AI is the end of programming

In this next post we're going to have a look at controversial tech talk on Youtube:

In a CD50 tech talk titled “Large Language Models and The End of Programming“ hosted by CS50 (2023)

Dr. Matt Welsh points out a couple of reasonable arguments why programmers are doomed when it comes to making a living with programming. The software engineering industry would drastically change according to Welsh. The video was watched about 600'000 times and had about 1800 comments after three month. I've watched already two videos, one from Travis Media (2023) and Mattthew Berman (2023)

 

that refer to his talk and assume other videos, blogs and podcasts do as well, which we will further investigate in upcoming blogs.

According to Wikipedia, Dr. Matt Welsh is a Computer Scientist and Software Engineer. His career and research is described as: "Welsh has led teams at Google and Apple Inc., and served a Professor of Computer Science at Harvard University." ("Matt Welsh (computer scientist)",2014).

What are Welsh's main arguments?

1. Humans are not as good at writing code as machines
2. AI models are much cheaper than a human software developer (about factor 10'000)

Ad 1) 

He says that computer science always had the assumption that the programs have to be implemented, maintained and understood by humans and he says the problem is "humans suck at all of these three things." (CS50, 2023).

Welsh believes that the human brain is not made for develop software. He gives the example that tricky problems solved with one language are hard to understand and translate into another language. It would take some time but humans to figure it out but is much better done by machines. 

So today he claims programming has to be done with natural language, using e.g. Github Copilot.

Ad 2)

Total Software Engineering Salary = 220'000 $
+ Benefits taxes, free breakfast, lunch... + 92'000 $
Total = 312'000 $
Number of working days per year: 260
=> Total cost for one human Sofware Engineer per day = 1200 

Lines of codes per day deliverd by human Software Engineer : 100 (assumption)
GPT tokens per line = 10
GPT-3 Costs per 1000 tokens = 0.02 $
=> Total cost for one human Software Engineer day equivalent work by AI : 0.12$
(I don't understand though how he calculates to 0.12$)

The final factor of 0.12$ to 1200$ is 10'000.
And so he claims AI is 10'000 times cheaper per day than humans and thus the industry would radically change.

He further describes in his talk a possible software team of the future formed by the following actors and flows:

PM (still human) writes requirements in natural language -> AI robot will program the code -> Human code reviewer will review the AI-generated code -> PM (starting again from the beginning).

He also thinks that first AI will produce human readable code, but than in the further future AI will evolve to generate directly machine instructions (without abstract code languages).

So let's build upon this controversial talk and analyze what other sources opinion are regarding the future of AI and Software Development.

Feel free to comment on your thoughts.

Cheers Simon

Sources:

CS50. (2023, September 29). Large Language Models and The End of Programming [Video].Youtube. https://www.youtube.com/watch?v=JhCl-GeT4jw&lc=UgwnlCZyElujU5w7AmF4AaABAg.9xgspYR4b0-A-1HPOwpgyz

Matt Welsh (computer scientist). (2014, September 23). In Wikipedia. https://en.wikipedia.org/wiki/Matt_Welsh_(computer_scientist)

Matthew Berman. (2023, December 22). Will I Will Make Programmers Obsolete? [Video]. Youtube. https://www.youtube.com/watch?v=ZV6Sz42l0hY

Travis Media. (2023, November 15). CS Professor Sounds Alarm on AI and Progammers [Video]. Youtube. https://www.youtube.com/watch?v=uuvjOvAdnOk