Hazey on Nostr: I used to do all my coding with local models, and anecdotally my experience is the ...

I used to do all my coding with local models, and anecdotally my experience is the performance gap never exceeded 10% compared to GPT (which I used for my fiat job). Since moving to OpenCode my productivity skyrocketed but so did token usage. The large context windows of cloud models are very nice but probably contribute to me burning up to 100k sats worth of tokens in a day of heavy use. It may be time to go back to local only. Those two 3090s are blazing fast but peg 700W during sustained inferencing. I haven't cost computed the power usage but I'm getting solar soon which will drop that to "free," at least during the day, when factoring out capital cost.