Why doesn't copilot lead to layoffs at coding jobs? [closed]

Question

Github argues that "developers who used GitHub Copilot completed the task significantly faster–55% faster than the developers who didn’t use GitHub Copilot."

How come we're not seeing a trimming of the workforce anywhere?

Answer 1

As a software developer, looking at my weekly schedule, "coding" is actually just a tiny part of my work. Things I am spending significant amount of time on are:

• Talking with domain experts and customers about which problems we need to solve.
• Reading bug reports from our testers and users, and analyze our code to find out why and where these problems occur.
• Coming up with possible solutions to the aforementioned problems.
• Discussing with domain experts and other developers which solutions would be best.
• Making time and cost estimates for our solutions.
• Explaining to stakeholders why these estimates are realistic.
• Actually implementing the solutions by writing code.
• Documenting how problems were solved in the end.
• Reading the documentations from other developers to keep up-to-date with how our software works and how it is implemented.
• Educating myself and others about the latest technological innovations and how we could use them to solve problems more efficiently.

So a tool that would make me 55% faster at writing code (assuming it were true) would only provide an insignificant boost to my overall productivity.

Answer 2

It may come as a shock to you, but companies who are advertising a product don't always adhere 100% to the truth in their advertising.

So, one possible explanation is that GitHub's advertising is simply not true.

Another problem is inherent in how LLMs work. LLMs can only regurgitate what has already been written elsewhere. LLMs have no intelligence, they have no creativity, they have no problem-solving skills. An LLM is essentially "spicy autocomplete", nothing more.

Just like regular autocomplete, LLMs sometimes generate garbage. They "hallucinate" or "confabulate" stuff that is incorrect. Remember, LLMs are trained on existing data; Copilot is trained on existing code – and, in line with Sturgeon's Law, most code out there is garbage.

So, even if a developer using Copilot completes the assigned task significantly faster, the result still has to be checked for correctness. With the added complication that the person checking the code can't ask the author for clarification, because there is no author.

Furthermore, LLMs can only regurgitate what they have already seen. This means, there is a chance that the output might violate someone's copyright. (There are documented examples of Copilot copying entire classes verbatim.) So, in addition to checking the code for correctness, the code also has to be checked for legality. Due to the nature of how an LLM works, it is impossible to know where that code came from, so that involves a significant amount of research.

So, even if GitHub's claims are true, those efficiency gains are at least partly offset by the additional work created. If you add to that the possibility that the advertised claims may be inflated, it should not be surprising that we don't see mass layoffs.

Lastly, your entire question hinges on the assumption that companies are actually using Copilot. I know many people who refuse to use LLMs and Generative AI in general. One quote I read said: "I want AI to do my dishes so I can focus on my creativity, not AI to do my creative work so I can focus on doing the dishes."

Answer 3

Even if we take their statistic at face value:

The automation vs labour debate has been going on since the invention of the threshing machine. There's a kind of notion that companies have X amount of work to do and the number of employees they have is dictated by that.

However, in many cases, it's not that their labour force is dictated by their output, it's that their output is limited by their labour costs (and other factors of course).

In other words, usually (though not always) if their employees take half the time to do the work, they'd rather - as it's more profitable - to double their output than half their employees.

Answer 4

This has been said so many times since computer programming has been a job...

In the 1950s when the first computer language was made there were fears that the 10x speed improvement that programming languages gave over machine code was going to put 9/10 programmers out of a job. It is now 70 years later, I don't think that is what happened. If anything, the more accessible/easy it is to program the more programming jobs have been created. Sure some people who refused to learn the new tools lost their jobs, because they were unwilling to adapt.

But the skills that are most required in programming aren't the ability to throw weird symbols/code words onto a piece of paper. It is to analyze and understand a problem, and then write out a precise enough specification that defines the problem that a literally minded idiot savant (computer) can understand it.

Answer 5

It just ain't that good.

Here is what Linus Torvalds thinks: https://www.instagram.com/reel/C7l4MQ7haxL/ Love him or hate him, his opinion is worth respecting.

AND software just isn't written that well. Who really wants to use the first or new version of anything?

So, maybe AI can help do boilerplate quicker and give the coders some extra time to ensure their work product is better? (Wishful thinking maybe on my part?)

This is my 50,000foot, 40 some years experience of software experience that is my HO.

And yes, I let copilot do some of my .net boilerplate for me.