5 ms·

I don’t know how they think that’s going to work. The best AI has been able to do for me is smarter Auto Complete, I doubt it can just fit into larger code base

I don’t know how they think that’s going to work. The best AI has been able to do for me is smarter Auto Complete, I doubt it can just fit into larger code bases without the searching a human can do. I know the job market has been rough which has emboldened these CEO’s but we are starting to see hiring pick up again, and not just Staff+, interns, SDE 1 and 2 as well, which bodes well for 2025.

bdcravens 2y ago
How many developers are doing little more than writing boilerplate? Those are the the kinds of jobs best automated. A few years ago behavior-driven development tried to take this as far as possible, taking stories and then turning them into runnable tests. LLMs can already do a pretty good job of being fed inputs and some context and create running code (for example, give ChatGPT a class file and tell it to write a unit test. It does a pretty good job of it)
- berkes 2y ago
  > How many developers are doing little more than writing boilerplate? Before you go down the road of what can be automated, let's try to answer this. I'm sure there's actual research, but I cannot find it. Anecdotal, but with 20+ years of professional development under my belt, at startups, ngos, enterprise, goverment etc: "insignificant amounts". Yes, some embed-images.sh or some invoice-extractor/src/main.rs - side projects related to a job. Every week some boilerplate like this. But the vast bulk of my work is: - finding out what customers/a spec/stakeholder really mean. And how important it really is (and encoding that in executable specs, preferably) - keeping the old stuff somewhat up-to-date. Fiddling with old Ruby runtimes, Pipenv/pyenv/ or whatever todays flavor is. Nodejs, npm juggling. Docker. Cargo. Deprecations. Security patches. CIs, infra, crashing on updates etc etc. - muddling through codebases with 10+ years of accumulated horrors. Touched by 100+ random freelancers, employees, most lost to the time. Crippling technical debt. Undocumented decisions everywhere. Crucial business requirements encoded only in these three LoCs amongst thousands of lines inside this update_user_projects() function. - Trying to get this old project running again on todays OS/runtimes/etc. - Stitching poorly documented, weirdly (or hardly) architectured, abandoned or outdated libraries together. The way I see it, LLMs-code-gens currently are little more than unpredictable but advanced compilers. Compiling instructions into instructions that computers can execute. BDD could be a good language to write these instructions. And works for new stuff, and much less for existing stuff. "Maintaining existing stuff" is the vast majority of most developers' work, I'd thing. - but cannot back up -.
rvz 2y ago
> we are starting to see hiring pick up again There is a different picture. You are now seeing the same jobs but this time it has a reduced salary + equity and fewer openings. It is not the same as before 2020. For every tech role open, it will be ultra competitive to the point where companies may start hiring olympians for the job that match the so-called "exceptional talent". The truth is, it will get worse before it gets any better.
jitl 2y ago
AI can search your code today. In cursor this is called “codebase indexing”. We have some million(s) lines of code, orders of magnitude smaller than Facebook, but definitely larger than the average startup. We search with AI tools, through Q&A, and for AI-driven code mods. Cursor also exposes LSP stuff to their AI system so it can look at that tool. Sourcegraph Cody is another example of a code AI that can search.
daniel_iversen 2y ago
I think what’ll happen when people say “AI will replace X people” (engineers in this case) it’s not that it’ll for example replace 100% of what 80 people might do (even if they’re kid-level), but rather it’ll make 8 people 10 times more efficient type of thing .. didn’t Tesla’s head of ai or self driving say publicly years ago that co-pilot wrote 80% of his code? It’s not quite apples to apples but you can see how it’s progressing. I know that’s not what Zuck said but I wonder if that might be what he meant?

bdcravens 2y ago

How many developers are doing little more than writing boilerplate? Those are the the kinds of jobs best automated. A few years ago behavior-driven development tried to take this as far as possible, taking stories and then turning them into runnable tests. LLMs can already do a pretty good job of being fed inputs and some context and create running code (for example, give ChatGPT a class file and tell it to write a unit test. It does a pretty good job of it)

berkes 2y ago
> How many developers are doing little more than writing boilerplate? Before you go down the road of what can be automated, let's try to answer this. I'm sure there's actual research, but I cannot find it. Anecdotal, but with 20+ years of professional development under my belt, at startups, ngos, enterprise, goverment etc: "insignificant amounts". Yes, some embed-images.sh or some invoice-extractor/src/main.rs - side projects related to a job. Every week some boilerplate like this. But the vast bulk of my work is: - finding out what customers/a spec/stakeholder really mean. And how important it really is (and encoding that in executable specs, preferably) - keeping the old stuff somewhat up-to-date. Fiddling with old Ruby runtimes, Pipenv/pyenv/ or whatever todays flavor is. Nodejs, npm juggling. Docker. Cargo. Deprecations. Security patches. CIs, infra, crashing on updates etc etc. - muddling through codebases with 10+ years of accumulated horrors. Touched by 100+ random freelancers, employees, most lost to the time. Crippling technical debt. Undocumented decisions everywhere. Crucial business requirements encoded only in these three LoCs amongst thousands of lines inside this update_user_projects() function. - Trying to get this old project running again on todays OS/runtimes/etc. - Stitching poorly documented, weirdly (or hardly) architectured, abandoned or outdated libraries together. The way I see it, LLMs-code-gens currently are little more than unpredictable but advanced compilers. Compiling instructions into instructions that computers can execute. BDD could be a good language to write these instructions. And works for new stuff, and much less for existing stuff. "Maintaining existing stuff" is the vast majority of most developers' work, I'd thing. - but cannot back up -.

rvz 2y ago

> we are starting to see hiring pick up again There is a different picture. You are now seeing the same jobs but this time it has a reduced salary + equity and fewer openings. It is not the same as before 2020. For every tech role open, it will be ultra competitive to the point where companies may start hiring olympians for the job that match the so-called "exceptional talent". The truth is, it will get worse before it gets any better.

jitl 2y ago

AI can search your code today. In cursor this is called “codebase indexing”. We have some million(s) lines of code, orders of magnitude smaller than Facebook, but definitely larger than the average startup. We search with AI tools, through Q&A, and for AI-driven code mods. Cursor also exposes LSP stuff to their AI system so it can look at that tool. Sourcegraph Cody is another example of a code AI that can search.

daniel_iversen 2y ago

I think what’ll happen when people say “AI will replace X people” (engineers in this case) it’s not that it’ll for example replace 100% of what 80 people might do (even if they’re kid-level), but rather it’ll make 8 people 10 times more efficient type of thing .. didn’t Tesla’s head of ai or self driving say publicly years ago that co-pilot wrote 80% of his code? It’s not quite apples to apples but you can see how it’s progressing. I know that’s not what Zuck said but I wonder if that might be what he meant?