the key is just an offsite machine with a stable IP plus a VPN. that way you can route all public traffic to that stable IP and send it where it needs to go regardless of its physical destination.

i have a VPS offsite to act as a gateway. it’s just a small piece of a machine somewhere in my region that routes requests to my home network via Tailscale. this has a few benefits:

• i don’t have to worry about my ISP changing my IP. my VPS has more stable IPs.
• i don’t have to expose ports directly to the internet. Tailscale authenticates the connection. plus i have Caddy routing the whole system. i use subdomains like foundry.chrash.net, jellyfin.chrash.net, etc.
• another benefit of Tailscale to point out is that you don’t need local IPs to be static either; Tailscale will allow you to access your machines by hostname or another static IP. this helps to decouple your local topology from your service network.

exactly! i basically live in the terminal, and this is my go-to shell for all platforms

i use Nushell for this! works with JSON, YAML, TOML, markdown, Polars Dataframes, SQLite, and a bunch of others including builtin parsing tools for whatever formats and a plugin ecosystem. i use it at work and for personal projects as my main shell, and it’s super handy for exploring, unpacking, sorting, and visualizing all sorts of data. i use it to:

• find specific parts of YAML cloud configs
• visualize JSON logs, including a parser that restructures journalctl logs.
• _re_structure data from CLIs to work with them as structured: git logs, Unix coreutils, etc
• script my environment: common kubectl queries, specific web API helpers, building and running and testing applications, etc

it is a slight learning curve, and technically you could do all of that with bash or zsh and jq or jc, but i appreciate the modern take on your base shell terminal env.

it’s replaced both Python and Bash for me.

in a container

well there’s your issue. i get not liking the OS, but actively crippling your project will cripple your project.

containers on macOS do kinda suck

honestly it’s hard to beat Macs these days in this space for two reasons:

• unified memory means that you don’t have to load up on RAM just to load the model and then also shell out for a video card with barely enough VRAM to fit a basic language model
• their supply chain is solid and has mostly avoided the constraints that other OEMs and parts manufacturers are struggling with

pricing is tough. sure, crypto is on its way out, but GPUs are still the platform of choice for most neural net workloads (outside of SoCs like Apple M-series). i built a PC in late 2024, and it’s easily worth twice what i paid for it.