"Remain in your cube - The Freedom Force is en route to administer freedom reeducation. Please be sure to provide proof of medical insurance prior to forced compliance."

2B is nothing, even 7B is tiny. Commercial API-based LLMs are like 130-200 billion parameters.

I mean yeah, training a 7B LLM from scratch on consumer-grade hardware could take weeks or months, and run up an enormous electric bill. With a decent GPU and enough VRAM you could probably shorten that to days or weeks, and you might want to power it on solar panels.

But I haven't calculated what it would take to do on rented compute.

Fuck, you almost sold me on GeForce Now. Owning is still a better value proposition for me because I get my games at... steep discounts.

It's nicer to develop anything on a beefy machine, I was rocking a 7950X until recently. The compile times are a huge boon, and for some modern bloated bullshit (looking at you, Android) you definitely need a beefy machine to build it in a realistic timeframe.

However, we can totally solve a lot of real-world problems with old cheap crappy hardware, we just never wanted to because it was "cheaper" for some poor soul in China to build a new PC every year than for a developer to spend an extra week thinking about efficiency. That appears to be changing now, especially if your code will be running on consumer hardware.

My dad used to "write" software for basic aerodynamic modelling on punchcards, on a mainframe that has about us much computing power as some modern microcontrollers. You wouldn't even consider it a potato by today's standards. I'm sure if we use our wit and combine it with arcane knowledge of efficient algorithms, we can optimize our stacks to compile code on a friggin 3.5GHz 10-core CPU (which are 10 year old now).

I can code my stuff ok on an older model. I’m sure there are some stacks that need more resources, but I’m having a hard time thinking of which.

Admittedly, on a laptop that is 20+ years old, I cannot surf the web AND run docker at the same time

Honestly it's fine. LSPs are nice but you don't need them per se. A combination of vim, tmux, entr, a fast incremental compiler, grep, and proper documentation can get you a long way there.

A lot of critically important code that's running the servers we're using to communicate was written this way. And, if capitalist decline continues long enough, we will all eventually be begging for vim while writing code with ed.

Personally I use helix with an LSP, because it helps speed up development quite a bit. I even have a local LLM for writing repetitive boilerplate bullshit. But I also understand that those are ultimately just tools that speed the process up, they do not fundamentally change what I'm doing.