Gentoo bans AI-created contributions

Today's LLMs are *heavily* subsidised by the companies pushing them, thanks to the Crypto Bros VC lot pumping money into the NextBigThing(TM) since sliced bread (and of course, crypto!).

Actual costs per query are very high. Not surprisingly, considering after the crypto collapse, nVidias "big market" is LLM based AI!

You forgot one intermediate, which is what they were doing right before LLMs came along - many-to-one chatting. The idea is that each CSR is in four or five (or maybe more, I don't really know the specifics) chats at once, and doing manual timesharing between them (i.e. looking at this chat for a few seconds, then looking at the next one, etc.). Whenever a response is due, they have a series of canned replies they can quickly input (presumably with copy and paste and/or some bespoke software), and most of those are more polite ways of saying "I'll get back to you in a minute," but often phrased in a way that makes it sound like the CSR is doing something other than juggling several chats at once (such as pulling up your file or consulting internal documentation).

You can tell that you're dealing with this kind of customer support based on the following:

1. They subtly encourage you to chat rather than getting on the phone (e.g. the phone has a longer wait time, they won't just give you a phone number and require you to fill out a form first, the phone has one of those awful voice-response menu thingies, etc.).
2. The chat has a queue, but it rarely takes more than a few minutes.
3. Every time you send a message to the CSR, it takes 5-10 seconds (sometimes longer) before they reply.
4. When they do reply, their reply does not meaningfully advance the conversation, and you have to wait another 5-10 seconds for them to give you a real reply.
5. Everything they say reads like it was meticulously proofread and approved in advance by some committee. There are never abbreviations, typos, missing capitalization or punctuation, or other common hallmarks of "real" chatting.

The obvious problem here is that this is much "safer" than using an LLM (see for example the Air Canada fiasco where their chatbot made up a policy and the company was forced to honor it), while simultaneously being relatively cheap (you don't need as many CSRs because they're multiplexed over several chats at once) and providing an experience that is not that much worse than an LLM can provide (the CSRs responses are already grammatically correct and professional, because they've been pre-vetted, so the most the LLM can do is cutting out the "give me a moment" intermediate replies).

We have a colleague who, after interviewing a candidate so dreadful we wondered if they did not show up to the *wrong* interview, once said something like "the light was on so I just let myself in" to give a satirizing summary of what the candidate did (not actually) say.

That stuck and we have many variants now, including "I saw there was light here" or "it looked warm in here" and they all mean the same. I don't know where that colleague got that from, whether it was divine inspiration or one French expression he brought with him.

> I'm a bit disappointed by your restrained reply, I hope I didn't offend you, which was surly not indented. That was supposed to be a harmless joke.

Well, my reply was - while true - intended to be tongue-in-cheek. And of course, when you're worried about your own English, and the LLM gets it worse, of course that's funny :-)

And no, I'm not offended. Speaking four languages to varying ability (and a brit to boot!), I know how hard it can be to express yourself in another language. It's when native speakers can't be bothered to try and get it right that I get upset. After all, English is reputed to be one of the hardest languages to learn (and then you've got American, Strine, Pidgin, and all the others with their quirks on top :-)

Cheers,
Wol

Of course, many people detest that game. But, on the other hand, money. Corporate jobs generally pay well and provide good benefits. If I have to write something bland and inoffensive every day, that's an entirely fair price of entry in my eyes.

* Freudian typo I actually made here: "writhe"

Yes, I understand. Nonetheless, in the domain of the technical references and descriptions, I don't want to see an authors personality shine through, neither polite nor rude nor funny. Just condensed, right to the point technical information. The text object, this subthread started with, was a commit message.

But any other text, even those which should show emotions and personality, can be improved if unintentional errors are removed. LLMs are good with text, they can point to possible errors. From spelling over grammar to even more abstract levels. Spell checkers on steroids. You can very much choose to what degree and how exactly you use the tool for. You don't need to let it rewrite your text into something without a soul.

---
Rephrase the following text into the style of a pirate from a fictional, funny, over the top pirate story. <above text>
---
Arrr, me hearties, listen well! When we be navigatin' the treacherous waters of technical manuscripts and parchments, I yearn not for a glimpse of the author's spirit, be it courteous, churlish, or full of jest. Give me straight, no-nonsense treasure maps of knowledge, I say, starting with the scrawl on the log that kicked off this here discussion.

But mark me words, any other bilge, even that which ought to be brimming with life and character, could stand to lose a few barnacles if it means cleaning up unintended blunders. Us sea dogs have our sea-lore chroniclers - LLMs, the mighty beasts of text, capable of sniffing out a whole array of missteps. From the cursed spellin' errors to the murky depths of grammar and beyond, they're like our spell-checkin' parrots on a hefty dose of grog. Ye be the captain of this vessel, choosing how to steer with this tool at your disposal. There's no need to set a course for blandness, devoid of the soul. Avast! Let your message fly true, without losin' its heart!

So tell it to use a different voice. It can't read your mind you know. If you want it to talk in Southern US slang, it can do that.

There are services out there that will, for a fee, lookup all public communications you've ever made and any private communications you provide, and then create a customised chatbot that will respond with your exact writing style. That's the thing that makes the transformer architecture so revolutionary: it's so easy to customise, anyone can do it.

People sometimes forget it's just a computer therefore doesn't have all the subtle context cues conversations with people have. So if you don't tell it to produce an output appropriate for a Linux kernel commit message, it's not going to figure that out itself. Learning how to configure an LLM to produce output suitable for the context is not very hard, and can be learned by anyone in an afternoon.

The idea you're going to be able to tell someone is using an LLM is on the same level as whether you can tell someone is using a spell checker. If done well, you're not going to notice at all.

And how many other languages have THREE different forms of the present tense? Which one do you use where? That's another classic "this guy is a foreigner" giveaway ...

(btw, khim, the difference between "a" and "the" is easy to explain and hard to apply. If you (could) have several, and don't care which, then it's "a". If you have several and *do* care which you're talking about, then it's "the". "I saw a/the cat in the street". "a" means there are a lot of cats around, and it's a casual comment you saw one of them. "the" means you were talking about a cat earlier, and it's that specific cat.)

Cheers,
Wol

If you don't understand the language, you won't be able to review the output.

If you barely understand the language, it's either going to take a loooong time to review or you're just not doing it right.

It's like writing code vs testing and debugging. Writing it out is the easy part (relatively speaking).