is the Verge’s play editor. He’s covered the tech manufacture for complete 18 years and knows a point aliases 2 astir synths.
China’s Zhipu AI (Z.ai) released its open-weight GLM-5.2, and immoderate researchers person claimed that it matches Mythos successful definite bug-finding and cybersecurity scenarios. While GLM lags down models from Anthropic and OpenAI successful other, much wide tasks, it seems that China has dramatically reduced the spread successful the capabilities betwixt its models and those of the US.
This level of advancement is peculiarly concerning to the US government, which has worked to restrict China’s entree to powerful models for illustration Anthropic’s Mythos and Fable, arsenic good arsenic the hardware basal to train and tally them. The Trump management views Mythos and different precocious AI models tin of identifying vulnerabilities arsenic superior nationalist information threats. Recently, OpenAI unveiled GPT-5.6, which has besides raised concerns astir its imaginable for misuse and has constricted entree to it.
Because GLM is an open-weight model, it tin beryllium downloaded and tally by anyone connected readily disposable hardware. That gives it awesome elasticity and allows powerfulness users heavy access, but it besides makes it ripe for maltreatment by bad actors who tin tally it pinch small oversight.
Follow topics and authors from this communicative to spot much for illustration this successful your personalized homepage provender and to person email updates.
English (US) · 
Indonesian (ID) ·