That nicely sums it up.

But I am sure some people will think this is just excellent and they totally have to do it.

What kind of utterly crappy CS program did you attend?

Well, if you think running a local toy and doing your own training for that is good, sure, go for it. And good luck.

Isolation is not actually an OO thing, because that isolation is fake. All you actually get is weak namespace separation and that you can get with other techniques as well.

That "tiny" part has stopped bein tiny a while ago. The likelihood of that not working when it gets slapped together is growing all the time and you may want to look up "cost of cyber attacks" as a reference.

To be fair, a team of competent experts can work with any stupid management "methods" as long as they can basically ignore them. I have seen that with a pretty good team being forced to adopt Agile. They just made the "sprints" 6 months long and proceeded to use what they knew worked.

Well, yes. A long time ago. When I studied CS, they had us take the Calculus and Linear Algebra for Mathematicians (same lecture, same exam). That did weed out about 50%. But it did nothing for CS/IT specific skills. Turns out mathematics is actually not that close to them. Friend of mine failed CS and then went on to get a Math Master's without any problems.

I do not think so, at least not with LLMs. The problem is that an LLM does not even begin to reach the level of reliability of a typical compiler. LLMs hallucinate, create bloated code, leave out functionality that was requested, have no clearly specified functionality, cannot deal with subtleties, etc. Hence LLM-generated code needs to be reviewed carefully and that is probably in the end more expensive than having somebody competent write the code from scratch. There are also indications that reviewing LLM code is mind-numbingly boring and people will burn out doing it, while even now there are not enough people that can do it in the first place.

Also note that the attempts of replacing programmers are _very_ old. When I studied CS almost 40 years ago, the 5GL project had just resoundingly failed. Its idea was to have program specifications in constraint logic and then have the machine create code for that. The problem was not that this did not work. The problem was that creating a specification on the level required was much harder and more time-consuming that writing the code directly with a typical informal or semi-formal specification. And, quite frankly, I see the same problem with LLMs as translation engine, as soon as you need actual production quality code. For producing low-quality slop, LLMs are fine. But that slop will not survive in the real world, especially as the one form of code LLMs are really good at creating is attack code, which does not need to be reliable, secure or maintainable. Hence low-skill hackers got a major upgrade here, while LLM use makes the defenders even weaker.

Indeed. Sure long-term, some jobs will get reduced in numbers, but the only real area where that will happen is low-skill paper pushers. Hallucinations and slop will kill it for the other uses and these two cannot be fixed for LLMs.

Not at all. Getting rid of the semi-skilled is how an engineering discipline eventually grows up. The second way LLMs help with that is that incompetent attackers can not get to the level of a somewhat mediocre expert using LLMs. That is going to drastically increase the number of attackers and put a lot of (very desirable) additional pressure on makers of software. Of course, "slop" code has basically no chance of survival.

Hence while things are unpleasant at the moment, I think that longer-term this is a good thing. Obviously, LLMs in their current form will not survive either, completely non-viable economically. Once the real costs have to be paid directly, they will find a few niches but the deranged hallucination of general applicability will be over.

That said, now may be an excellent time to start studying CS/IT for somebody with the right talents and interests.

Outsourcing is actually an excellent comparison. Having had the misfortune of reviewing code from supposedly "quality" outsourcing several times about 15 years back, I now recognize this as an early form of "slop". Code produced with no insight, exactly matching the spec and the spec not being very good (which is normal).

The legal angle for LLM "code" is going to be even funnier. First, it does not have your copyright, as LLMs results do not qualify for copyright. But second, it may have somebody else's copyright, if the LLM produced training data too directly! One of the reasons the GCC project has prohibited LLM code completely and the Linux Kernel requires it to be clearly labelled (so they can rip it out later if needed).

Probably. On the other hand, any serious software project is like 10-20% initial coding and the rest is what makes it a success or failure. Well, with LLM use (and LLM supported attackers) the coding itself will probably also become a serious source of failures now. It is still hilarious how incapable LLMs are at producing code. Not that many supposedly professional coders are any better.

He is right that OO never really delivered on its many promises. I regard its as mostly a failure these days and my perception is that doing OO often only serves to increase project cost. "OO for everything" is one of these bad ideas that refuse to die in low-insight environments. That said, OO has its place but it is not everywhere.

Yes. That Agile approach has really never worked. Competent people usually hate it and incompetent ones just produce trash a bit faster, which is not a win.

The fact of the matter is that when you produce stuff more complex than simple web interfaces or business logic, you cannot accommodate changing specifications during the project. It will mess up everything. And, funny thing, that has been known for a long, long time. It is just that we have far too many incompetents in the industry.