3 ms·

It's interesting to see people who have been working as software engineers worrying that codegen LLMs will take their job. In the past week the amount of work t

by lsy 4y ago

It's interesting to see people who have been working as software engineers worrying that codegen LLMs will take their job. In the past week the amount of work that consisted of writing well-defined code was pretty much zero, and the actual core logic I have written lately has been almost mindlessly trivial. I spent much more time discussing requirements, arriving at consensus on where components should live, determining whether change requests should go forward, determining how changes affected latency or disk space, and debugging nuanced issues that required multiple engineers to slowly reveal the problem. If your programming job consists of simply receiving a greenfield spec and banging out the code for it over and over, LLMs might be an issue. But most jobs are not like this. The bulk of the work, as always, is in dealing with the people at either end of the process. And knowing what the computer is doing and how it works will continue to be invaluable when an LLM writes logically perfect code that happens to run slowly or do the wrong thing once it makes contact with the real world.

vsareto 4y ago
Hot take, but if you're doing a lot of soft skills work with non-developers, you might be carrying the team too much. You're very much right that those jobs won't go away because that would mean people around the developer would then have to do more work and be more technical. What would be great is for senior ICs to push into non-IC roles (ideally: ICs can use AI, so you need less ICs but this is unproven). Technical skills and knowledge applied to writing requirements cuts down on the number of conversations around those requirements. Then you get closer to a "greenfield spec -> bang out AI-generated code" situation because you're not slogging and negotiating through requirements written by non-technical people.
alexpotato 4y ago
This reminds me of a quote from Joel Spolsky in the early 2000's but he was talking about offshoring: "There is always a global shortage of developers who know what they are doing" At the time, I wasn't sure if this was the case but two things have helped me believe that this is true: 1. a few years ago there was a stat going around on HN that something like less than 5% of the US population could even write a Hello World application in ANY language 2. 15+ years of working in tech and 5+ years of being a hiring manager have shown me that even people with great looking resumes often fail the Fizz Buzz test
- steve_adams_86 4y ago
  I suppose what people are afraid of is that the devs who can do Hello World + pass Fizz Buzz might use AI to orchestrate sophisticated systems which actually are effective in reducing a lot of menial work, which could eliminate a non-trivial number of jobs through sheer reduction of necessary labour/brain power. I agree though, off-shoring could have been the cheap labour which displaced a lot of devs, but I actually found a lot of work cleaning up off-shored applications and made great money doing it. It's perfectly possible that some horrible amalgamations of software will be puked out by AIs for a few years and ultimately create as much work as it displaced.
onlyrealcuzzo 4y ago
At many of the FAANGs, it's not uncommon to spend 4 hours of testing for every one 1 of programming. Considering how important it is to not have bugs - we're extremely unlikely to have an LLM writing code at an acceptable level of quality, and anyone being qualified to verify that quality beside an eng, before I'm retired. But I'm confident many 22 year-old new grads will reply to tell me how wrong I am, and how I'll be starving to death under the burden of unemployment by next week.
- pfdietz 4y ago
  How good are the LLMs at creating tests?
lonelyasacloud 4y ago
In one of GPT4's release demos, it is shown being fed a photograph of a hand-drawn web page. It then image analyses that, works out that it's a picture of a web page and turns it into a working web page (https://youtube.com/live/outcGtbnMuQ?feature=shares&t=972 https://youtube.com/live/outcGtbnMuQ?feature=shares&t=972). So here's the thing; what in that list of non-codegen tasks is more complicated than that which GPT can already automate?
pfdietz 4y ago
> If your programming job consists of simply receiving a greenfield spec and banging out the code for it over and over, LLMs might be an issue. I think LLMs are going to be very powerful for brownfield coding too.

by lsy 4y ago

vsareto 4y ago

Hot take, but if you're doing a lot of soft skills work with non-developers, you might be carrying the team too much. You're very much right that those jobs won't go away because that would mean people around the developer would then have to do more work and be more technical. What would be great is for senior ICs to push into non-IC roles (ideally: ICs can use AI, so you need less ICs but this is unproven). Technical skills and knowledge applied to writing requirements cuts down on the number of conversations around those requirements. Then you get closer to a "greenfield spec -> bang out AI-generated code" situation because you're not slogging and negotiating through requirements written by non-technical people.

alexpotato 4y ago

This reminds me of a quote from Joel Spolsky in the early 2000's but he was talking about offshoring: "There is always a global shortage of developers who know what they are doing" At the time, I wasn't sure if this was the case but two things have helped me believe that this is true: 1. a few years ago there was a stat going around on HN that something like less than 5% of the US population could even write a Hello World application in ANY language 2. 15+ years of working in tech and 5+ years of being a hiring manager have shown me that even people with great looking resumes often fail the Fizz Buzz test

steve_adams_86 4y ago
I suppose what people are afraid of is that the devs who can do Hello World + pass Fizz Buzz might use AI to orchestrate sophisticated systems which actually are effective in reducing a lot of menial work, which could eliminate a non-trivial number of jobs through sheer reduction of necessary labour/brain power. I agree though, off-shoring could have been the cheap labour which displaced a lot of devs, but I actually found a lot of work cleaning up off-shored applications and made great money doing it. It's perfectly possible that some horrible amalgamations of software will be puked out by AIs for a few years and ultimately create as much work as it displaced.

onlyrealcuzzo 4y ago

At many of the FAANGs, it's not uncommon to spend 4 hours of testing for every one 1 of programming. Considering how important it is to not have bugs - we're extremely unlikely to have an LLM writing code at an acceptable level of quality, and anyone being qualified to verify that quality beside an eng, before I'm retired. But I'm confident many 22 year-old new grads will reply to tell me how wrong I am, and how I'll be starving to death under the burden of unemployment by next week.

pfdietz 4y ago
How good are the LLMs at creating tests?

lonelyasacloud 4y ago

In one of GPT4's release demos, it is shown being fed a photograph of a hand-drawn web page. It then image analyses that, works out that it's a picture of a web page and turns it into a working web page (https://youtube.com/live/outcGtbnMuQ?feature=shares&t=972 https://youtube.com/live/outcGtbnMuQ?feature=shares&t=972). So here's the thing; what in that list of non-codegen tasks is more complicated than that which GPT can already automate?

pfdietz 4y ago

> If your programming job consists of simply receiving a greenfield spec and banging out the code for it over and over, LLMs might be an issue. I think LLMs are going to be very powerful for brownfield coding too.