- Author
- Posted by azhenley, with replies from many commenters
- Published
- 30 August 2025
- Kind
- Hacker News discussion thread on a blog post
- Original
- news.ycombinator.com/item?id=45074467
- Subject
- Harnesses and agent loops
What it is
The linked article is "AI models need a virtual machine" on the SIGPLAN blog. The thread has 234 points and was posted by azhenley. Replies debate sandboxes, permissions and whether a virtual machine is the right frame.
What it says
- Commenter mehulashah says systems folks should build abstractions on LLMs, as for CPUs.
- Commenter csmpltn replies that LLMs are one tool among many, not a universal computer.
- Commenter singron says the article lacks specifics about what is being proposed.
- Commenter tylerhou reads the proposal as a standard virtual machine with clear permission rules.
- Commenter spankalee says isolation alone is not enough and data flow must be tracked.
- Commenter nostrademons says capability security never won in mainstream systems. A virtual machine is the nearest option.
- Commenter daxfohl doubts a sandbox helps, since prompt injection can still leak context.
What evidence it gives
The thread offers opinion, analogy and links, with no measurements. A few commenters describe their own tools, such as Dagger and OpenHands. Others cite papers and Wikipedia pages on capability security.
Interests and limits
Commenters are self-selected. The linked article was not read, so its claims reach this page only through commenters. Some commenters promote their own products. Many replies reason from security analogies and none run tests.