Written by VPs of sales from Anthropic and OpenAI?
Where this article fails the worse: in my experience smaller local models are not often used in agentic tasks that involve code execution so much of the otherwise OK points don't apply. Also, when I have played with, for example, the Agno agent library with local models, I have the application code print/display any generated Python code before execution, and local sandboxing is not difficult to do!
Local models and embedded models excel at data transformation, NLP tasks, etc.
Especially with agentic browsers like OpenAI Atlas, Comet, etc., there are real security concerns. Probably more of a concern that running local models.