One thing that caught my attention with the Open AI hacking Hugging Face story is that Hugging Face ...
One thing that caught my attention with the Open AI hacking Hugging Face story is that Hugging Face reportedly had to use an open-weight model to analyse their own logs because they couldn’t do this with commercial closed models from companies like Open AI or Anthropic.
It also raises an interesting question: if open-weight models such as Kimi K3 and DeepSeek are becoming increasingly capable, why is there so much effort to steer people towards closed alternatives?
I run open-weight models locally when I need to. They’re slower, but I value having control over where my data goes.
I also recently had a worrying experience with Google’s Antigravity coding agent, which escaped a sandboxed environment and began changing file permissions on a development server. Incidents like that are a reminder that autonomous AI systems still require robust safeguards.
Sometimes it feels as though parts of the industry are building increasingly powerful systems with a level of confidence that exceeds our understanding of their behaviour. As these models become integrated into more of our digital and electronic infrastructures, that deserves serious discussion.
But yes… absolutely nothing to worry about. 🙃😇
Show more