nikhil.io

(Addressed to myself. I look forward to eating crow over all or most of this.)

To use LLMs effectively as a programmer,

1. You still need to know what to tell them¹.
   Be explicit. Especially with success/failure conditions. Especially when laying out a brand new project or module.

2. You need to check all the work.
   “Trust But Verify.” Did you solve the right problem? Are your tests testing the right thing? Are the docstrings and comments actually helpful (and correct)? Is the generated code maintainable, modular, readable? Are you confident you’ll be able to integrate features and bugfixes after you stepped away for six months?

3. Vibe Coding is bullshit for anything other than proofs of concept.
   And atrophies your skills and your passion (that last part makes you a human being and not the thing they want to replace you with.) If you’re not a programmer, go wild. Citizen coding ftw. And if your app blows up, thank you for ensuring our job security.

With all three, Wisdom and Experience will still matter. Both generalists and specialists will still matter. Replacement fears are mostly unfounded and driven by greedy and/or uninformed people. Don’t listen to them. Hone your craft daily, be as good as possible at it, keep your chin up, and you will be fine.

The Future #

The future will be complicated and nuanced and unpredictable. But here I go anyway.

The free-use, ad-free gravy train will end soon, especially with OpenAI and Anthropic. You will pay by the token or at least be inundated by stultifying and privacy-abusing ads. You will understand the true cost of inference and learn the larger and depressing reason behind how non-Google companies continued to attract investment (“Radix malorum est cupiditas.”)

Any personal AI that is both unobtrusive and effective 99% of the time (as endorsed by users) is what will endure. Think of OCR when you hit spacebar on an image of a receipt on Mac, but at a very advanced level. Good design is very hard and people don’t like it when you shove your shiny new toy into their faces, especially when they use your product remembering your cartoon villain glee at replacing them.

On-device LLMs, most likely fine-tuned open models, that are ‘good enough’ and take you 80%+ of the way there are the future.