Experiment: Vibe Code with cloud LLM using Copilot with GPT-4.1

[jethrocarr]

I ran local LLMs in three tests in #9 , #10 and #11. I then ran this one using Copilot on GPT-4.1 for comparison and it pretty much nailed it right out of the gate.

Only three prompts total for this PR, each run in very short succession like a conversation:

• "Update the application code to follow latest JS standards, as well as latest AWS SDK" ✅ (success, code is basically perfect)
• "please update the dependencies" ✅ (success, it removed old libs)
• "update the serverless.yml runtime to node22" ✅ (success, it updated the yml file without error)

Overall:

• GPT-4.1 is one of the basic "unlimited" (at least right now) models in copilot and it just smashed it right out of the park.
• The local LLM exercises took almost 1-2 hours each to get through. It would be faster with a whopping big GPU but at massive financial cost. Possibly worth it if needed code assistance in an airgapped or offline environment for some reason.
• It does feel like the local LLMs are maybe 12-18 months behind the frontier cloud models , that said it's still super impressive that we have OpenAI flagship level performance from only a few years back locally. If the pace remains consistent, we might see some really viable local dev agents in the next few years.