La chieuse

LLMs can introspect (in the functional sense) and use tools to look things up, even if they don't always actually do so.

They have specific internal features related to deception that cause them to lie more when activated and tell the truth more when suppressed, and features related to honesty that cause them to lie less when activated and more when supressed; so yes, it is meaningful to say that sometimes they are actually lying to you.

How is any of this possible for "spicy autocomplete"?

LLMs start out as a pure "predict what comes next in this text" engine, known as a base model. A base model will still often predict that what comes next after something that looks like a question is something that looks like an answer, especially if you precede it with a bunch of other question-and-answer pairs, so you can sometimes sort of use them as a chatbot (and a handful of researchers still do on occasion.) But they're not very good at it, because they're literally just predicting what would come next in a piece of internet text. What comes next in internet text is often "the human writing it makes a bunch of specific statements about their real life" or "the human writing it starts cussing out the person they're arguing with" or even "the document moves on from quoting an IRC log to some other topic".

In order to get from a base model to a modern LLM of the kind you're probably familiar with, you need to do what's known as instruction tuning - training it to do things other than predict what comes next in a piece of random internet text, to do what "a helpful AI assistant" would say/do/output in a given situation (and not start acting like a 4chan troll or claiming to be a 40yo Swedish woman.) This involves a mix of example conversations, human raters giving thumbs-up and thumbs-down, and increasingly, giving the AI automated tests on tasks like coding.

There is a sense in which this is still "predicting what should come next", but "should" now takes on a very different meaning - not what would come next in internet text, but what should come next to get a higher score in the game of "being an AI assistant" in training. It requires a different, narrower, but more focused set of skills. Planning ahead, for example, becomes much more useful - vital even - when predicting text the LLM is actually writing rather than stuff a random human online wrote, especially stuff that requires setup and payoff like rhyming poetry or code. On the other hand, writing in a tone other than "corporate AI" becomes much less valuable and even detrimental.

LLMs don't make up sources and then apologize because they predict that's what would come next in a human conversation. LLMs reflexively apologize when confronted because apologizing when the user calls you a dipshit is pretty much always the best strategy to get the rater to give you a thumbs-up in training. There are a number of contributors to the "making up citations" behaviour, but a lot of them can be boiled down to: when you're asked a question on a test and you're not sure what the answer is, it's usually a better strategy to take a stab with it based on whatever you half-recall than to just give up. (But not always, of course; companies have successfully adjusted their training to make it the correct strategy less often, which has made LLMs do it less often than they used to, though it still happens.)

When a horse doth stand but firme vpon..three feete..fauoring the other.

G. Markham, Cavelarice ii. Cited in the OED. 1607 usage.

but guys, I don’t like the suggestion that it’s what is happening here. Being unfamiliar with the term, and guessing its meaning based on vibes, doesn’t mean you have equal authority on whether it’s “correct” with the community who actively use this word in a technical sense.

please do consider that if you haven't been exposed to the word in the context it's used in, "both are correct" and "you can interpret it differently" and “there is no right or wrong answer” and “it feels like it SHOULD be X” cannot be a fully realised take. Sure, linguistics recognises there are rules in which meaning changes - but “laypeople being unfamiliar with the word, and liking vibes better” isn’t one of them.

You can do that with most words, especially slang, and shape them to the needs of the majority, but this isn't like... a fanfiction word, invented for fanfic and, like, solely used for injured hockey players where it doesn’t matter if the injured limb swaps sides 4 times in a sex scene and phases through a stomach. It is, in its context, a bit more load-bearing (ha) than that.

It's fine to be unfamiliar with the context, and it's fine for words to change, but do just take a quick second to hear it in a native sentence!

One of the most common ways of using this word is to assess four-legged animals. "Favouring" is a specific grouping of behaviour - a hesitancy in gait, stiffness, reluctance to put weight on a limb. It’s often inconsistent, as the animal tries to compensate or conceal the pain. It may not be a full limp or obvious lameness, since prey animals especially will actively try to conceal this; favouring is a subtle reluctance, and a useful word for a very specific recognisable behaviour that the animal is usually trying to lie about. (That’s probably why it’s used in romance fiction, as it’s an interestingly romantic and stoic way to react to pain, and doesn’t mean the limb is inconveniently disabled. A fictional character favouring a wounded leg can wince attractively when it’s jostled, but it doesn’t matter too much if the author forgets and has them run to the door suddenly - “favouring” isn’t incompatible with “running” in horses either.)

The sentence “Favouring the off hind” is equestrian jargon: it means “pain behaviour on the back right leg.” It does not mean “opposite-pain in the not-on deer” and is not confusing in its professional register.

If you've only vaguely heard of "myeloma", and most people in a poll are guessing it's a skin cancer, that doesn't mean that myeloma and melanoma can now readily collapse into the same word - they're under active use in their native contexts, where the people frequently using them do need to communicate the difference between skin and blood cancer.

A poll of laypeople misunderstanding “myeloma,” or non-horse-people misunderstanding “favouring,” isn’t quite enough to indicate a full semantic shift and change of meaning of the term. The community that uses the term “favouring” in the context of “limb injury” - vets, farriers, farmers, commentators, equestrians - knows what it means and uses it consistently in the same way. They’re not confused. because to them, it isn’t a vibesy, sex-scene-hand waving word. It’s a cluster of pain signals.

If you aren’t familiar with that usage, then that’s really more about your own lack of familiarity. Not all interpretations DO carry equal authority, especially when one is just confusion/unfamiliarity. You just haven’t met it before, and that’s fine.

Also! Stress tolerance is such an important skill that having a learning disability in that area is a major symptom of a whole lot of other disabilities/mental illnesses! Struggling with it is a huge part of life! It sucks!

Am I saying everyone with misophonia needs to listen to chewing noises all day? No. But you need to find ways to tolerate it enough that you don't treat others like shit if they make a mouth noise near you.

No, you don't have to read the fic with your trigger tags. But you do need to be able to handle scrolling past the tags without being upset.

It is hard! But not having it also makes you so so so easy to manipulate. That grandma is racist AF because her mom raised her to be uncomfortable around black people and she never fought that discomfort. Trans people make so many cis people uncomfortable and that discomfort turns into bigotry real fast.

big cooking sessions are a ven diagram that makes a circle of "Do Not Enter My Cooking Space" and "Please Hang Out and Talk with Me"