Research Radar/AI agents & MCP/China
Tencent Zhuque Lab released AI-Infra-Guard, an open-source scanner for AI agents. Its skill checker scored 0.9848 loose F1 on a new public test set of 5,520 cases.
Research Radar/AI agents & MCP/South Korea · USA
Researchers hid fake fields inside data that AI agents read, using quote-like characters. The attack worked 49.1 percent of the time, while old-style prompt injection worked 0.2 percent.
Research Radar/Quantum computing/Singapore · China
A review of quantum error mitigation prices the cleanup. At a half percent error rate, fixing 100 spots costs 7.46 times more runs and fixing 1,000 spots costs about 537 million times more.
Research Radar/Edge computing/China · United Kingdom · Singapore
A 28nm edge chip design for DeepSeek reports 109.4 TFLOPS per watt, which the authors put at 19.35 times an H100 at FP8. No chip was built, so the figures come from design tools.
Research Radar/Semiconductors & chips/India
A review of AI chips shows memory bandwidth, not arithmetic, sets the speed of language models. A flagship phone tops out near 19 tokens a second, and an H100 near 24 on a 70 billion parameter model.
Research Radar/Cloud computing/China · Canada
One default setting in the software that serves chatbots does not lower peak power, but it cuts how fast that power swings by up to 42.6 percent. A grid would need about a fifth less emergency reserve.
Research Radar/Robotics/China · Singapore · USA
A new study trained robots on 5,000 hours of video shot from people's own eyes. Those robots beat robots trained on robot video. The gap was biggest on tasks the robots had never seen before.