4 ms·
Personally, if I could run capable-enough inference on hardware I control, and could rely on the harness asking me for mechanistic confirmation before the agent
by lxgr 8mo ago
Personally, if I could run capable-enough inference on hardware I control, and could rely on the harness asking me for mechanistic confirmation before the agent can take consequential actions, I'd do it immediately.
- taurath 7mo agoConsequential actions like searching the web or downloading packages or dependencies or doing most anything useful?
- lxgr 7mo agoNo, these are all fine for me (my agent is sandboxed in a container, so it can install all the node modules or Debian packages it wants). I was thinking more of sending outgoing emails, publishing anything on the web, spending my money etc.