I'll write a longer post on this in a few weeks, but I wanted to respond to all the recent AI FUD regarding whether or not AI LLMs can actually code anything valuable.
Yes. They can.
I have been using SWE LLMs heavily for the past 2-3 years.
But there's a huge caveat:
You must already know how to code whatever it is you want coded yourself before you delegate the actual code-writing to an SWE.
This is because the model will, 100% of the time and with any of the "coding models" in the top-tier, deviate off course unless you know very specifically everything you would need to do as a seasoned, veteran software engineer already yourself.
Otherwise, you get something out the other end that strictly meets whatever prompts you gave it, but which doesn't actually come anywhere near a real MVP of any sort.
This is because in the real world you must account for a ginormous amount of factors, toolings, infra, edge-cases, logging requirements, monitoring requirements, inter-process messaging, thread concurrency, database recovery, etc.
No existing model will just do all that for you unless you specify everything and then correct the model as it "thinks" and produces code iterations.
Unfortunately, this means that these models are effective at replacing entry-level and junior engineers. But they aren't capable of replacing senior engineers, architects, and the like.
Maybe someday they will replace that, too. But I believe that's a long way off still because it really equates to these models learning a much broader context of business-requirements which are adjacent to or marginal factors in the core technical-requirements.
Referenced:
- Claude Code
- Grok Code Fast
- Gemini AI for Coders
- Copilot
(Right now Claude Code blows the others away, irrespective of Elon's bragging. Grok is good at JS, Python, etc., but sucks at C++, Rust, or functional languages)