Mark Zuckerberg says AI could soon do the work of Meta's midlevel engineers

Not so much to tell: example: some company has been running for the past 20 years on some erp and things have been attached to it and built around it over that time: suddenly it stops working for no clear reason and the company basically is not functioning fully or at all, losing whatever amount per hour. The original lead dev is gone or (in a recent case) died years ago and there is not much documentation, no issue tracking or even version management (it's a little HN myth that non IT companies all use that stuff as well; they don't, including some IT/SaaS ones we encountered). We get asked to fix it and make it run again. Sometimes we get ask to figure out a plan to fix it long term as well.

> when you can direct your 24/7 mid engineers to just replicate the Salesforce offering

What happens when AI devs can't manage it anymore and it suddenly falls apart, with every new release shipping more critical bugs than it fixed?

I think you better pray that it happens sooner rather than later, because the larger the mountain of crap code becomes, the more it's going to hurt when it unravels, and less likely it is that it's fixable. Do you then just shut the company down?

> 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 -.