It's nice to see some recognition of the idea that interpreters are safe by default (excluding infinite loops/OOM), and we can avoid many security concerns (access to files/network/etc.) by simply not including that functionality in the interpreter (unless opted in via a startup parameter, as described).
I'm also a fan of using env vars for configuration and locating dependencies, as mentioned in the talk. Much simpler and easily extensible compared to e.g. vendored directories (requires messing with the contents of the source directory, which requires write access and breaks hashes, etc.), hard-coded system paths like /usr or ~/.some-default-location (causes conflicts when running multiple incompatible versions), etc. A simple env var of paths, e.g. colon-separated, maybe with some sane quoting convention, can be used for all of those if desired, whilst making it super easy to extend-with or restrict-to any other location(s) instead. It's also trivial to "bake in" an env var to an application, just call `PACKAGE_PATH=foo:bar my_app` rather than `my_app` (or make a one-line wrapper script).
I agree with others that importing from URLs seems like a bad idea: network I/O is one of the least reliable actions we can take, which would make importing far more complicated than necessary (what if we're offline? should we follow redirects? should we check for proxy settings? how should we report errors? etc.). All of this complexity and the inescapable problems of network failures are completely avoidable by just downloading things up-front. Package managers/build tools can fetch whatever they like (URLs, git repos, etc.), however they like (with proxies, caches, etc.), to wherever they like (one big cache, project-specific vendor dirs, whatever), and just stick the resulting directory (possibly of symlinks) in the program's environment (e.g. via a wrapper script, as above).