Large language models and the end of programming.

How much does it cost to replace one human developer with AI? Matt Welsh (co-founder of Fixie.ai and former Harvard, where this talk is given, computer science professor) did the math.

So let's say that a typical software engineer salary in Silicon Valley or Seattle is around 220,000 a year. That's just the base salary, doesn't include blah blah blah. He figures it adds up to $1,200/day. That sounds a bit high to me, but whatever -- it's the order of magnitude that matters here. Because he next figures the cost for a GPT model for the same output is $0.12. Twelve cents.

Ratio is about 10,000.

This suggests "a very large shift in our industry".

If I can compress and paraphrase the "very large shift in our industry" that he predicts is coming, it would be this: What we today call "prompt engineering" will become "engineering", and what we today call "engineering" will be done by the machines.

Humans will continue to play the role of "product manager" -- deciding what software needs to be written based on what customers want. AI will write the software. AI will maintain the software on an ongoing basis. Human programmers will become AI-generated software proofreaders. But eventually people will stop caring about whether the software is readable or maintainable to humans. As AI models improve, the software will achieve the same reliability as software written by humans, and people will gain confidence in it. AI will also write the tests. AI will go from low-level abstractions to high-level abstractions. AI will be able to write code to do tasks humans can't write the code for, like "transform this text to make it kid-safe."

Large language models and the end of programming - CS50

#solidstatelife #ai #genai #llms #economics #computerscience