No amount of claude can handle the key thing that Tailscale fixes about wireguard, all of the connection to connection logic is handled be it direct, via all sorts of weird hole punching or via relays. Wireguard itself is horrible to do this with, half the time you're just looking at `wg status` wondering why the packet counter isn't incrementing today. It can never achieve the magic that is a direct connection between plane wifi and a host that's behind NAT on the ground.

▲

hdgvhicv an hour ago | parent | next [-]

I can’t think of a single time I’ve wanted to connect between two hosts when I haven’t specifically allowed one through a firewall

	▲	drum55 44 minutes ago \| parent [-]
		NAT are not firewalls.

▲

swingandamiss 3 hours ago | parent | prev [-]

[dead]

▲

LoganDark 2 hours ago | parent [-]

It makes me happy that I can automate some menial tasks or replace a search engine sometimes, but this is just because I already have knowledge of the field and the skill to apply it. You're not wrong that it'll probably get even better, but I don't see a path right now to unskilled people being able to effectively use an agent. Only time will tell if that can happen.

I do disagree with GP's assertion that Claude can never make something comparable to Tailscale, but that's just because I believe I (and probably many others here) could probably prompt that, not because I think just anyone could. LLMs can one-shot a fair few things based on unskilled prompts, but I don't believe this is one of them yet.

▲

swingandamiss 2 hours ago | parent [-]

Right now it's not unskilled. It's managers, PMs, TPM's, or people with some tech knowledge but who previously couldn't do these types tasks (not that they weren't capable, but it takes a lot of effort and time to get skilled in these areas and they didn't devote the time). They know what they want at a high level and now can do it with AI.

	▲	LoganDark 2 hours ago \| parent [-]
		This is mostly true, I think. I think this is a sweet spot for LLMs, too. Maybe with the right approach, a user could know or figure out how to articulate what they're missing, if the first attempt doesn't work well enough for them. I just worry that e.g. if they forget to review things like protocol details, they could silently end up with flaws and such, or jank. But yeah, there's potential where there wasn't before, and I'm all for that.