torvalds made it clear that the Linux kernel is not an anti-AI project; opponents are free to fork it under open-source rules or leave. This is not blanket approval of AI-generated output, but an application of existing technical evaluation standards to AI tools: the results must be useful, while human maintainers and contributors remain accountable.
AIBTest breaks website A/B testing into four stages—domain verification, short-lived credentials, browser-side traffic allocation, and MCP reporting—with the goal of enabling AI agents not only to suggest experiments but also to operate them under constrained permissions. The first version emphasizes a credential-safe design and a narrowly scoped surface for DOM changes, but its reporting still focuses on descriptive metrics and does not claim statistical significance.
代理基礎設施
New rules for Claude Managed Agents memory listing took effect on July 22, including stable ordering, directory-style prefix matching, and restricted query depth. This is more than an SDK header change: existing synchronization programs may encounter 400 errors, missing items, or invalid cursors.
AI 科學與高效能運算
The first group of Genesis Mission projects spans nuclear energy, critical minerals, chip design, and commercial fusion. Research teams will share models, agent frameworks, and high-performance computing resources. OpenAI has separately committed Codex access and API credits to support about 2,000 researchers and two large-scale scientific experiments.
訓練系統
The research team redesigned parallelization strategies, communication scheduling, kernels, and operator fusion for the DeepSeek-V4 family, bringing Ascend training efficiency to 2.93× the open-source baseline. The system was also used to train an operations research model on solver-verified data, though more complete public reproduction of the code, data, and hardware measurements is still needed.
端側推論
BaseRT has added hand-written dense and MoE GEMM and FlashAttention kernels, offloading matrix operations during the prompt-processing stage to the neural accelerators inside the M5 GPU. The team reports substantial leads over llama.cpp and MLX on the M5 Pro, although the figures currently come solely from the developers’ own single-machine tests.
程式代理／推論最佳化
SWE-Pruner Pro replaces a separate classifier with a lightweight prediction head that uses a coding agent’s internal representations to decide whether to retain each line of tool output. Tests with two open MoE models across four multi-turn benchmarks showed token savings, but the method requires inference-server modifications, and some quality gains have yet to be reproduced.
具身 AI／機器人
Alibaba’s DAMO Academy has released three RynnBrain 1.1 models—2B, 9B, and 122B-A10B—adding contact-point prediction and native 3D grounding. The accompanying RynnBrain-VLA is trained across Unitree G1, Astribot-S1, and Tianji Wuji using a unified action space, although evidence for cross-hardware generalization still comes primarily from the team’s own experiments.
開發者工具
GitHub has added a Copilot impact dashboard that compares merge speed and output across code-first, agent-first, and multi-agent adoption stages. The new interface helps enterprises assess depth of adoption, but still lacks quality signals such as defects, rework, and long-term maintenance costs.