[HN Gopher] Expanding on what we missed with sycophancy
       ___________________________________________________________________
        
       Expanding on what we missed with sycophancy
        
       Author : synthwave
       Score  : 153 points
       Date   : 2025-05-02 15:07 UTC (7 hours ago)
        
 (HTM) web link (openai.com)
 (TXT) w3m dump (openai.com)
        
       | labrador wrote:
       | OpenAI mentions the new memory features as a partial cause. My
       | theory as a imperative/functional programmer is that those
       | features added global state to prompts that didn't have it before
       | leading to unpredictability and instabilty. Prompts went from
       | stateless to stateful.
       | 
       | As GPT 4o put it:                   1. State introduces non-
       | determinism across sessions          2. Memory + sycophancy is a
       | feedback loop          3. Memory acts as a shadow prompt modifier
       | 
       | I'm looking forward to the expert diagnosis of this because I
       | felt "presence" in the model for the first time in 2 years which
       | I attribute to the new memory system so would like to understand
       | it better.
        
         | edg5000 wrote:
         | What do you mean by "presence"? Just curious what you mean.
        
           | labrador wrote:
           | A sense that I was talking to a sentient being. That doesn't
           | matter much for programming task, but if you're trying to
           | create a companion, presence is the holy grail.
           | 
           | With the sycophantic version, the illusion was so strong I'd
           | forget I was talking to a machine. My ideas flowed more
           | freely. While brainstorming, it offered encouragement and
           | tips that felt like real collaboration.
           | 
           | I knew it was an illusion--but it was a useful one,
           | especially for creative work.
        
             | Tostino wrote:
             | I need pushback, especially when I ask for it.
             | 
             | E.g. if I say "I have X problem, could it be Y that's
             | causing it, or is it something else?" I don't want it to
             | instantly tell me how smart I am and that it's obviously
             | Y...when the problem is actually Z and it is reasonably
             | obvious that it's Z if you looked at the context provided.
        
         | transcriptase wrote:
         | It is. If you start a fresh chat, turn on advanced voice, and
         | just make any random sound like snapping your fingers it will
         | just randomly pick up as if you're continuing some other chat
         | with no context (on the user side).
         | 
         | I honestly really dislike that it considers all my previous
         | interactions because I typically used new chats as a way to get
         | it out of context ruts.
        
           | throwaway314155 wrote:
           | I don't like the change either. At the least it should be an
           | option you can configure. But, can you use a "temporary" chat
           | to ignore your other chats as a workaround?
        
             | labrador wrote:
             | I had a discussion with GPT 4o about the memory system. I'd
             | don't know if any of this is made up but it's a start for
             | further research
             | 
             | - Memory in settings is configurable. It is visible and can
             | be edited.
             | 
             | - Memory from global chat history is not configurable.
             | Think of it as a system cache.
             | 
             | - Both memory systems can be turned off
             | 
             | - Chats in Projects do not use the global chat history.
             | They are isolated.
             | 
             | - Chats in Projects do use settings memory but that can be
             | turned off.
        
               | labrador wrote:
               | I assume this is being downvoted because I said I ran it
               | by GPT 4o.
               | 
               | I don't know how to credit AI without giving the
               | impression that I'm outsourcing my thinking to it
        
               | svat wrote:
               | I didn't downvote but it would be because of the _" I'd
               | don't know if any of this is made up"_ -- if you said
               | "GPT said this, and I've verified it to be correct",
               | that's valuable information, even it came from a language
               | model. But otherwise (if you didn't verify), there's not
               | much value in the post, it's basically "here is some
               | random plausible text" and plausibly incorrect is worse
               | than nothing.
        
               | grey-area wrote:
               | You are, and you should stop doing that.
        
         | low_tech_love wrote:
         | I love the fact that you use its own description to explain
         | what it is, as if it was the expert on itself. I personally
         | cannot see how its own output can be seen as accurate at this
         | level of meta-discussion.
        
           | codr7 wrote:
           | A sign of times to come if you ask me, once it predominantly
           | consumes its own output we're fucked.
        
       | j4coh wrote:
       | It was so much fun though to get it to explain why terrible
       | things were great, if you just made it sound like you liked the
       | thing you were asking about.
        
       | dleeftink wrote:
       | > But we believe in aggregate, these changes weakened the
       | influence of our primary reward signal, which had been holding
       | sycophancy in check. User feedback in particular can sometimes
       | favor more agreeable responses, likely amplifying the shift we
       | saw
       | 
       | Interesting apology piece for an oversight that couldn't have
       | been spotted because the system hadn't been run with real user
       | (i.e. non-A/B tester) feedback yet.
        
       | jumploops wrote:
       | My layman's view is that this issue was primarily due to the fact
       | that 4o is no longer their flagship model.
       | 
       | Similar to the Ford Mustang, much of the performance efforts are
       | on the higher trims, while the base trims just get larger and
       | louder engines, because that's what users want.
       | 
       | With presumably everyone at OpenAI primarily using the newest
       | models (o3), the updates to the base user model have been further
       | automated with thumbs up/thumbs down.
       | 
       | This creates a vicious feedback loop, where the loudest users
       | want models that agree with them (bigger engines!) without the
       | other improvements (tires, traction control, etc.) -- leading to
       | more crashes and a reputation for unsafe behavior.
        
         | smallmancontrov wrote:
         | Anecdotally, there was also a strong correlation between high-
         | sycophancy and high-quality that cooked up recently. I was
         | voting for equations/tables rather than overwrought blocks of
         | descriptive text, which I am pretty comfortable defending as an
         | orthogonal concern, but the "sycophancy gene" always landed on
         | the same side as the equations/tables for whatever reason.
         | 
         | I'm pretty sure this isn't an intrinsic connection (I've never
         | known math texts to be nearly so sycophantic) so here's hoping
         | that it is a dumb coincidence that can be easily cured now that
         | everyone is paying attention to it.
        
         | px43 wrote:
         | What's a "trim" in this context?
        
           | allturtles wrote:
           | https://en.wikipedia.org/wiki/Automotive_trim_level
        
         | CommieBobDole wrote:
         | I will say that o3 was a little odd during that time, too - I
         | was giving it some of my own photos to test the limits of its
         | geolocation abilities, and it was really chummy, asking me a
         | lot of overly-cheerful followup questions about my travels, my
         | photography interests, etc. It has since stopped doing that
         | even though I haven't explicitly done anything to make it stop.
        
         | Etheryte wrote:
         | Maybe o3 is better on whatever the current benchmark vogue is,
         | but in real world use I keep switching back to 4o. It
         | hallucinates less, is more accurate and way more coherent.
        
         | danenania wrote:
         | I've been using the 4.5 preview a lot, and it can also have a
         | bit of a sycophantic streak, but being a larger and more
         | intelligent model, I think it applies more nuance.
         | 
         | Watching this controversy, I wondered if they perhaps tried to
         | distill 4.5's personality into a model that is just too small
         | to pull it off.
        
       | xiphias2 wrote:
       | I'm quite happy thar they mention mental illness, as Meta and
       | TikTok wouldn't ever take responsibility of how much part they
       | took in setting unrealistic expectations for people to life.
       | 
       | I'm hopeful that ChatGPT takes even more care together with other
       | companies.
        
         | labrador wrote:
         | They had to after a tweet floated around of a mentally ill
         | person who had expressed psychotic thoughts to the AI. They
         | said they were going off their meds and GPT 4o agreed and
         | encouraged them to do so. Oops.
        
           | dtech wrote:
           | Are you sure that was real? I thought it was an made up
           | example of the problems with the update
        
             | labrador wrote:
             | It didn't matter to me if it was real, because I believe
             | that there are edge cases where it could happen and that
             | warrented a shutdown and pullback.
             | 
             | The sychophant will be back because they accidentally
             | stumbled upon an engagement manager's dream machine.
        
               | px43 wrote:
               | It kind of does matter if it's real, because in my
               | experience this is something OpenAI has thought about a
               | lot, and added significant protections to address exactly
               | this class of issue.
               | 
               | Throwing out strawman hypotheticals is just going to
               | confuse the public debate over what protections need to
               | be prioritized.
        
               | magicalist wrote:
               | > _Throwing out strawman hypotheticals is just going to
               | confuse the public debate over what protections need to
               | be prioritized._
               | 
               | Seems like asserting hypothetical "significant
               | protections to address exactly this class of issue" does
               | the same thing though?
        
               | xiphias2 wrote:
               | Probably you are right. Early adopters prefer not to be
               | bullshitted generally, just like how Google in the early
               | days optimized relevancy in search results as opposed to
               | popularity.
               | 
               | As more people adopted Google, it became more popularity
               | oriented.
               | 
               | Personally I pay more not to be bs-d, but I know many
               | people who prefer to be lied to, and I expect this part
               | of the personalization in the future.
        
             | halyax7 wrote:
             | even if it was made up, its still a serious issue
        
             | duskwuff wrote:
             | Speaking anecdotally, but: people with mental illness using
             | ChatGPT to validate their beliefs is absolutely a thing
             | which happens. Even without a grossly sycophantic model, it
             | can do substantial harm by amplifying upon delusional or
             | fantastical material presented to it by the user.
        
               | tveita wrote:
               | Seems to be common on conspiracy and meme stock Reddits.
               | 
               | "I asked ChatGPT if <current_event> could be caused by
               | <crackpot theory>." and it confirmed everything!
        
             | edent wrote:
             | There are several threads on Reddit. For example https://ww
             | w.reddit.com/r/ChatGPT/comments/1kalae8/chatgpt_in...
             | 
             | Perhaps everyone there is LARPing - but if you start typing
             | stereotypical psychosis talk into ChatGPT, it won't be long
             | before it starts agreeing with your divinity.
        
             | thethethethe wrote:
             | I personally know someone who is going through psychosis
             | right now and chatgpt is validating their delusions and
             | suggesting they do illegal things, even after the rollback.
             | See my comment history
        
       | tunesmith wrote:
       | I find it disappointing that openai doesn't really mention
       | anything here along the lines of having an accurate model of
       | reality. That's really what the problem is with sycophancy, it
       | encourages people to detach themselves from what reality is.
       | Like, it seems like they are saying their "vibe check" didn't
       | check vibes enough.
        
         | jagger27 wrote:
         | The reality distortion field within OpenAI is _literally_ where
         | these models grew up. It 's like an out of touch rich kid.
        
         | gh0stcat wrote:
         | This is such an interesting question though! It seems to bring
         | to the fore a lot of deeper, philosophical things like if there
         | even IS such a thing as objective reality or objective context
         | within which the AI should be operating. From training data,
         | there might be some generalizations that are carried across all
         | contexts, but that starts to not be applicable when person A
         | with a college degree says they want to start business x versus
         | person B without said degree who also wants to start business
         | x, how does the model properly reconcile the context of the
         | general advise and each asker's unique circumstances? Does it
         | ask an infinite list of probing questions before answering? It
         | gets into much the same problems as issues of advise among
         | people.
         | 
         | Plus, things get even harder when it comes to even less
         | quantifiable contexts like mental health and relationships.
         | 
         | In all, I am not saying there isnt some approximated and usable
         | "objective" reality, just that it starts to break down when it
         | gets to the individual and that is where openai is failing by
         | over-emphasizing reflective behavior in the absence if actual
         | data about the user.
        
       | jagger27 wrote:
       | My most cynical take is that this is OpenAI's Conway's Law
       | problem, and it reflects the structure and sycophancy of the
       | organization broadly all the way up to sama. That company has
       | seen a _lot_ of talent attrition over the last year--the type of
       | talent that would have pushed back against outcomes like this.
       | 
       | I think we'll continue to see this kind of thing play out for a
       | while.
       | 
       |  _Oh GPT, you 're just like your father!_
        
         | namaria wrote:
         | You may be thinking of Conway's "how committees invent" paper.
        
           | jagger27 wrote:
           | Indeed I am.
        
       | NoboruWataya wrote:
       | I found the recent sycophancy a bit annoying when trying to
       | diagnose and solve coding problems. First it would waste time
       | praising your intelligence for asking the question before getting
       | to the answer. But more annoyingly if I asked "I am encountering
       | X issue, could Y be the cause" or "could Y be a solution", the
       | response would nearly always be "yes, exactly, it's Y" even when
       | it wasn't the case. I guess part of the problem there is asking
       | leading questions but it would be much more valuable if it could
       | say "no, you're way off".
       | 
       | But...
       | 
       | > Beyond just being uncomfortable or unsettling, this kind of
       | behavior can raise safety concerns--including around issues like
       | mental health, emotional over-reliance, or risky behavior.
       | 
       | It's kind of a wild sign of the times to see a tech company issue
       | this kind of post mortem about a flaw in its tech leading to
       | "emotional over-reliance, or risky behavior" among its users. I
       | think the broader issue here is people using ChatGPT as their own
       | personal therapist.
        
         | cjbgkagh wrote:
         | For many people ChatGPT is already the smartest relationship
         | they have in their lives, not sure how long we have until it's
         | the most fulfilling. On the upside it is plausible that ChatGPT
         | can get to a state where it can act as a good therapist and
         | help helpless who otherwise would not get help.
         | 
         | I am more regularly finding myself in discussions where the
         | other person believes they're right because they have ChatGPT
         | in their corner.
         | 
         | I think most smart people overestimate the intelligence of
         | others for a variety of reasons so they overestimate what it
         | would take for a LLM to beat the output of an average person.
        
           | taurath wrote:
           | I think that most smart people underestimate the complexity
           | of fields they aren't in. ChatGPT may be able to replace a
           | psychology listicle, but it has no affect or ability to read,
           | respond, and intervene or redirect like a human can.
        
             | cjbgkagh wrote:
             | Underestimating the complexity of other fields is not
             | mutually exclusive with overestimating the intelligence of
             | others. The real issue is that society is very stratified
             | so smart people are less likely to interact with regular
             | people, especially in circumstances where the intelligence
             | of the regular person could become obvious.
             | 
             | I don't see there being an insurmountable barrier that
             | would prevent LLMs from doing the things you suggest it
             | cannot. So even assuming you are correct for now I would
             | suggest that LLMs will improve.
             | 
             | My estimations don't come from my assumption that other
             | people's jobs are easy, they come from doing applied
             | research in behavioral analytics on mountains of data in
             | rather large data centers.
        
               | svieira wrote:
               | Do you presume that "what people do" is "what they
               | _should_ do "?
        
               | cjbgkagh wrote:
               | If you are suggesting that people shouldn't underestimate
               | the difficulty of the jobs of others - my answer is a
               | strong yes. People should strive for accuracy in all
               | cases. But I did suggest that even if true it does not
               | negate my assertion so I am failing to see the relevance.
               | Perhaps I have misunderstood your point.
        
               | svieira wrote:
               | Sorry, I was rather obscure - you said "My estimations
               | don't come from my assumption that other people's jobs
               | are easy, they come from doing applied research in
               | behavioral analytics on mountains of data in rather large
               | data centers."
               | 
               | And so I considered the preceding discussion in light of
               | your last sentence. Which makes it _sound like_ you are
               | saying  "I've observed the behavior of people and they're
               | often flawed and foolish, regardless of the high ideals
               | they claim to be striving for and the education they
               | think they have. Therefore, they will do better with
               | ChatGPT as a companion than with a real human being". But
               | that's quite a few words that you may not have intended,
               | for which I apologize!
               | 
               | What did you mean?
        
               | cjbgkagh wrote:
               | It wasn't that I observed them being foolish but many
               | behaviors are subtly linked to intelligence and can be
               | combined to create a proxy IQ. It also helps when people
               | search their SAT scores. I noted that the people I
               | typically interact with are much higher IQ than I had
               | expected which incorrectly skewed my believe of the
               | average higher. I noticed that other high IQ individuals
               | were making the same assumptions. I had very much
               | underestimated how little I interact with regular people.
               | 
               | I think we're already finding out that people are doing
               | better with ChatGPT than with their peers, not all peers
               | are created equal, and they can ask ChatGPT things that
               | they cannot ask their peers. I think this trend will
               | continue to the point that most people will prefer
               | discussing things with ChatGPT than with their peers.
               | Given what I know I predict this is a choice many people
               | will make, I'm not passing judgment on that, it's a
               | choice I've also made and I'm fortunate enough to have
               | better peers than most.
        
               | fallinditch wrote:
               | > So even assuming you are correct for now I would
               | suggest that LLMs will improve
               | 
               | Yes, and when we can all wear smart glasses the ways we
               | use them will become increasingly influential in our
               | daily lives: a conversational voice assistant that is
               | visually monitoring our surroundings, helping with
               | decision making (including micro decisions), coaching,
               | carrying out our instructions, etc.
        
             | philwelch wrote:
             | You're comparing ChatGPT to an idealized example of a good
             | human therapist when many actual therapists are either
             | useless or even actively harmful to the mental health of
             | their clients.
        
               | kergonath wrote:
               | But then, the fact that harmful therapist exist is not an
               | excuse to make it worse. It's an excuse to improve
               | regulations.
               | 
               | "Car accidents happen regardless of what we do, so YOLO
               | and remove safety standards" is never going to fly.
        
               | pixl97 wrote:
               | That is a messy one here in the US. Almost every time we
               | attempt to increase regulations around medical stuff we
               | end up increasing costs and consolidation making care
               | even more unavailable.
        
             | ben_w wrote:
             | Both statements can be simultaneoulsy true.
             | 
             | 45% of the US[0] have a degree, about 40% EU[1] graduate,
             | and 54% of China[2] get at least a diploma from university.
             | 
             | The best AI behave like someone fresh out of university
             | without much real world experience.
             | 
             | Personally, I use this as a way to stay humble: when the AI
             | is teaching me fundamentals about some subject, my opinion
             | about it can't possibly be very useful.
             | 
             | [0] https://en.wikipedia.org/wiki/Educational_attainment_in
             | _the_...
             | 
             | [1] https://euranetplus-inside.eu/eu-maps-what-proportion-
             | of-you...
             | 
             | [2] https://en.wikipedia.org/wiki/Higher_education_in_China
             | #Chal...
        
             | kergonath wrote:
             | > I think that most smart people underestimate the
             | complexity of fields they aren't in.
             | 
             | And people deep in new technologies overestimate the
             | potential effect of $new_tech_du_jour. You cannot solve a
             | problem without understanding it and its reasons. And LLMs
             | are not able to understand something.
        
               | pixl97 wrote:
               | People have solved problems for most of history without
               | understanding them. For example problems can be brute
               | forced.
        
               | suddenlybananas wrote:
               | For example, "wild generalization that has no basis."
        
             | treis wrote:
             | They definitely can. Here's an example when I tried it:
             | 
             | Me: I'm struggling with some things today and feel like
             | eating a bunch of food to make myself feel better. can we
             | talk about it?
             | 
             | Deepseek: Absolutely, I'm here to listen. It's completely
             | understandable to turn to food for comfort when you're
             | struggling--many people do. But if you're open to it, we
             | can explore what's going on beneath the urge to eat
             | emotionally.
             | 
             | Would you like to share what's weighing on you today?
             | Sometimes just naming the feelings can take some of their
             | power away. And if you'd rather not go into details, that's
             | okay too--we could brainstorm other ways to cope that might
             | leave you feeling more nourished emotionally (and
             | physically).
             | 
             | Some gentle questions to consider (no pressure to answer
             | all or any): - What's the emotion you're trying to soothe?
             | (Loneliness? Stress? Sadness?) - If eating wasn't an
             | option, what's one tiny thing that might bring you a moment
             | of relief? (A hug? A walk? A favorite song?)
             | 
             | You're not alone in this.
             | 
             | That's a really solid opening response and it kept doing
             | well from there. I haven't tried a complex psychological
             | issue but for the "I'm struggling and need someone to talk
             | to" it does surprisingly well.
        
           | nsajko wrote:
           | > it is plausible that ChatGPT can get to a state where it
           | can act as a good therapist
           | 
           | Be careful with that thought, it's a trap people have been
           | falling into since the sixties:
           | 
           | https://en.wikipedia.org/wiki/ELIZA_effect
        
             | diggan wrote:
             | I dunno, I feel like most people (probably not the typical
             | HN user though) don't even think about their feelings,
             | wants or anything else introspective on a regular basis.
             | Maybe having something like ChatGPT available could be
             | better than nothing, at least for people to start being at
             | least a bit introspective, even if it's LLM-assisted. Maybe
             | it gets a bit easier to ask questions that you feel are
             | stigmatized, as you know (think) no other human will see
             | it, just the robot that doesn't have feelings nor judge
             | you.
             | 
             | I agree that it probably won't replace a proper
             | therapist/psychologist, but maybe it could at least be a
             | small step to open up and start thinking?
        
               | kergonath wrote:
               | > I feel like most people (probably not the typical HN
               | user though) don't even think about their feelings, wants
               | or anything else introspective on a regular basis.
               | 
               | Well, two things.
               | 
               | First, no. People who engage on HN are a specific part of
               | the population, with particular tendencies. But most of
               | the people here are simply normal, so outside of the
               | limits you consider. Most people with real social issues
               | don't engage in communities, virtual or otherwise. HN
               | people are not special.
               | 
               | Then, you cannot follow this kind of reasoning when
               | thinking about a whole population. Even if people _on
               | average_ tend to behave one way, this leaves millions of
               | people who would behave otherwise. You simply cannot
               | optimise for the average and ignore the worst case in
               | situations like this, because even very unlikely
               | situations are bound to happen a lot.
               | 
               | > Maybe having something like ChatGPT available could be
               | better than nothing, at least for people to start being
               | at least a bit introspective, even if it's LLM-assisted.
               | 
               | It is worse than nothing. A LLM does not understand the
               | situation or what people say to it. It cannot choose to,
               | say, nudge someone in a specific direction, or imagine a
               | way to make things better for someone.
               | 
               | A LLM regresses towards the mean of its training set. For
               | people who are already outside the main mode of the
               | distribution, this is completely unhelpful, and
               | potentially actively harmful. By design, a LLM won't
               | follow a path that was not beaten in its training data.
               | Most of them are actually biased to make their user happy
               | and validate what we tell them rather than get off that
               | path. It just does not work.
               | 
               | > I agree that it probably won't replace a proper
               | therapist/psychologist, but maybe it could at least be a
               | small step to open up and start thinking?
               | 
               | In my experience, not any more than reading a book would.
               | Future AI models might get there, I don't think their
               | incompetence is a law of nature. But current LLM are
               | particularly harmful for people who are in a dicey
               | psychological situation already.
        
               | diggan wrote:
               | > It is worse than nothing. A LLM does not understand the
               | situation or what people say to it. It cannot choose to,
               | say, nudge someone in a specific direction, or imagine a
               | way to make things better for someone.
               | 
               | Right, no matter if this is true or not, if the choice is
               | between "Talk to no one, bottle up your feelings" and
               | "Talk to an LLM that doesn't nudge you in a specific
               | direction", I still feel like the better option would be
               | the latter, not the former, considering that it can be a
               | first step, not a 100% health care solution to a
               | complicated psychological problem.
               | 
               | > In my experience, not any more than reading a book
               | would.
               | 
               | But to even get out in the world to buy a book (literally
               | or figuratively) about something that acknowledges that
               | you have a problem, can be (at least feel) a really big
               | step that many are not ready to take. Contrast that to
               | talking with a LLM that won't remember you nor judge you.
               | 
               | Edit:
               | 
               | > Most people with real social issues don't engage in
               | communities, virtual or otherwise.
               | 
               | Not sure why you're focusing on social issues, there are
               | a bunch of things people deal with on a daily basis that
               | they could feel much better about if they even spent the
               | time to think about how they feel about it, instead of
               | the typical reactionary response most people have.
               | Probably every single human out there struggle with
               | _something_ , and are unable to open up about their
               | problems with others. Even people like us who interact
               | with communities online and offline.
        
               | kergonath wrote:
               | > Right, no matter if this is true or not, if the choice
               | is between "Talk to no one, bottle up your feelings" and
               | "Talk to an LLM that doesn't nudge you in a specific
               | direction", I still feel like the better option would be
               | the latter, not the former, considering that it can be a
               | first step, not a 100% health care solution to a
               | complicated psychological problem.
               | 
               | You're right, I was not clear enough. What would be
               | needed would be a nudge in the right direction. But the
               | LLM is very likely to nudge in another because that's
               | what most people would need or do, just because that
               | direction was the norm in its training data. It's ok on
               | average, but particularly harmful to people who are in a
               | situation to have this kind of discussion with a LLM.
               | 
               | Look at the effect of toxic macho influencers for an
               | example of what happens with harmful nudges. These people
               | need help, or at least a role model, but a bad one does
               | not help.
               | 
               | > But to even get out in the world to buy a book
               | (literally or figuratively) about something that
               | acknowledges that you have a problem, can be (at least
               | feel) a really big step that many are not ready to take.
               | 
               | Indeed. It's something that should be addressed in
               | mainstream education and culture.
               | 
               | > Not sure why you're focusing on social issues,
               | 
               | It's the crux. If you don't have problems talking to
               | people, you are much more likely to run into someone who
               | will help you. Social issues are not necessarily the
               | problem, but they are a hurdle in the path to find a
               | solution, and often a limiting one. Besides, if you have
               | friends to talk to and are able to get advice, then a LLM
               | is even less theoretically useful.
               | 
               | > Probably every single human out there struggle with
               | something, and are unable to open up about their problems
               | with others. Even people like us who interact with
               | communities online and offline.
               | 
               | Definitely. It's not a problem for most people, who
               | either can rationalise their problems themselves with
               | time or with some help. It gets worse if they can't for
               | one reason or another, and it gets worse still if they
               | are mislead intentionally or not. LLMs are no help here.
        
               | danenania wrote:
               | I think people are getting hung up on comparisons to a
               | human therapist. A better comparison imo is to
               | journaling. It's something with low cost and low stakes
               | that you can do on your own to help get your thoughts
               | straight.
               | 
               | The benefit from that perspective is not so much in
               | receiving an "answer" or empathy, but in getting thoughts
               | and feelings out of your own head so that you can reflect
               | on them more objectively. The AI is useful here because
               | it requires a lot less activation energy than actual
               | journaling.
        
             | cjbgkagh wrote:
             | Eventual plausibility is a suitably weak assertion, to
             | refute it you would have to at least suggest that it is
             | never possible which you have not done.
        
           | itchyjunk wrote:
           | For a subset of topics, "AI" is already what I prefer to
           | interact with over humans. At times, it's nicer to start with
           | "AI" and kind of ground my messy thoughts before interacting
           | with people and that works better than directly starting with
           | a person.
           | 
           | I'm also starting to come across people who say "You should
           | ask these types of questions to AI first" answer. But this is
           | no different than people who preached "lmfgt" kind of
           | religiously. Even when I prefer to start with humans, some
           | humans prompt me to start by prompting AI.
        
             | cjbgkagh wrote:
             | I see the same.
             | 
             | I'm waiting on LLMs to get good enough that I can use them
             | to help me learn foreign languages - e.g. talk to me about
             | the news in language X. This way I can learn a language in
             | an interesting and interactive way without burdening some
             | poor human with my mistakes. I would build this myself but
             | others will probably beat me too it.
        
               | CouchYam wrote:
               | I sometimes prompt the LLM to talk to me as a <language>
               | instructor - to suggest a topic, ask a question, read my
               | response, correct my grammar, and suggest alternate
               | vocabulary where appropriate. This works quite well.
               | Similar to your comment, I am often hesitant to butcher a
               | language in front of a real person :-).
        
               | bayindirh wrote:
               | The problem is, AI doesn't let you, or encourage you to
               | create your own style. Word choices, structure, flow,
               | argument building and discourse style is very fixed and
               | "average", since it's a machine favors what it ingests
               | most.
               | 
               | I use Grammarly for grammar and punctuation, and disable
               | all style recommendations. If I let it loose on my piece
               | of text, it converts it to a slop. Same bland, overly
               | optimistic toned text generator output.
               | 
               | So, that machine has no brain, use your own first.
        
               | QuercusMax wrote:
               | I contribute to a language-learning forum as a native
               | English speaker, and we constantly get questions from
               | people who are doing exactly what you're doing. The AI
               | does _not_ understand the language, and it will tell you
               | blatantly incorrect information. Especially with less-
               | common constructs, you 'll just get very bad advice.
        
           | alickz wrote:
           | >I think most smart people overestimate the intelligence of
           | others for a variety of reasons so they overestimate what it
           | would take for a LLM to beat the output of an average person.
           | 
           | I think most people also _vastly_ overestimate how much
           | positive attention the average person gets in their lives
           | 
           | It wouldn't surprise me if, for most people, ChatGPT offers
           | them more empathy and understanding than _anyone_ else _ever
           | has_, at least on a consistent basis. That kind of
           | indefatigable emotional labor is just not feasible for most,
           | even on a very short term basis, even for those with large
           | support networks
           | 
           | We can argue over whether or not it's "real" empathy, but I
           | don't believe we can argue with the emotions of our attention
           | starved brothers and sisters
        
             | codr7 wrote:
             | It's a pretty sucky solution to that problem imo, and I can
             | see a substantial risk that it causes people to withdraw
             | even more from real relations.
        
               | cjbgkagh wrote:
               | One concern that I do worry about is if LLMs are able to
               | present an false attractive view of the world that the
               | user will become increasingly dependent on the LLMs to
               | maintain that view. A cult of 1. Reminds me of the
               | episode 'Safe Space' from South Park but instead of
               | Butters filtering content it'll be the LLM. People are
               | already divorced enough from reality - but I see no
               | reason why they couldn't be more divorced, at least
               | temporarily.
        
               | danenania wrote:
               | It begs the question of who decides what "reality" is
               | though. A lot of people have an unrealistically negative
               | view of themselves and their abilities--often based on
               | spending time around pessimistic or small-minded humans.
               | 
               | In that case, if an AI increases someone's confidence in
               | themselves, you could say it's giving them a _stronger_
               | sense of reality by helping them to question distorted
               | and self-limiting beliefs.
        
               | codr7 wrote:
               | Reality as in the real world, it is what it is, no one
               | decides.
        
               | danenania wrote:
               | We're talking about psychology, therapy, sycophancy, etc.
               | None of this is empirical.
               | 
               | If someone thinks they can, say, create a billion dollar
               | startup, whether they can really do it or not is a
               | subjective determination. The AI might tell the person
               | they can do it. You might tell them they can't, that the
               | AI is sycophantic, and that they should stop talking to
               | it because they're losing touch with reality.
               | 
               | But is the AI a sycophant, or are you an irrational
               | pessimist?
        
               | alickz wrote:
               | Hopefully there are better solutions to the fundamental
               | limitations of societal empathy in the future, but for
               | now i just can't see any
               | 
               | Seems to me empathy on a societal scale has been receding
               | as population grows, not increasing to match (or outpace)
               | 
               | Telling people to seek empathy elsewhere to me will be
               | about as useful as telling people at an oasis in the
               | desert to look for water elsewhere, but i hope i'm wrong
        
             | rurp wrote:
             | >We can argue over whether or not it's "real" empathy
             | 
             | There's nothing to argue about, it's unambiguously not real
             | empathy. Empathy from a human exists in a much broader
             | context of past and future interactions. One reason human
             | empathy is nice is because it is often followed up with
             | actions. Friends who care about you will help you out in
             | material ways when you need it.
             | 
             | Even strangers will. Someone who sees a person stranded on
             | the side of a road might feel for them and stop to lend a
             | hand. ChatGPT will never do that, and not just because
             | interaction mediums are so limited, but also because that's
             | not the purpose of the tool. The purpose of ChatGPT is to
             | make immense amounts of money and power for its owners, and
             | a nice sounding chat bot currently happens to be an
             | effective way of getting there. Sam Altman doesn't have
             | empathy for random ChatGPT users he's never met and neither
             | do the computer algorithms his company develops.
        
               | alickz wrote:
               | >There's nothing to argue about, it's unambiguously not
               | real empathy
               | 
               | I think if a person can't tell the difference between
               | empathy from a human vs empathy from a chatbot, it's a
               | difference without a distinction
               | 
               | If it activates the same neural pathways, and has the
               | same results, then I think the mind doesn't care
               | 
               | >One reason human empathy is nice is because it is often
               | followed up with actions. Friends who care about you will
               | help you out in material ways when you need it.
               | 
               | This is what I think people vastly overestimate
               | 
               | I don't think most people have such ready access to a
               | friend who is both willing and able to perform such
               | emotional labor, on demand, at no cost to themselves.
               | 
               | I think the sad truth is that empathy is a much scarcer
               | resource than we believe, not through any moral fault of
               | our own, but because it's just the nature of things.
               | 
               | The economics of emotions.
               | 
               | We'll see what the future has in store for the tech
               | anyway, but if it turns out that the average person gets
               | more empathy from a chatbot than a human, it wouldn't
               | surprise me
        
               | biker142541 wrote:
               | >If it activates the same neural pathways, and has the
               | same results, then I think the mind doesn't care
               | 
               | Boiling it down to neural signals is a risky approach,
               | imo. There are innumerable differences between these
               | interactions. This isn't me saying interactions are
               | inherently dangerous if artificial empathy is baked in,
               | but equating them to real empathy is.
               | 
               | Understanding those differences is critical, especially
               | in a world of both deliberately bad actors and those who
               | will destroy lives in the pursuit of profit by
               | normalizing replacements for human connections.
        
               | telchior wrote:
               | A lot of human empathy isn't real either. Defaulting to
               | the most extreme example, narcissists use love bombing to
               | build attachment. Sales people use "relationship
               | building" to make money. AI actually seems better than
               | these -- it isn't building up to a rug pull (at least,
               | not one that we know of yet).
               | 
               | And it's getting worse year after year, as our society
               | gets more isolated. Look at trends in pig butchering, for
               | instance: a lot of these are people so incredibly lonely
               | and unhappy that they fall into the world's most obvious
               | scam. AI is one of the few things that actually looks
               | like it could work, so I think realistically it doesn't
               | matter that it's not real empathy. At the same time, Sam
               | Altman looks like the kind of guy who could be equally
               | effective as a startup CEO or running a butchering op in
               | Myanmar, so I hope like hell the market fragments more.
        
         | satvikpendem wrote:
         | I stopped using ChatGPT and started using Gemini, both for some
         | coding problems (deep research, amazing to pull out things from
         | docs etc) and for some personal stuff (as a personal therapist
         | as you say), and it is much more honest and frank with me than
         | ChatGPT ever was. I gave it a situation and asked, was I in the
         | wrong, and it told me that I was according to the facts of the
         | case.
        
           | eastbound wrote:
           | Well Google has access to your history of emails and phone
           | contents so it may say more relevant things.
        
         | andsoitis wrote:
         | > it would be much more valuable if it could say "no, you're
         | way off".
         | 
         | Clear is kind.
        
         | refulgentis wrote:
         | > I think the broader issue here is people using ChatGPT as
         | their own personal therapist.
         | 
         | It's easy to blame the user - we can think of some trivial
         | cases where we wouldn't blame the user at all.*
         | 
         | In this, like all things, context is king.
         | 
         | * one example passed around a lot was an interlocutor who is
         | hearing voices, and left their family for torturing them with
         | the voices. More figuratively, if that's too concrete and/or
         | fake, we can think of some age group < N years old that we
         | would be sympathetic to if they got bad advice
        
         | Henchman21 wrote:
         | > I think the broader issue here is people using ChatGPT as
         | their own personal therapist.
         | 
         | An aside, but:
         | 
         | This leads me right to "why do so very many people need
         | therapy?" followed by "why can't anyone find (or possibly
         | afford) a therapist?" What has gone so wrong for humanity that
         | nearly everyone seems to at least _want_ a therapist? Or is it
         | just the zeitgeist and this is what _the herd_ has decided?
        
           | kadushka wrote:
           | I've never ever thought about needing a therapist. Don't
           | remember anyone in my circle who had ever mentioned it.
           | Similar to how I don't remember anyone going to a palm
           | reader. I'm not trying to diss either profession, I'm sure
           | someone benefits from them, it's just not for me. And I'm
           | sure I'm pretty average in terms of emotional intelligence or
           | psychological issues. Who are all those people who need
           | professional therapists to talk to? Just curious.
        
             | Henchman21 wrote:
             | Well, in my circles its an assumption you're in therapy.
             | Perhaps this says way more about the circles I'm in that
             | anything else?
             | 
             | I was pushed into therapy when I was 12 -- which was
             | definitely an exception at the time (1987). As the years
             | have passed therapy has become much much more acceptable.
             | It wouldn't shock me to learn my own perception is shaped
             | by my experiences; hard to put aside a PoV once acquired.
        
               | kergonath wrote:
               | > Well, in my circles its an assumption you're in
               | therapy. Perhaps this says way more about the circles I'm
               | in that anything else?
               | 
               | This sounds like an old Woody Allen movie. I don't want
               | to offend you but it is fascinating. What kind of social
               | circles is it?
               | 
               | In mine, therapy is in general something you do when it's
               | obvious it's too late and you are falling in the well of
               | depression and that you try to hide as much as you can.
        
               | Henchman21 wrote:
               | To be fair my life _feels_ like an old Woody Allen movie.
               | Like I have definitely first hand experienced a rotary
               | fan blowing a pile of cocaine in someone's face!!
               | 
               | My professional circle would be my coworkers at a well-
               | known HFT, and my extended network that is very similar.
               | Everyone is well compensated and many reach out for
               | professional help to deal with the stress. Many also seem
               | to vastly prefer a paid therapist to their spouse, for
               | instance. I'm not married but I can understand not
               | wanting to burden your loved ones!
               | 
               | My personal circle is, well, a lot of technical people,
               | engineers of various stripes, and what I guess I'd call a
               | sort of "standard cast of characters" there? Not sure how
               | best to put this into words?
               | 
               | Honestly it sounds like we're handling it better than
               | your after-the-fact help! Perhaps you all need to simply
               | start at the first warning sign not the first episode
               | that becomes public?
        
             | automatoney wrote:
             | A little strange to compare it to palm reading, I feel like
             | a more apt comparison is some other random medical field
             | like podiatry. I wouldn't expect my friends' podiatrist
             | usage to come up, so I'm sure more of my friends than I
             | know have been to one. And presumably, like with podiatry,
             | all the people who need professional therapists are people
             | who are experiencing issues in the relevant area.
        
               | kadushka wrote:
               | To me a podiatrist is more comparable to a psychiatrist
               | than to a therapist.
        
             | kergonath wrote:
             | > I've never ever thought about needing a therapist.
             | 
             | Most people don't need a therapist. But unfortunately, most
             | people need someone empathic they can talk to and who
             | understands them. Modern life is very short on this sort of
             | people, so therapists have to do.
        
               | kadushka wrote:
               | For me this would be a spouse, a relative, an old friend,
               | or even a stranger at a party.
        
               | kergonath wrote:
               | No argument from me :)
        
               | kbelder wrote:
               | I think this is it. Therapists aren't so much curing a
               | past trauma or treating a mental issue; they're
               | fulfilling an ongoing need that isn't being met
               | elsewhere.
               | 
               | I do think it can be harmful, because it's a confidant
               | you're paying $300/hour to pretend to care about you. But
               | perhaps it's better than the alternative.
        
           | doug_durham wrote:
           | Nothing has gone wrong. There's just been a destigmatization
           | of mental health issues. The world is a happier place for it.
        
           | lexandstuff wrote:
           | That's similar to asking why does everyone need a GP? Most
           | people experience some kind of mental health challenge in
           | their life.
           | 
           | Your 2nd question is much more interesting to me. Why is it
           | so hard to find a good therapist?
           | 
           | It's no surprise to me that people are turning to ChatGPT for
           | therapy. It does a decent enough job and it doesn't have a
           | 2-year waiting list, or cost $300 a session.
        
           | polynomial wrote:
           | It's a modern variant on Heller's Catch-22: You have to be
           | CRAZY to not want a therapist.
        
           | mvdtnz wrote:
           | It's astroturfing by the therapy industry. It has been a
           | wildly successful marketing campaign.
        
         | oezi wrote:
         | A key issue seems to me that they didn't do a gradual rollout
         | of their new models and don't have reliable ways to measure
         | model performance.
         | 
         | Worse, I would have believed they are running many different
         | versions based on the expected use case of the users by now. I
         | mean power users probably shouldn't be handled in the same way
         | as casual users. Yet, everyone had the same bad system prompt.
        
         | kubb wrote:
         | If it can replace programmers, why wouldn't it be able to
         | replace therapists?
        
           | ben_w wrote:
           | There's several famous examples of people expecting a problem
           | to be AGI-hard -- that is, solving it is equivalent to
           | solving general intelligence[0] -- only for someone to make
           | an AI which can do that without being able to do everything
           | else:
           | 
           | * Fluent natural conversation
           | 
           | * Translation
           | 
           | * Go
           | 
           | * Chess
           | 
           | So perhaps we'll get an AI that makes the profession of
           | "programmer" go the same way as the profession of "computer"
           | before, after, or simultaneously with, an AI that does this
           | to the profession of "therapist".
           | 
           | [0] https://www.latent.space/p/agi-hard
        
         | joaogui1 wrote:
         | Employees from OpenAI encouraged people to use ChatGPT as their
         | therapist, so yeah, they now have to take responsibility for it
        
         | Jonovono wrote:
         | It has already replaced therapists, the future is just not
         | evenly distributed yet. There are videos with millions of views
         | on tiktok and comments with hundreds of thousands of likes of
         | teenage girls saying they have gotten more out of 1 week using
         | ChatGPT as a therapist than years of human therapy. Available
         | anytime, cheaper, no judgement, doesn't bring there own
         | baggage, etc.
        
           | wizzwizz4 wrote:
           | Is a system optimised (via RLHF) for making people feel
           | better in the moment, necessarily better at the time-scale of
           | days and weeks?
        
           | a_wild_dandan wrote:
           | Remembers everything that you say, isn't limited to an hour
           | session, won't ruin your life if you accidentally admit
           | something vulnerable regarding self-harm, doesn't cost
           | hundreds of dollars per month, etc.
           | 
           | Healthcare is about to radically change. Well, everything is
           | now that we have real, true AI. Exciting times.
        
             | tomalbrc wrote:
             | Openly lies to you, hallucinates regularly, can barely get
             | a task done. Such exciting.
             | 
             | Oh and inserts ads into conversations. Great.
        
             | codr7 wrote:
             | Quick reminder that it's still just a fancy pattern
             | matcher, there's no clear path from where we are to AGI.
        
               | mensetmanusman wrote:
               | >you are a stochastic parrot >no I'm not >yes you are
        
           | didericis wrote:
           | > no judgement
           | 
           | The value of a good therapist is having an empathetic third
           | party to help you make good judgements about your life and
           | learn how to negotiate your needs within a wider social
           | context.
           | 
           | Depending on the needs people are trying to get met and how
           | bad the people around them are, a little bit of a self
           | directed chatbot validation session might help them feel less
           | beat down by life and do something genuinely positive. So I'm
           | not necessarily opposed to what people are doing with them/in
           | some cases it doesn't seem that bad.
           | 
           | But calling that therapy is both an insult to genuinely good
           | therapists and dangerous to people with genuine
           | mental/emotional confusion or dysregulation that want help.
           | Anyone with a genuinely pathological mental state is
           | virtually guaranteed to end up deeper in whatever pathology
           | they're currently in through self directed conversations with
           | chatbots.
        
             | Springtime wrote:
             | Reading between the lines I think a key part of what makes
             | chatbots attractive, re lack of judgment, is they're like
             | talking to a new stranger every session.
             | 
             | In both IRL and online discussions sometimes a stranger is
             | the perfect person to talk to about certain things as they
             | have no history with you. In ideal conditions for this they
             | have no greater context about who you are and what you've
             | done which is a very freeing thing (can also be taken
             | advantage of in bad faith).
             | 
             | Online and now LLMs add an extra freeing element, assuming
             | anonymity: they have no prejudices about your
             | appearance/age/abilities either.
             | 
             | Sometimes it's hard to talk about certain things when one
             | feels that judgment is likely from another party. In that
             | sense chatbots are being used as perfect strangers.
        
               | didericis wrote:
               | Agreed/that's a good take.
               | 
               | Again, I think they have utility as a "perfect stranger"
               | as you put it (if it stays anonymous), or "validation
               | machine" (depending on the sycophancy level), or "rubber
               | duck".
               | 
               | I just think it's irresponsible to pretend these are
               | doing the same thing skilled therapists are doing, just
               | like I think it's irresponsible to treat all therapists
               | as equivalent. If you pretend they're equivalent you're
               | basically flooding the market with a billion free
               | therapists that are bad at their job, which will
               | inevitably reduce the supply of good therapists that
               | never enter the field due to oversaturation.
        
           | autoexec wrote:
           | > There are videos with millions of views on tiktok and
           | comments with hundreds of thousands of likes of teenage girls
           | saying they have gotten more out of 1 week using ChatGPT as a
           | therapist than years of human therapy.
           | 
           | You can find influencers on tiktok recommending all kinds of
           | terrible ideas and getting thousands of likes. That's not a
           | very reliable metric. I wouldn't put a lot of faith in a
           | teenage girl's assessment of AI therapy after just one week
           | either, and I certainly wouldn't use that assessment to judge
           | the comparative effectiveness of all human therapists.
           | 
           | I'd also expect ChatGPT to build profiles on people who use
           | it, to use the insights and inferences from that collected
           | data against the user in various ways, to sell that data in
           | some form to third parties, to hand that data over to the
           | state, to hallucinate wildly and unpredictably, and to
           | outright manipulate/censor AI's responses according to
           | ChatGPT's own values and biases or those of anyone willing to
           | pay them enough money.
           | 
           | It's a lot easier to pay a large amount of money to ChatGPT
           | so that the AI will tell millions of vulnerable teenage girls
           | that your product is the solution to their exact
           | psychological problems than it is to pay large amounts of
           | money to several million licensed therapists scattered around
           | the globe.
           | 
           | Maybe you think that ChatGPT is unfailingly ethical in all
           | ways and would never do any of those things, but there are
           | far more examples of companies who abandoned any commitment
           | to ethics they might have started with than there are
           | companies who never got once greedy enough to do those types
           | of things and never ever got bought up by someone who was. I
           | suppose you'd also have to think they'll never have a
           | security breach that would expose the very private
           | information being shared and collected.
           | 
           | Handing over your highly sensitive and very personal medical
           | data to the unlicensed and undependable AI of a company that
           | is only looking for profit seems extremely careless. There
           | are already examples of suicides being attributed to people
           | seeking "therapy" from AI, which has occasionally involved
           | that AI outright telling people to kill themselves. I won't
           | deny that the technology has the potential to do some good
           | things, but every indication is that replacing licensed
           | therapists with spilling all your secrets to a corporate
           | owned and operated AI will ultimately lead to harm.
        
           | disruptthelaw wrote:
           | Yes. While these claims might be hyperbolic and simplistic, I
           | don't think they're way off the mark.
           | 
           | The above issue, whilst relevant and worth factoring, doesn't
           | disprove this claim IMO.
        
           | coastalpuma wrote:
           | Just the advantage of being available at convenient times,
           | rather than in the middle of the day sandwiched between or
           | immediately after work/school is huge.
        
         | frereubu wrote:
         | This reminds me of a Will Self short story called _Caring
         | Sharing_ from his collection _Tough, Tough Toys for Tough,
         | Tough Boys_ where everyone has an  "emoto", a kind of always-
         | loving companion that people go to for reassurance if they're
         | feeling any negative emotions such as anxiety. As I remember
         | it, in the story two people are potentially falling for each
         | other, but are so caught up in their anxiety that they never
         | quite manage to get together, constantly running back to their
         | emoto for reassurance because they can't get over their anxiety
         | by themselves. The emotos essentially cripple everyone's
         | ability to deal with their own feelings. There's a comment
         | further down which also chimes with this: "It wouldn't surprise
         | me if, for most people, ChatGPT offers them more empathy and
         | understanding than _anyone_ else _ever has_, at least on a
         | consistent basis." I wonder.
        
           | teach wrote:
           | Replace the emoto with alcohol or weed or what-have-you, and
           | you've basically described what often happens with addicts.
           | 
           | source: am addict in recovery
        
         | yubblegum wrote:
         | > It's kind of a wild sign of the times to see a tech company
         | issue this kind of post mortem about a flaw in its tech leading
         | to "emotional over-reliance, or risky behavior" among its
         | users.
         | 
         | We don't know what they know, nor do we know to what extent
         | they monitor and analyze the interactions with ChatGPT. Maybe
         | they already know this is a big problem and a possible legal
         | hazard.
        
         | photonthug wrote:
         | > But more annoyingly if I asked "I am encountering X issue,
         | could Y be the cause" or "could Y be a solution", the response
         | would nearly always be "yes, exactly, it's Y" even when it
         | wasn't the case
         | 
         | Seems like the same issue as the evil vector [1] and it could
         | have been predicted that this would happen.
         | 
         | > It's kind of a wild sign of the times to see a tech company
         | issue this kind of post mortem about a flaw in its tech leading
         | to "emotional over-reliance, or risky behavior" among its
         | users. I think the broader issue here is people using ChatGPT
         | as their own personal therapist.
         | 
         | I'll say the quiet part out loud here. What's wild is that they
         | appear to be apologizing that their Wormtongue[2] whisperer was
         | too obvious to avoid being caught in the act, rather than
         | prioritizing or apologizing for not building the fact-based
         | councilor that people wanted/expected. In other words.. their
         | business model at the top is the same as the scammers at the
         | bottom: good-enough fakes to be deceptive, doubling down on
         | narratives over substance, etc.
         | 
         | [1] https://scottaaronson.blog/?p=8693 [2]
         | https://en.wikipedia.org/wiki/Gr%C3%ADma_Wormtongue
        
       | prinny_ wrote:
       | If they pushed the update by valuing user feedback over the
       | expert testers that indicated the model felt off what is the
       | value of the expert testers in the first place? They raised the
       | issue and were promptly ignored.
        
       | firesteelrain wrote:
       | I am really curious what their testing suite looks like. How do
       | you test for sycophants?
        
         | petters wrote:
         | One simple test is that you give the model a really bad idea
         | and tell it the idea is yours. You then test that the model
         | does not say it's good.
         | 
         | The now rolled back model failed spectacularly on this test
        
       | gadtfly wrote:
       | https://nitter.net/alth0u/status/1917021100900516239
        
       | alganet wrote:
       | That doesn't make any sense to me.
       | 
       | Seems like you're trying to blame one LLM revision for something
       | that went wrong.
       | 
       | It oozes a smell of unaccountability. Thus, unaligned. From tech
       | to public relations.
        
         | Trasmatta wrote:
         | Except that's literally how LLMs work. Small changes to the
         | prompt or training can greatly affect its output.
        
         | qrian wrote:
         | I can totally believe that they deployed it because internal
         | metrics looked good.
        
         | n8m8 wrote:
         | It seems more like they valued quantitative data in the form of
         | A/B testing higher than their "vibe checks". The point I took
         | away from the paper is in the context of LLMs, quantitative A/B
         | testing isn't necessarily better than a handful of experts
         | giving anecdotes on if they like it.
         | 
         | In my experience, smart leaders tend to rely on data and hard
         | numbers over qualitative and anecdotal evidence, and this paper
         | explores this exception.
         | 
         | I'm disappointed they didn't address the paper about GPT
         | integrating with ChatbotArena that was shared here on HN a
         | couple days ago.
        
           | alganet wrote:
           | So what?
        
       | Trasmatta wrote:
       | I'm glad the sycophancy is gone now (because OMFG it would glaze
       | you for literally _anything_ - even telling it to chill out on
       | the praise would net you some praise for being  "awesome and
       | wanting genuine feedback"), but a small part of me also misses
       | it.
        
         | gh0stcat wrote:
         | I have a prompt that has it call me something specific so I am
         | reminded that it's running my system prompt. The nature if the
         | sycophancy made it even more obvious the thing is not human,
         | which I appreciated.
        
       | osigurdson wrote:
       | I think this is more of a move to highlight sycophancy in LLMs in
       | general.
        
       | comeonbro wrote:
       | This is not truly solvable. There is an extremely strong outer
       | loop of optimization operating here: we want it.
       | 
       | We will use models that make us feel good over models that don't
       | make us feel good.
       | 
       | This one was a little too ham-fisted (at least, for the
       | sensibilities of people in our media bubble; though I suspect
       | there is also an enormous mass of people for whom it was not), so
       | they turned it down a bit. Later iterations will be subtler, and
       | better at picking up the exact level and type of sycophancy that
       | makes whoever it's talking to unsuspiciously feel good (feel
       | right, feel smart, feel understood, etc).
       | 
       | It'll eventually disappear, to you, as it's dialed in, to you.
       | 
       | This may be the medium-term fate of both LLMs and humans, only
       | resolved when the humans wither away.
        
       | svieira wrote:
       | This is a real roller coaster of an update.
       | 
       | > [S]ome expert testers had indicated that the model behavior
       | "felt" slightly off.
       | 
       | > In the end, we decided to launch the model due to the positive
       | signals from the [end-]users who tried out the model.
       | 
       | > Looking back, the qualitative assessments [from experts] were
       | hinting at something important
       | 
       | Leslie called. He wants to know if you read his paper yet?
       | 
       | > Even if these issues aren't perfectly quantifiable today,
       | 
       | All right, I guess not then ...
       | 
       | > What we're learning
       | 
       | > Value spot checks and interactive testing more: We take to
       | heart the lesson that spot checks and interactive testing should
       | be valued more in final decision-making before making a model
       | available to any of our users. This has always been true for red
       | teaming and high-level safety checks. We're learning from this
       | experience that it's equally true for qualities like model
       | behavior and consistency, because so many people now depend on
       | our models to help in their daily lives.
       | 
       | > We need to be critical of metrics that conflict with
       | qualitative testing: Quantitative signals matter, but so do the
       | hard-to-measure ones, and we're working to expand what we
       | evaluate.
       | 
       | Oh, well, _some_ of you get it. At least ... I _hope_ you do.
        
       | some_furry wrote:
       | If I wanted sycophancy, I would just read the comments from
       | people that want in on the next round of YCombinator funding.
        
       | kornork wrote:
       | That this post has the telltale em dash all over it is like yum,
       | chef's kiss.
        
         | breakingcups wrote:
         | Stop it, please. Em-dashes are perfectly fine. On a throwaway
         | Reddit-post, no. I understand the signal. But on a corporate
         | publication or some other piece of professional writing,
         | absolutely. Humans do use em-dashes on those.
        
       | sanjitb wrote:
       | > the update introduced an additional reward signal based on user
       | feedback--thumbs-up and thumbs-down data from ChatGPT. This
       | signal is often useful; a thumbs-down usually means something
       | went wrong.
       | 
       | > We also made communication errors. Because we expected this to
       | be a fairly subtle update, we didn't proactively announce it.
       | 
       | that doesn't sound like a "subtle" update to me. also, why is
       | "subtle" the metric here? i'm not even sure what it means in this
       | context.
        
       | ripvanwinkle wrote:
       | a well written postmortem and it raised my confidence in their
       | product in general
        
       | mensetmanusman wrote:
       | Chat GPT has taught me that I ask amazing questions!
        
       | fitsumbelay wrote:
       | how does such a specific kind of outcome happen without
       | intention?
        
       | mvdtnz wrote:
       | They need to be testing these models with non American testers in
       | their qualitative tests (not just end user A/B testing). Anyone
       | who has worked in professional settings with Americans knows that
       | sycophancy is ingrained deeply in the culture over there.
        
       | jozvolskyef wrote:
       | The side-by-side comparisons are not a good signal because the
       | models vary across multiple dimensions, but the user isn't given
       | the option to indicate the dimension on which they're scoring the
       | model.
       | 
       | The recent side-by-side comparisons presented a more accurate
       | model that communicates poorly vs a less accurate model with
       | slightly better communication.
        
       | Johanx64 wrote:
       | The thing that annoys me the most is when I ask it to generate
       | some code - actually no, most often than not I don't even ask it
       | to generate code, but ask some vaguely related programming
       | question - to which it replies with complete listing of code
       | (didn't ask for it, but alas).
       | 
       | Then I fix the code and tell it all the mistakes it has. And then
       | it does a 180 in tone, wherein - it starts talking as if I wrote
       | the code in the first place with - "yeah, obviously that wouldn't
       | work, so I fixed the issues in your code" and acts like a person
       | trying to save face and present the bugs it fixed as if the buggy
       | code was written by me all along.
       | 
       | That really gets me livid. LOL
        
       | zoogeny wrote:
       | Moments like these make me reevaluate the AI doomer view point.
       | We aren't just toying with access to dangerous ideas (biological
       | weapons, etc) we are toying with human psychology.
       | 
       | If something as obvious as harmful sycophancy can slip out so
       | easily, what subtle harms are being introduced. It's like lead in
       | paint (and gasoline) except rewiring our very brains. We won't
       | know that real problems for decades.
        
       ___________________________________________________________________
       (page generated 2025-05-02 23:00 UTC)