Ask HN: Why does ChatGPT love the word "eager" so much?
I have used English as a second language for many years, but I have never heard of this word.
Suddenly, while talking with ChatGPT, it appears almost in every other conversation...
Is it because I talk (sometimes) about PHP/Laravel/Eloquent, and it somehow fixed the "Eager loading of relationships," or is the language changed, and I did not notice?
Did you notice other words like this?
I found this article about the word "delve":
https://www.theguardian.com/technology/2024/apr/16/techscape...
Little off-topic: Lastly, I learned that children in Portugal have started to speak the Brazilian variant of Portuguese, as videos from there are flooding the Internet. It is interesting how technology affects our lives in more surprising areas. One of the funniest things I've ever heard/read about chatGPT "writing" is that it writes like a highschool student trying to inflate their word count on an essay. It typically uses a lot of words to say very little, and the style is hard to un-see once you recognize it. ChatGPT will write in essentially any style you ask it to. The default behaviour is an artifact of RLHF training and the hidden pre-prompt, not a reflection of the capabilities of the underlying model. A pre-prompt along the following lines will radically change the nature of the output: "You are a subject matter expert in x. You communicate in a terse, information-dense manner. You make clear, strong arguments without vacillation." Yeah Open AI, Anthropic etc are all intentionally making LLMs sound more robotic than they are capable of sounding. It's pretty awful because the effects it has on fiction writing is disastrous. >> like a highschool student trying to inflate their word count on an essay. So the chameleon is a bad chameleon because it looks too much like a leaf, too much like the thing it is attempting to imitate? A few years ago I was tasked with editing the performance reviews for my unit (military). Every supervisor had a finite number of characters in which to describe each soldier's performance. I went through and removed all the extra/useless words. Oh the anger! While supervisors agreed I had improved their writing, they now felt obligated to fill up all the blank space I had created. I'm glad you asked about chameleons! Determining whether a chameleon is good or bad can be very important for lizard rankings. As everyone knows, a chameleon is able to camouflage to look more like a leaf. This can help protect it from predators! But sometimes looking like a leaf can go too far.... [and so on]. I hooked ChatGPT up to a speech recognizer and far field mics etc, trying to build my own alexa, and I had to add "Please be terse" to the prompt. And that wasn't enough, so I said "Please limit yourself to as few words as possible to convey the answer. Be terse. Try to keep your speech very short." before I finally started getting reasonable replies. Your example showed it that by "terse" you meant explain the same thing three different ways. ;-) I have had good luck telling it to use simplified technical English. It's called thinking out loud. Gets you better results. Try it out yourself. ChatGPT doesn’t think Prove to my satisfaction that you think. ;-) I think that comments like this with emoticons are vague in intention and difficult to reply to especially given parent context - but basically, I think because I am thinking. No proof exists within the LLM flavor of AI products that they do any form of thinking as we understand it, because we haven’t even defined what thinking is. Charlatans tend to use vague definitions to advance whatever agenda they have, and when a subject is already poorly understood, it is prone to wild, unsubstantiated (and prone to misinterpretation) claims. Whatever you want to make of AI’s emergent abilities or whatever dumb AI meaningless buzzwords are the trend right now, you have to first prove to me you know how I think before I will accept your claim AI knows how to and as far as I’m aware that’s still an open question. Hope that explains my viewpoint. > are vague in intention Just poking fun at a flaw in your reasoning. You think you are thinking because of your subjective experience. However, you say that something else cannot be thinking, despite the fact that you cannot share its subjective experience- therefore ChatGPT could say the same thing about you. This is turning into the type of pedantic argument I find exhausting on here and gets into complex philosophy theories of mind, but basically, I believe the flaw in your reasoning is that you are assuming ChatGPT has a subjective experience. To me that's absurd to the point of comedy and I haven't seen any evidence of it, but I'm completely uninterested in debating that. > the flaw in your reasoning is that you are assuming ChatGPT has a subjective experience No I'm not. I don't think ChatGPT thinks either. But I can express that claim without making untestable, unscientific, circular claims. Your explanations of why it can't be conscious have all boiled down to "because it isn't [synonym for conscious]". But you are conscious "because you do [thing synonymous with consciousness]". I'm not trying to attack you as a person, just poking fun at this all-too-common style of reasoning. I apologize if I offended you by doing that. https://starwars.fandom.com/wiki/Therefore_I_Am:_The_Tale_of... Only old heads will know this but this is a Star wars old canon short story from a collection of short stories that explores the topic of what consciousness and thoughts mean and in the context of a malevolent AI. “I think, therefore I am.” However, “think” is a bit ouroboric in definition and the reason this is an interesting question. Although "eager" isn't called out, a recent study of academic publications shows that the use of LLMs can be measured through word frequency analysis [1], finding certain words are disproportionally represented: > We study vocabulary changes in 14 million PubMed abstracts from 2010–2024, and show how the appearance of LLMs led to an abrupt increase in the frequency of certain style words. I don't want to look for the source of analysis right now, but I recall reading a study demonstrating that a large part, if not most of the word frequency shift was caused by RLHF training done on data predominantly generated by people hired from lower income English-speaking countries which simply have a different dialect of English with a noticeably different frequency of certain phrases and expressions, so e.g. at least some versions of ChatGPT got RLHF-trained to speak more in a Nigerian English dialect. Since there isn't a single English (English learners generally get informed about the choice of UK vs US English only, but most English is spoken outside of UK and USA in other places and other dialects), but multiple different Englishes, any English speaker will probably find something to be surprised by, and there is an economic incentive to get data from people other than the relatively expensive native speakers of UK or USA English. There wasn't a study or analysis. It was just lazy speculation that felt good because it could be bound up in a "evil white countries exploiting the developing world" narrative. Where exploiting was "paying to do a job". It was submitted as https://news.ycombinator.com/item?id=40623629 Again, there is effectively zero real data showing this. Further, RLHF isn't likely to reinforce such word selection regardless. A more logical, likely scenario is that training data is biased heavily towards higher grade level material, so word selection veers towards writings that you find in those realms. > It was just lazy speculation that felt good because it could be bound up in a "evil white countries exploiting the developing world" narrative. Where exploiting was "paying to do a job". Exploitation like that is in fact happening (see pretty much everything having to do with social media content moderation and RLHF to avoid disturbing content. Also "paying to do a job" is not the moral panacea you seem to think it is. tinfoil had theory: they implanted watermarks already, so that AI generated text can be flagged for future training runs or as a service, such that some phrases are coaxed to become statistical beacons. That's not really a tinfoil hat theory. That's been possible for some years and OpenAI reportedly does watermark their outputs, and can detect it. They just haven't released it as a service because it'd annoy all the users who are using it for cheating :) I believe that if that was possible to do on purpose, they wouldn’t have so much trouble preventing the LLMs from talking about things they shouldn’t. Yeah I would like to see some evidence of this too. It's just asserted as truth in the article. Delve doesn't seem like a particularly unusual word to me, especially in the context of scientific abstracts, and LLMs could totally learn random weird things. How common is "it's important to remember" in Nigeria? Wait, why wouldn’t RLHF influence word choices? I didn't say it wouldn't (or rather couldn't), I said it was unlikely for the selected hypothesis given standard training data vs RLHF iterations. then again, most history consists of whitewashing back when northern countries were exploiting everywhere else in various ways: imperialism, colonialism, neocolonialism, capitalism, financialization,... typical people prefer to pretend this is simply "order" and "progress"; seemingly blind to their own ideological baggage like fish in water Yeah, right. ChatGPT was trained on Pidgin English dialect. Have a look at BBC translation to get a taste, and tell me its not hoax:
https://www.bbc.com/pidgin The window of time where word frequency of chatgpt's favourites and usage of chatgpt is closely related is rather small I think. Academic language has a number of 'marker' words that are basically just style and will be more or less copied once you read many papers. 'Rigorous' is a general example, but most fields have their own. If many papers you read while writing your own paper use words like 'delve', you will be much more likely to use it yourself. On another note, while the paper itself is pretty cool, in discussions on it I thought people where kind of looking down on using LLM's to help you write. There's a philistine moat in many fields around writing style. While writing well is in my experience correlated with paper quality, it is not predicated by it. And introducing tools that help people write more readable papers is probably a net benefit overall. I wonder why some words are overrepresented. Isn't the whole idea of language models to model word distribution as close as possible? Does it have something to do with RLHF? Or it's the training data? Language models would be fairly useless for most people if they accurately modelled the source distribution, no better than autocomplete. In fact, they were fairly useless when they modelled the source distribution, that's why ChatGPT was an instant hit whereas GPT-3 was mainly only interesting to other AI reasearchers. What made LLMs suddenly interesting was that the responses were much more like answers and much less like additional questions in the same vein as the prompt. >In fact, they were fairly useless when they modelled the source distribution, that's why ChatGPT was an instant hit whereas GPT-3 was mainly only interesting to other AI reasearchers. I had a bot which used the original GPT3 (i.e. the completion model, not the chat model) and its answers were pretty decent (with the right prompting). Often even better than GPT3.5, whose answers were overly formulaic in comparison ("as an AI language model...", "it's important to ..." all the time) I think that means you would count as "another AI developer" ^_^; To what extent can this style be overcome by prompting? If it can be overcome in existing models, it’s probably going to involve different aspects including vocabulary, style, and organization. Can it just be a frequency illusion, where you tend to notice a new-to-you phenomenon again and again at first https://en.wikipedia.org/wiki/Frequency_illusion?wprov=sfla1 ? Eager isn't an especially uncommon word (eg "eager beavers" is a somewhat common saying), even though it's not used in most convos. I feel like "delve" is a YouTube phenomenon (as in "let's delve into this topic") as a weird proxy for "deep dive". Maybe a side effect of D&D's resurgence over the last decade, where it's often used to describe small adventures/dungeons...? This, and ChatGPT was trained on a lot of web content produced by businesses, so in my view this may lead it to overuse "buzzword" words such as "delve," "deep dive," "leverage," "optimize," "synergy," "impact," and others. My new frequency illusion is the phrase "ride or die". I'm 32 and I had never heard that in my life until I watched that show Cruel Summer and now it seems like I hear it everywhere, mostly on tv/ads/podcasts and almost never when talking to people I know. I refuse to say it because I don't fully understand what it means and when I do hear it, it always sounds a little cringey. A Will Smith movie called ride or die came out in May and grossed $400M. Probably explains the jump in that phrase. I had heard it before but you only hear that phrase every couple years or so. "Paramount" has been that word for me. Last year it seemed like every response from ChatGPT included it. So I just heard about Frequency Illusion earlier today, and now I see it here again. Such meta. (Not kidding, from today’s NYT crossword column: https://www.nytimes.com/2024/08/11/crosswords/daily-puzzle-2...) > Eager isn't an especially uncommon word I'd say it's very common, at least in my part of the US. It's one of the words I hear on a daily basis, anyway. "Delve" used to be a very commonly used term before "deep dive" largely replaced it. I'm sure there are a whole lot of writings online that use "delve" because of the time period they were produced in. As a graybeard, I'm personally still much more likely to say "delve" than "deep dive". I think niche words like "delve" get replaced by phrases like "deep dive" to accommodate ESLs, especially in big business and software development. "Delve" is the word to use, of course, but if you're going to lose (or annoy/insult) your ESL audience by using fancy words, maybe just being accommodating has value. Seems like "chef's kiss" is replacing "icing on the cake" or "cherry on top", although I think it really means "stamp of approval", so that one has been bugging me since I hear it all the time now, it seems. As a former academic in tech I'm tickled that you think 'a deep dive' is normative while 'to delve' strikes your ear as strange. No judgment! I'm delighted, however, that language is so supple ("leverages domain-local synergies") Probably a generational thing? OP cited youtube as 'delve'ers, which skews young, so I'm guessing it's that your cognitive 'ear' is tuned to the technosphere All of that’s way more common in (American) business English than other registers, I’d say, including “eager”. ChatGPT itself is eager, or plays the role of an eager companion. Why? - it's conversationally-aligned with dumping large amounts of information - it's an easy emotional state to hold unilaterally (without factoring in the other participant) - it's unlikely to offend or cause a PR nightmare - it's flattering! i compiled a list of overused words you can stick in a "please do not use these words unless you absolutely have to" antiprompt https://gist.github.com/swyxio/8ac555e88ad153764051012d2db27... (we use these in ainews summaries so that we dont delve too much https://buttondown.email/ainews) I feel like ChatGPT is specifically 'eager' to help me, that may be an instruction to the LLM that overflows somewhat in to it's answers. But generally, 'eager' isn't particularly rare in English. It’s word use fashion, which LLM has influenced. Not uncommon pre-gpt either. Hence we suddenly started using two words “reaching out” rather than one “contact”. Eager loading is a technical term with a specific meaning, contrasting with lazy loading. Be the text came out of an llm the real question for the user is, does this technical term actually to this situation. If it does, then it's an appropriate word choice carrying additional information. It is generally the antonym, in technical contexts, to "lazy", as in "lazy evaluation" etc. > Lastly, I learned that children in Portugal have started to speak the Brazilian variant of Portuguese, as videos from there are flooding the Internet. It is interesting how technology affects our lives in more surprising areas. As a non-native speaker of English, I speak and write some weird mix of British and US English, and I always keep forgetting how strong the words "bugger" and "cunt" are in each context. Here's globalization for you. Another giveaway of GPT content for me are overusing importance adjectives like "crucial" or "essential", and of course an extreme overuse of enumerations. An extreme overuse of everything. Like "give me a formula to color every second Excel row blue". "Sure! First make sure you have installed Excel, here are the steps..." I often find that it's easy for me to explain things in enumeration. I'm wondering if a) there is a latent middle manager living inside me, always shuffling his little powerpoint decks, or b) there is a little LM living inside me, making me compulsively enumerate things. I don't see that words often. Maybe it depends on the why question is asked. What about "certainly" and "seasoned developer"? Sorry, this is a British company, we only hire unseasoned developerd here. The word ChatGPT uses the most is "apologize". Whenever I ask it to clean up a code and it either screws up the syntax or removes some necessary parts then starts apologizing until it is fixed or I give up. I specifically ask ChatGPT to stop apologizing because it becomes insufferable after a while. For texts, it uses "furthermore" more than any other word followed by "lastly" imho. The one that keeps popping up for me is "fosters"... Also see: tapestry, testament, delve And So and thrilled "Tapestry" is another frequently used word I've noticed.