They cannot be trusted to produce output that works (let alone works well) because they are just statistical models, without any actual understanding of what they produce. That means that you have to carefully review every single line of code they produce, because you don't know where the hallucinations will be. But by the time you do that, you have saved no time at all (indeed, in my experience you lose time), because typing the code was never the part that took time. It was understanding the problem. So if you use an LLM, you spend a bunch of money for zero gain in productivity, or you sacrifice quality and pray there aren't nasty bugs lurking. I certainly think it's fair to call that state of affairs "it doesn't work".

There has never yet been a new model which actually improved over the previous ones. They suck just as much, and in the same ways, as the models of 3 years ago.

> In the past, an engineer who deeply understood the internals of a DB and how memory management worked in Java would be indispensable.

That engineer still is indispensable. Any organization foolish enough to replace such a person with an LLM is going to find itself in deep water when the pile of hallucinations becomes too much to endure.

It is not remotely reasonable to ask "but why didn't he feed it to ChatGPT?". It is pretty silly to assume that ChatGPT should always be consulted.

It's a good starting place. As an analogy imagine someone wanted to look up the definition of a word. If someone wanted to know the definition and they went out and crawled the entire internet, built an LLM, and then asked it for the definition you would wonder why they didn't check an existing dictionary first. I wouldn't consider it silly at all to always check a dictionary or existing LLM first when you want to know the definition of a word.

Wanting to know the definition of a word is not an original problem. Similarly wanting to know what's in an image is not a new problem either.

I understand this is more about the process than the result, but note that: a) his result is completely wrong, he identifies a living land snail as the 100+MYa fossil; b) a conversation with Claude helps with some decent knowledge and provides a few possible candidates, that can be then double checked. c) Claude could have talked the author out of trying to identify a Jurassic fossil against a database of living species.

I hear you, brother. The prospect of being just a babysitter for the clanker strikes me as hell on earth. I'll try to do it - better to be miserable but employed, after all. But if that happens I will mourn the death of a job which I truly loved for the rest of my life.

There is still, to this very day, not a good reason for most businesses to run their workloads in the cloud (startups being a notable exception). So, your argument isn't as compelling as you think.

> BSD license is unrestricted, it tolerates taking open source and closing it, thus always being at risk of things closing down.

There is no such risk. If someone wishes to make a closed source derivative of the BSD-licensed original, it does not deprive anyone of the original. That remains there, just as open as before.

It deprive us of their improvements, while they get to build off other people’s work.

With the GPL, if you want to modify, and built on others work, you have to share.

Share and share alike, vs take if you like share if you like.

It is not "performative hand wringing" to observe that a tool sucks and to reject its use. You cannot, at present, write quality software with AI tools. At best you get something you could've made yourself, slower than you could've made it yourself. Only a fool insists on using a tool when it has been proven to not work.

Unsure what you mean by "quality", but I've personally created projects using AI which work to my satisfaction, and which I didn't make "myself" as I considered it to be too much manual work for not enough benefit. Now using it for even greater things. The tool works fine once you've learned how to use it.

> proven to not work.

what twitter account proved that?

Considering the Catholic Church also teaches that ecumenical councils are infallible, if you propose that Vatican II taught error, then you must also reject a church doctrine which predates that council.

Here's the train of thought:

ecumenical councils where a pope presides with bishops are infallible (something of what you are saying);

Vatican 2 appears to be such a council but also taught error contrary to teachings of infallibility (seemingly impossible);

The only proposition we can think of as to how Catholicism can be consistent without falling in to contraction with the above facts then, would be to conclude that such a "pope" that taught error could not have been a pope in the first place, but was an heretic who then taught those heresies in a false council

(There are some other arguments about popes who can fall in to error or heresy but they are more speculative)

Thus the OP document about AI we wouldn't expect to be reliable from a Catholic standpoint and it goes out of its way to make all kinds of statements not related to AI which we are also critiquing here in this comment chain