AI News Roundup: Meta Brain2Qwerty v2 Brain-to-Text, Cursor iOS Remote Agents, and DeepSeek V4 in llama.cpp
English summary
Meta released Brain2Qwerty v2, a non-invasive brain-to-text decoder achieving up to 78% word accuracy on its best participant and 61% average across nine volunteers, with training code and the v1 dataset made public. Cursor launched an iOS app featuring always-on cloud agents, remote control of desktop agents, and in-app PR diff review. Open-weight model access was productized: cline introduced a $9.99/month pass bundling GLM 5.2, DeepSeek, Kimi, MiniMax, and Qwen, while Cognition's Devin Fusion harness claimed a 35% cost reduction for coding via hybrid-model orchestration. DeepSeek V4 support and DFlash speculative decoding were merged into llama.cpp, with community demonstrations of GLM-5.2 753B running at 16 tok/s across two Macs using llama.cpp RPC. Arena exceeded $100M ARR and 700M conversations, pivoting toward post-deployment agent evaluation, and a Claude Code safety incident prompted calls for sandboxing AI agents.
Chinese summary
Meta发布了Brain2Qwerty v2，一种非侵入式脑到文本解码器，在最佳受试者上达到78%的单词准确率，九名志愿者平均61%，同时公开了训练代码和v1数据集。Cursor推出iOS应用，支持常驻云端代理、远程控制桌面代理及应用内PR差异审查。开源模型访问被产品化：cline推出9.99美元/月的订阅，捆绑GLM 5.2、DeepSeek、Kimi、MiniMax和Qwen；Cognition的Devin Fusion混合模型框架声称通过编排降低了35%的编码成本。DeepSeek V4支持和DFlash投机解码已合并到llama.cpp，社区展示GLM-5.2 753B在两台Mac间通过llama.cpp RPC达到16 tok/s。Arena年经常性收入突破1亿美元，对话超7亿次，正在转向部署后代理评估；一起Claude Code安全事件引发了对AI代理沙箱化的呼吁。
Key points
Meta’s Brain2Qwerty v2 decoded sentences from non-invasive MEG/EEG signals in real time, reaching 61% average word accuracy and 78% for the best participant, with training code and dataset released.
Meta的Brain2Qwerty v2从非侵入式MEG/EEG信号中实时解码句子，平均单词准确率61%，最佳受试者78%，并发布了训练代码和数据集。
Cursor launched iOS remote agents: always-on cloud agents, remote control of desktop computer-bound agents, and mobile PR diff review.
Cursor推出iOS远程代理：常驻云端代理、远程控制桌面计算机代理以及移动端PR差异审查。
Open-weight model access was productized via cline’s $9.99/month pass for GLM 5.2, DeepSeek, Kimi, MiniMax, Qwen, and Cognition’s Devin Fusion hybrid coding harness claiming 35% cost reduction.
开源模型访问被产品化：cline推出9.99美元/月订阅，包含GLM 5.2、DeepSeek、Kimi、MiniMax、Qwen；Cognition的Devin Fusion混合编码框架声称成本降低35%。
DeepSeek V4 PR and DFlash support were merged into llama.cpp, enabling local inference of large models; community achieved 16 tok/s on GLM-5.2 753B via multi-Mac RPC setup.
DeepSeek V4 PR和DFlash支持合并到llama.cpp，实现大型模型本地推理；社区通过多Mac RPC配置在GLM-5.2 753B上达到16 tok/s。