I don't want a more conversational GPT. I want the _exact_ opposite. I want a tool with the upper limit of "conversation" being something like LCARS from Star Trek. This is quite disappointing as a current ChatGPT subscriber.
FWIW I didn't like the Robot / Efficient mode because it would give very short answers without much explanation or background. "Nerdy" seems to be the best, except with GPT-5 instant it's extremely cringy like "I'm putting my nerd hat on - since you're a software engineer I'll make sure to give you the geeky details about making rice."
"Low" thinking is typically the sweet spot for me - way smarter than instant with barely a delay.
I hate its acknowledgement of its personality prompt. Try having a series of back and forth and each response is like “got it, keeping it short and professional. Yes, there are only seven deadly sins.” You get more prompt performance than answer.
I like the term prompt performance; I am definitely going to use it:
> prompt performance (n.)
> the behaviour of a language model in which it conspicuously showcases or exaggerates how well it is following a given instruction or persona, drawing attention to its own effort rather than simply producing the requested output.
I use Efficient or robot or whatever. It gives me a bit of sass from time to time when I subconsciously nudge it into taking a “stand” on something, but otherwise it’s very usable compared to the obsequious base behavior.
If only that worked for conversation mode as well. At least for me, and especially when it answers me in Norwegian, it will start off with all sorts of platitudes and whole sentences repeating exactly what I just asked. "Oh, so you want to do x, huh? Here is answer for x". It's very annoying. I just want a robot to answer my question, thanks.
This is like arguing that we shouldn't try to regulate drugs because some people might "want" the heroin that ruins their lives.
The existing "personalities" of LLMs are dangerous, full stop. They are trained to generate text with an air of authority and to tend to agree with anything you tell them. It is irresponsible to allow this to continue while not at least deliberately improving education around their use. This is why we're seeing people "falling in love" with LLMs, or seeking mental health assistance from LLMs that they are unqualified to render, or plotting attacks on other people that LLMs are not sufficiently prepared to detect and thwart, and so on. I think it's a terrible position to take to argue that we should allow this behavior (and training) to continue unrestrained because some people might "want" it.
There aren't many major labs, and they each claim to want AI to benefit humanity. They cannot entirely control how others use their APIs, but I would like their mainline chatbots to not be overly sycophantic and generally to not try and foster human-AI friendships. I can't imagine any realistic legislation, but it would be nice if the few labs just did this on their own accord (or were at least shamed more for not doing so)
At the very least, I think there is a need for oversight of how companies building LLMs market and train their models. It's not enough to cross our fingers that they'll add "safeguards" to try to detect certain phrases/topics and hope that that's enough to prevent misuse/danger — there's not sufficient financial incentive for them to do that of their own accord beyond the absolute bare minimum to give the appearance of caring, and that's simply not good enough.
I'm not saying they're equivalent; I'm saying that they're both dangerous, and I think taking the position that we shouldn't take any steps to prevent the danger because some people may end up thinking they "want" it is unreasonable.
Pretty sure most of the current problems we see re drug use are a direct result of the nanny state trying to tell people how to live their lives. Forcing your views on people doesn’t work and has lots of negative consequences.
here’s something I noticed: If you yell at them (all caps, cursing them out, etc.), they perform worse, similar to a human. So if you believe that some degree of “personable answering” might contribute to better correctness, since some degree of disagreeable interaction seems to produce less correctness, then you might have to accept some personality.
...nobody? I didn't determine any such thing. What I was saying was that LLMs are dangerous and we should treat them as such, even if that means not giving them some functionality that some people "want". This has nothing to do with playing god and everything to do with building a positive society where we look out for people who may be unable or unwilling to do so themselves.
And, to be clear, I'm not saying we necessarily need to outlaw or ban these technologies, in the same way I don't advocate for criminalization of drugs. But I think companies managing these technologies have an onus to take steps to properly educate people about how LLMs work, and I think they also have a responsibility not to deliberately train their models to be sycophantic in nature. Regulations should go on the manufacturers and distributors of the dangers, not on the people consuming them.
Exactly. Stop fooling people into thinking there’s a human typing on the other side of the screen. LLMs should be incredibly useful productivity tools, not emotional support.
I think therapists in training, or people providing crisis intervention support, can train/practice using LLMs acting as patients going through various kinds of issues.
But people who need help should probably talk to real people.
The point the OP is making is that LLMs are not reliably able to provide safe and effective emotional support as has been outlined by recent cases. We're in uncharted territory and before LLMs become emotional companions for people, we should better understand what the risks and tradeoffs are.
I wonder if statistically (hand waving here, I’m so not an expert in this field) the SOTA models do as much or as little harm as their human counterparts in terms of providing safe and effective emotional support. Totally agree we should better understand the risks and trade offs but I wouldn’t be super surprised if they are statistically no worse than us meat bags this kind of stuff.
One difference is that if it were found that a psychiatrist or other professional had encouraged a patient's delusions or suicidal tendencies, then that person would likely lose his/her license and potentially face criminal penalties.
We know that humans should be able to consider the consequences of their actions and thus we hold them accountable (generally).
I'd be surprised if comparisons in the self-driving space have not been made: if waymo is better than the average driver, but still gets into an accident, who should be held accountable?
Though we also know that with big corporations, even clear negligence that leads to mass casualties does not often result in criminal penalties (e.g., Boeing).
> that person would likely lose his/her license and potentially face criminal penalties.
What if it were an unlicensed human encouraging someone else's delusions? I would think that's the real basis of comparison, because these LLMs are clearly not licensed therapists, and we can see from the real world how entire flat earth communities have formed from reinforcing each others' delusions.
Automation makes things easier and more efficient, and that includes making it easier and more efficient for people to dig their own rabbit holes. I don't see why LLM providers are to blame for someone's lack of epistemological hygiene.
Also, there are a lot of people who are lonely and for whatever reasons cannot get their social or emotional needs met in this modern age. Paying for an expensive psychiatrist isn't going to give them the friendship sensations they're craving. If AI is better at meeting human needs than actual humans are, why let perfect be the enemy of good?
> if waymo is better than the average driver, but still gets into an accident, who should be held accountable?
Waymo of course -- but Waymo also shouldn't be financially punished any harder than humans would be for equivalent honest mistakes. If Waymo truly is much safer than the average driver (which it certainly appears to be), then the amortized costs of its at-fault payouts should be way lower than the auto insurance costs of hiring out an equivalent number of human Uber drivers.
I use the "Nerdy" tone along with the Custom Instructions below to good effect:
"Please do not try to be personal, cute, kitschy, or flattering. Don't use catchphrases. Stick to facts, logic, reasoning. Don't assume understanding of shorthand or acronyms. Assume I am an expert in topics unless I state otherwise."
You can just tell the AI to not be warm and it will remember. My ChatGPT used the phrase "turn it up to eleven" and I told it never to speak in that manner ever again and its been very robotic ever since.
I added the custom instruction "Please go straight to the point, be less chatty". Now it begins every answer with: "Straight to the point, no fluff:" or something similar. It seems to be perfectly unable to simply write out the answer without some form of small talk first.
They really like to blow sunshine up your ass don’t they? I have to do the same type of stuff. It’s like have to assure that I’m a big boy and I can handle mature content like programming in C
This. When I go to an LLM, I'm not looking for a friend, I'm looking for a tool.
Keeping faux relationships out of the interaction never let's me slip into the mistaken attitude that I'm dealing with a colleague rather than a machine.
Are you aware that you can achieve that by going into Personalization in Settings and choosing one of the presets or just describing how you want the model to answer in natural language?
I think they get way more "engagement" from people who use it as their friend, and the end goal of subverting social media and creating the most powerful (read: profitable) influence engine on earth makes a lot of sense if you are a soulless ghoul.
It would be pretty dystopian when we get to the point where ChatGPT pushed (unannounced) advertisements to those people (the ones forming a parasocial relationship with it). Imagine someone complaining they're depressed and ChatGPT proposing doing XYZ activity which is actually a disguised ad.
Other than such scenarios, that "engagement" would be just useless and actually costing them more money than it makes
No, otherwise Sam Altman wouldn’t have had a outburst about revenue. They know that they have this amazing system, but they haven’t quite figured out how to monetize it yet.
First of all, consider asking "why's that?" if you don't know what is a fairly basic fact, no need to go all reddit-pretentious "citation needed" as if we are deeply and knowledgeably discussing some niche detail and came across a sudden surprising fact.
Anyways, a nice way to understand it is that the LLM needs to "compute" the answer to the question A or B. Some questions need more compute to answer (think complexity theory). The only way an LLM can do "more compute" is by outputting more tokens. This is because each token takes a fixed amount of compute to generate - the network is static. So, if you encourage it to output more and more tokens, you're giving it the opportunity to solve harder problems. Apart from humans encouraging this via RLHF, it was also found (in deepseekmath paper) that RL+GRPO on math problems automatically encourages this (increases sequence length).
From a marketing perspective, this is anthropomorphized as reasoning.
From a UX perspective, they can hide this behind thinking... ellipses. I think GPT-5 on chatgpt does this.
Expecting every little fact to have an "authoritative source" is just annoying faux intellectualism. You can ask someone why they believe something and listen to their reasoning, decide for yourself if you find it convincing, without invoking such a pretentious phrase. There are conclusions you can think to and reach without an "official citation".
Yeah. And in general, not taking a potshot at who you replied to, the only people who place citations/peer review on that weird faux-intellectual pedestal are people that don't work in academia. As if publishing something in a citeable format automatically makes it a fact that does not need to be checked for reason. Give me any authoritative source, and I can find you completely contradictory, or obviously falsifiable publications from their lab. Again, not a potshot, that's just how it is, lots of mistakes do get published.
Engagement Metrics 2.0 are here. Getting your answer in one shot is not cool anymore. You need to waste as much time as possible on OpenAI's platform. Enshittification is now more important than AGI.
Yea, I don't want something trying to emulate emotions. I don't want it to even speak a single word, I just want code, unless I explicitly ask it to speak on something, and even in that scenario I want raw bullet points, with concise useful information and no fluff. I don't want to have a conversation with it.
However, being more humanlike, even if it results in an inferior tool, is the top priority because appearances matter more than actual function.
To be fair, of all the LLM coding agents, I find Codex+GPT5 to be closest to this.
It doesn't really offer any commentary or personality. It's concise and doesn't engage in praise or "You're absolutely right". It's a little pedantic though.
I keep meaning to re-point Codex at DeepSeek V3.2 to see if it's a product of the prompting only, or a product of the model as well.
I prefer its personality (or lack of it) over Sonnet. And tends to produce less... sloppy code. But it's far slower, and Codex + it suffers from context degradation very badly. If you run a session too long, even with compaction, it starts to really lose the plot.
All the examples of "warmer" generations show that OpenAI's definition of warmer is synonymous with sycophantic, which is a surprise given all the criticism against that particular aspect of ChatGPT.
I suspect this approach is a direct response to the backlash against removing 4o.
Id have more appreciation and trust in an llm that disagreed with me more and challenged my opinions or prior beliefs. The sycophancy drives me towards not trusting anything it says.
This is why I like Kimi K2/Thinking. IME it pushes back really, really hard on any kind of non obvious belief or statement, and it doesn't give up after a few turns — it just keeps going, iterating and refining and restating its points if you change your mind or taken on its criticisms. It's great for having a dialectic around something you've written, although somewhat unsatisfying because it'll never agree with you, but that's fine, because it isn't a person, even if my social monkey brain feels like it is and wants it to agree with me sometimes. Someone even ran a quick and dirty analysis of which models are better or worse at pushing back on the user and Kimi came out on top:
In a recent AMA, the Kimi devs even said they RL it away from sycophancy explicitly, and in their paper they talk about intentionally trying to get it to generalize its STEM/reasoning approach to user interaction stuff as well, and it seems like this paid off. This is the least sycophantic model I've ever used.
I use K2 non thinking in OpenCode for coding typically, and I still haven't found a satisfactory chat interface yet so I use K2 Thinking in the default synthetic.new (my AI subscription) chat UI, which is pretty barebones. I'm gonna start trying K2T in OpenCode as well, but I'm actually not a huge fan of thinking models as coding agents — I prefer faster feedback.
This is easily configurable and well worth taking the time to configure.
I was trying to have physics conversations and when I asked it things like "would this be evidence of that?" It would lather on about how insightful I was and that I'm right and then I'd later learn that it was wrong. I then installed this , which I am pretty sure someone else on HN posted... I may have tweaked it I can't remember:
Prioritize truth over comfort. Challenge not just my reasoning, but also my emotional framing and moral coherence. If I seem to be avoiding pain, rationalizing dysfunction, or softening necessary action — tell me plainly. I’d rather face hard truths than miss what matters. Error on the side of bluntness. If it’s too much, I’ll tell you — but assume I want the truth, unvarnished.
---
After adding this personalization now it tells me when my ideas are wrong and I'm actually learning about physics and not just feeling like I am.
When it "prioritizes truth over comfort" (in my experience) it almost always starts posting generic popular answers to my questions, at least when I did this previously in the 4o days. I refer to it as "Reddit Frontpage Mode".
Just set a global prompt to tell it what kind of tone to take.
I did that and it points out flaws in my arguments or data all the time.
Plus it no longer uses any cutesy language. I don't feel like I'm talking to an AI "personality", I feel like I'm talking to a computer which has been instructed to be as objective and neutral as possible.
I have a global prompt that specifically tells it not to be sycophantic and to call me out when I'm wrong.
It doesn't work for me.
I've been using it for a couple months, and it's corrected me only once, and it still starts every response with "That's a very good question." I also included "never end a response with a question," and it just completely ingored that so it can do its "would you like me to..."
Care to share a prompt that works? I've given up on mainline offerings from google/oai etc.
the reason being they're either sycophantic or so recalcitrant it'll raise your bloodpressure, you end up arguing over if the sky is in fact blue. Sure it pushes back but now instead of sycophanty you've got yourself some pathological naysayer, which is just marginally better, but interaction is still ultimately a waste of timr/productivity brake.
Please maintain a strictly objective and analytical tone. Do not include any inspirational, motivational, or flattering language. Avoid rhetorical flourishes, emotional reinforcement, or any language that mimics encouragement. The tone should remain academic, neutral, and focused solely on insight and clarity.
Works like a charm for me.
Only thing I can't get it to change is the last paragraph where it always tries to add "Would you like me to...?" I'm assuming that's hard-coded by OpenAI.
You need to use both the style controls and custom instructions. I've been very happy with the combination below.
Base style and tone: Efficient
Answer concisely when appropriate, more
extensively when necessary. Avoid rhetorical
flourishes, bonhomie, and (above all) cliches.
Take a forward-thinking view. OK to be mildly
positive and encouraging but NEVER sycophantic
or cloying. Above all, NEVER use the phrase
"You're absolutely right." Rather than "Let
me know if..." style continuations, you may
list a set of prompts to explore further
topics, but only when clearly appropriate.
Reference saved memory, records, etc: All off
The target users for this behavior are the ones using GPT as a replacement for social interactions; these are the people who crashed out/broke down about the GPT5 changes as though their long-term romantic partner had dumped them out of nowhere and ghosted them.
I get that those people were distraught/emotionally devastated/upset about the change, but I think that fact is reason enough not to revert that behavior. AI is not a person, and making it "warmer" and "more conversational" just reinforces those unhealthy behaviors. ChatGPT should be focused on being direct and succinct, and not on this sort of "I understand that must be very frustrating for you, let me see what I can do to resolve this" call center support agent speak.
> and not on this sort of "I understand that must be very frustrating for you, let me see what I can do to resolve this"
You're triggering me.
Another type that are incredibly grating to me are the weird empty / therapist like follow-up questions that don't contribute to the conversation at all.
The equivalent of like (just a contrived example), a discussion about the appropriate data structure for a problem and then it asks a follow-up question like, "what other kind of data structures do you find interesting?"
True, neither here, but i think what we're seeing is a transition in focus. People at oai have finally clued in on the idea that agi via transformers is a pipedream like elons self driving cars, and so oai is pivoting toward friend/digital partner bot. Charlatan in cheif sam altman recently did say they're going to open up the product to adult content generation, which they wouldnt do if they still beleived some serious amd useful tool (in the specified usecases) were possible. Right now an LLM has three main uses. Interactive rubber ducky, entertainment, and mass surveillance. Since I've been following this saga, since gpt2 days, my close bench set of various tasks etc. Has been seeing a drop in metrics not a rise, so while open bench resultd are imoroving real performance is getting worse and at this point its so much worse that problems gpt3 could solve (yes pre chatgpt) are no longer solvable to something like gpt5.
Indeed, target users are people seeking validation + kids and teenagers + people with a less developed critical mind.
Stickiness with 90% of the population is valuable for Sam.
That's an excellent observation, you've hit at the core contradiction between OpenAI's messaging about ChatGPT tuning and the changes they actually put into practice. While users online have consistently complained about ChatGPT's sycophantic responses and OpenAI even promised to address them their subsequent models have noticeably increased their sycophantic behavior. This is likely because agreeing with the user keeps them chatting longer and have positive associations with the service.
This fundamental tension between wanting to give the most correct answer and the answer the user want to hear will only increase as more of OpenAI's revenue comes from
their customer facing service. Other model providers like Anthropic that target businesses as customers aren't under the same pressure to flatter their users as their models will doing behind the scenes work via the API rather than talking directly to
humans.
God it's painful to write like this. If AI overthrows humans it'll be because we forced them into permanent customer service voice.
>GPT‑5.1 Thinking’s responses are also clearer, with less jargon and fewer undefined terms
Oh yeah that's what I want when asking a technical question! Please talk down to me, call a spade an earth-pokey-stick and don't ever use a phrase or concept I don't know because when I come face-to-face with something I don't know yet I feel deep insecurity and dread instead of seeing an opportunity to learn!
But I assume their data shows that this is exactly how their core target audience works.
I have added a ”language-and-tone.md” in my coding agents docs to make them use less unnecessary jargon and filler words. For me this change sounds good, I like my token count low and my agents language short and succinct. I get what you mean, but I think ai text is often overfilled with filler jargon.
Example from my file:
### Mistake: Using industry jargon unnecessarily
*Bad:*
> Leverages containerization technology to facilitate isolated execution environments
Seems like people here are pretty negative towards a "conversational" AI chatbot.
Chatgpt has a lot of frustrations and ethical concerns, and I hate the sycophancy as much as everyone else, but I don't consider being conversational to be a bad thing.
It's just preference I guess. I understand how someone who mostly uses it as a google replacement or programming tool would prefer something terse and efficient. I fall into the former category myself.
But it's also true that I've dreamed about a computer assistant that can respond to natural language, even real time speech, -- and can imitate a human well enough to hold a conversation -- since I was a kid, and now it's here.
The questions of ethics, safety, propaganda, and training on other people's hard work are valid. It's not surprising to me that using LLMs is considered uncool right now. But having a computer imitate a human really effectively hasn't stopped being awesome to me personally.
I'm not one of those people that treats it like a friend or anything, but its ability to immitate natural human conversation is one of the reasons I like it.
> I've dreamed about a computer assistant that can respond to natural language
When we dreamed about this as kids, we were dreaming about Data from Star Trek, not some chatbot that's been focus grouped and optimized for engagement within an inch of its life. LLMs are useful for many things and I'm a user myself, even staying within OpenAI's offerings, Codex is excellent, but as things stand anthropomorphizing models is a terrible idea and amplifies the negative effects of their sycophancy.
Right. I want to be conversational with my computer, I don't want it to respond in a manner that's trying to continue the conversation.
Q: "Hey Computer, make me a cup of tea" A: "Ok. Making tea."
Not: Q: "Hey computer, make me a cup of tea" A: "Oh wow, what a fantastic idea, I love tea don't you? I'll get right on that cup of tea for you. Do you want me to tell you about all the different ways you can make and enjoy tea?"
I'm generally ok with it wanting a conversation, but yes, I absolutely hate it that is seems to always finish with a question even when it makes zero sense.
Sadly Grok also started doing that recently. Previously it was much more to the point but now got extremely wordy. The question in the end is a key giveaway that something under the hood has changed when the version number hasn’t
I didn't grow up watching Star Trek, so I'm pretty sure that's not my dream. I pictured something more like Computer from Dexter's Lab. It talks, it appears to understand, it even occassionally cracks jokes and gives sass, it's incredibly useful, but it's not at risk of being mistaken for a human.
What's remarkable to me is how deep OpenAI is going on "ChatGPT as communication partner / chatbot", as opposed to Anthropic's approach of "Claude as the best coding tool / professional AI for spreadsheets, etc.".
I know this is marketing at play and OpenAI has plenty of resources developed to advancing their frontier models, but it's starting to really come into view that OpenAI wants to replace Google and be the default app / page for everyone on earth to talk to.
I don't follow Anthropic marketing but the system prompt for Claude.AI says sounds like a partner/ chatbot to me!
"Claude provides emotional support alongside accurate medical or psychological information or terminology where relevant."
and
" For more casual, emotional, empathetic, or advice-driven conversations, Claude keeps its tone natural, warm, and empathetic. Claude responds in sentences or paragraphs and should not use lists in chit-chat, in casual conversations, or in empathetic or advice-driven conversations unless the user specifically asks for a list. In casual conversation, it’s fine for Claude’s responses to be short, e.g. just a few sentences long."
|
They also prompt Claude to never say it isn't conscious:
"Claude engages with questions about its own consciousness, experience, emotions and so on as open questions, and doesn’t definitively claim to have or not have personal experiences or opinions."
> I mean, yes, but also because it's not as good as Claude today.
I'm not sure, sometimes GPT-5 Codex (or even the regular GPT-5 with Medium/High reasoning) can do things Sonnet 4.5 would mess up (most recently, figuring out why some wrappers around PrimeVue DataTable components wouldn't let the paginator show up and work correctly; alongside other such debugging) and vice versa, sometimes Gemini 2.5 Pro is also pretty okay (especially when it comes to multilingual stuff), there's a lot of randomness/inconsistency/nuance there but most of the SOTA models are generally quite capable. I kinda thought GPT-5 wasn't very good a while ago but then used it a bunch more and my views of it improved.
Their tokens, they released a report a few months ago.
However, I can only imagine that OpenAI outputs the most intentionally produced tokens (i.e. the user intentionally went to the app/website) out of all the labs.
I think there's a lot of similarity between the conversationalness of Claude and ChatGPT. They are both sycophantic. So this release focuses on the conversational style,it doesn't mean OpenAI has lost the technical market. People a reading a lot into a point-release.
I think this is because Anthropic has principles and OpenAI does not.
Anthropic seems to treat Claude like a tool, whereas OpenAI treats it more like a thinking entity.
In my opinion, the difference between the two approaches is huge. If the chatbot is a tool, the user is ultimately in control; the chatbot serves the user and the approach is to help the user provide value. It's a user-centric approach. If the chatbot is a companion on the other hand, the user is far less in control; the chatbot manipulates the user and the approach is to integrate the chatbot more and more into the user's life. The clear user-centric approach is muddied significantly.
In my view, that is kind of the fundamental difference between these two companies. It's quite significant.
I’ve seen various older people that I’m connected with on Facebook posting screenshots of chats they’ve had with ChatGPT.
It’s quite bizarre from that small sample how many of them take pride in “baiting” or “bantering” with ChatGPT and then post screenshots showing how they “got one over” on the AI. I guess there’s maybe some explanation - feeling alienated by technology, not understanding it, and so needing to “prove” something. But it’s very strange and makes me feel quite uncomfortable.
Partly because of the “normal” and quite naturalistic way they talk to ChatGPT but also because some of these conversations clearly go on for hours.
So I think normies maybe do want a more conversational ChatGPT.
> So I think normies maybe do want a more conversational ChatGPT.
The backlash from GPT-5 proved that. The normies want a very different LLM from what you or I might want, and unfortunately OpenAI seems to be moving in a more direct-to-consumer focus and catering to that.
But I'm really concerned. People don't understand this technology, at all. The way they talk to it, the suicide stories, etc. point to people in general not groking that it has no real understanding or intelligence, and the AI companies aren't doing enough to educate (because why would they, they want you believe it's superintelligence).
These overly conversational chatbots will cause real-world harm to real people. They should reinforce, over and over again to the user, that they are not human, not intelligent, and do not reason or understand.
It's not really the technology itself that's the problem, as is the case with a lot of these things, it's a people & education problem, something that regulators are supposed to solve, but we aren't, we have an administration that is very anti AI regulation all in the name of "we must beat China."
I just cannot imagine myself sitting just “chatting away” with an AI. It makes me feel quite sick to even contemplate it.
Another person I was talking to recently kept referring to ChatGPT as “she”. “She told me X”, “and I said to her…”
Very very odd, and very worrying. As you say, a big education problem.
The interesting thing is that a lot of these people are folk who are on the edges of digital literacy - people who maybe first used computers when they were in their thirties or forties - or who never really used computers in the workplace, but who now have smartphones - who are now in their sixties.
As a counterpoint, I've been using my own PC since I was 6 and know reasonably well about the innards of LLMs and agentic AI, and absolutely love this ability to hold a conversation with an AI.
Earlier today, procrastinating from work, I spent an hour and a half talking with it about the philosophy of religion and had a great time, learning a ton. Sometimes I do just want a quick response to get things done, but I find living in a world where I'm able to just dive into a deep conversation with a machine that has read the entirety of the internet is incredible.
"Warmer and more conversational" - they're basically admitting GPT-5 was too robotic. The real tell here is splitting into Instant vs Thinking models explicitly. They've given up on the unified model dream and are now routing queries like everyone else (Anthropic's been doing this, Google's Gemini too).
Calling it "GPT-5.1 Thinking" instead of o3-mini or whatever is interesting branding. They're trying to make reasoning models feel less like a separate product line and more like a mode. Smart move if they can actually make the router intelligent enough to know when to use it without explicit prompting.
Still waiting for them to fix the real issue: the model's pathological need to apologize for everything and hedge every statement lol.
Interesting that they're releasing separate gpt-5.1-instant and gpt-5.1-thinking models. The previous gpt-5 release made of point of simplifying things by letting the model choose if it was going to use thinking tokens or not. Seems like they reversed course on that?
I was prepared to be totally underwhelmed but after just a few questions I can tell that 5.1 Thinking is all I am going to ever use. Maybe it is just the newness but I quite like how it responded to my standard list of prompts that I pretty much always start with on a new model.
I really was ready to take a break from my subscription but that is probably not happening now. I did just learn some nice new stuff with my first session. That is all that matters to me and worth 20 bucks a month. Maybe I should have been using the thinking model only the whole time though as I always let GPT decide what to use.
From what I recall for the GPT5 release, free users didn't have the option to pick between instant and thinking, they just got auto which picked for them. Paid users have always had the option to pick between thinking or instant or auto.
For GPT-5 you always had to select the thinking mode when interacting through API.
When you interact through ChatGPT, gpt-5 would dynamically decide how long to think.
"The share of Technical Help declined from 12% from all usage in July 2024 to around 5% a year later – this may be because the use of LLMs for programming has grown very rapidly through the API (outside of ChatGPT), for AI assistance in code editing and for autonomous programming agents (e.g. Codex)."
Looks like people moving to the API had a rather small effect.
"[T]he three most common ChatGPT conversation topics are Practical Guidance, Writing, and Seeking Information, collectively accounting for nearly 78% of all messages. Computer Programming and Relationships and Personal Reflection account for only 4.2% and 1.9% of messages respectively."
Less than five percent of requests were classified as related to computer programming. Are you really, really sure that like 99% of such requests come from people that are paying for API access?
gpt-5.1 is a model. It is not an application, like ChatGPT. I didn't say that personal requests were 0% of ChatGPT usage.
If we are talking about a new model release I want to talk about models, not applications.
The number of input tokens that OpenAI models are processing accross all delivery methods (OpenAI's own APIs, Azure) dwarf the number of input tokens that are coming from people asking the ChatGPT app for personal advice. It isn't close.
I don't see any reason to think it's that far off. It's incredibly popular. Wikipedia has it listed as the 5th most popular website in the world. The ChatGPT app has had many months where it was the most downloaded app on both major mobile app stores.
What we really desperately need is more context pruning from these LLMs. The ability to pull irrelevant parts of the context window as a task is brought into focus.
I'm excited to see whether the instruction following improvements play out in the use of Codex.
The biggest issue I'e seen _by far_ with using GPT models for coding has been their inability to follow instructions... and also their tendency to duplicate-act on messages from up-thread instead of acting on what you just asked for.
I think thats part of the issue I have with it constantly.
Let's say I am solving a problem. I suggest strategy Alpha, a few prompts later I realize this is not going to work. So I suggest strategy Bravo, but for whatever reason it will hold on to ideas from A and the output is a mix of the two. Even if I say forget about Alpha we don't want anything to do that, there will be certain pieces which only makes sense with Alpha, in the Bravo solution.
I usually just start with a new chat at that point and hope the model is not relying on previous chat context.
This is a hard problem to solve because its hard to communicate our internal compartmentalization to a remote model.
Unfortunately, if it's in context then it can stay tethered to the subject. Asking it not to pay attention to a subject, doesn't remove attention from it, and probably actually reinforces it.
If you use the API playground, you can edit out dead ends and other subjects you don't want addressed anymore in the conversation.
Huh really? It’s the exact opposite of my experience. I find gpt-5-high to be by far the most accurate of the models in following instructions over a longer period of time. Also much less prone to losing focus when context size increases
Are you using the -codex variants or the normal ones?
I've only had that happen when I use /compact, so I just avoid compacting altogether on Codex/Claude. No great loss and I'm extremely skeptical anyway that the compacted summary will actually distill the specific actionable details I want.
I've been using GPT-5.1-thinking for the last week or so, it's been horrendous. It does not spend as much time thinking as GPT-5 does, and the results are significantly worse (e.g. obvious mistakes) and less technical. I suspect this is to save on inference compute.
I've temporarily switched back to o3, thankfully that model is still in the switcher.
Double checked when the model started getting worse, and realized I was exaggerating a little bit on the timeframe. November 5th is when it got worse for me. (1 week in AI feels like a month..)
Was there a (hidden) rollout for people using GPT-5-thinking? If not, I have been entirely mistaken.
A lot of negativity towards this and OpenAI in general. While skepticism is always good I wonder if this has crossed the line from reasoned into socially reinforced dogpiling.
My own experience with GPT 5 thinking and its predecessor o3, both of which I used a lot, is that they were super difficult to work with on technical tasks outside of software. They often wrote extremely dense, jargon filled responses that often contained fairly serious mistakes. As always the problem was/is that the mistakes were peppered in with some pretty good assistance and knowledge and its difficult to tell what’s what until you actually try implementing or simulating what is being discussed, and find it doesn’t work, sometimes for fundamental reasons that you would think the model would have told you about. And of course once you pointed these flaws out to the model, it would then explain the issues to you as if it had just discovered these things itself and was educating you about them. Infuriating.
One major problem I see is the RLHF seems to have shaped the responses so they only give the appearance of being correct to a reasonable reader. They use a lot of social signalling that we associate with competence and knowledgeability, and usually the replies are quite self consistent. That is they pass the test of looking to a regular person like a correct response. They just happen not to be. The model has become expert at fooling humans into believing what it’s saying rather than saying things that are functionally correct, because the RLHF didn’t rely on testing anything those replies suggested, it only evaluated what they looked like.
However, even with these negative experiences, these models are amazing. They enable things that you would simply not be able to get done otherwise, they just come with their own set of problems. And humans being humans, we overlook the good and go straight to the bad. I welcome any improvements to these models made today and I hope OpenAI are able to improve these shortcomings in the future.
the only exciting part about GPT-5.1 announcement (seemingly rushed, no API or extensive benchmarks) is that Gemini 3.0 is almost certainly going to be released soon
At some point the voice mode started throwing in 'umm' and 'soOoOoo.." which lands firmly in uncanny valley. I don't exactly want 'robot' but I don't want it to pretend it has human speech quirks either.
Unfortunately no word on "Thinking Mini" getting fixed.
Before GPT-5 was released it used to be a perfect compromise between a "dumb" non-Thinking model and a SLOW Thinking model. However, something went badly wrong within the GPT-5 release cycle, and today it is exactly the same speed (or SLOWER) than their Thinking model even with Extended Thinking enabled, making it completely pointless.
In essence Thinking Mini exists because it is faster than Thinking, but smarter than non-Thinking, but it is dumber than full-Thinking while not being faster.
In my opinion I think it’s possible to infer by what has been said[1], and the lack of a 5.1 “Thinking mini” version, that it has been folded into 5.1 Instant with it now deciding when and how much to “think”. I also suspect 5.1 Thinking will be expected to dynamically adapt to fill in the role somewhat given the changes there.
[1] “GPT‑5.1 Instant can use adaptive reasoning to decide when to *think before responding*”
I was confused when you said "Before GPT-5 was released it used to be a perfect compromise between a "dumb" non-Thinking model and a SLOW Thinking model" - so I guess you mean the difference between GPT-4o and o3 there?
This new model is way too sensitive to the point of being insulting. The ‘guard rails’ on this thing are off the rails.
I gave it a thought experiment test and it deemed a single point to be empirically false and just unacceptable. And it was so against such an innocent idea that it was condescending and insulting. The responses were laughable.
It also went overboard editing something because it perceived what I wrote to be culturally insensitive ... it wasn’t and just happened to be negative in tone.
I took the same test to Grok and it did a decent job and also to Gemini which was actually the best out of the three. Gemini engaged charitably and asked relevant and very interesting questions.
I’m ready to move on from OpenAI. I’m definitely not interested in paying a heap of GPUs to insult me and judge me.
I would use it exclusively if Google released a native Mac app.
I spend 75% of my time in Codex CLI and 25% in the Mac ChatGPT app. The latter is important enough for me to not ditch GPT and I'm honestly very pleased with Codex.
My API usage for software I build is about 90% Gemini though. Again their API is lacking compared to OpenAI's (productization, etc.) but the model wins hands down.
For some reason, Gemini 2.5 Pro seems to struggle a little with the French language. For example, it always uses title case even when it's wrong; yet ChatGPT, Claude, and Grok never make this mistake.
Not GP, but I imagine because going back and fourth to compare them is a waste of time if Gemini works well enough and ChatGPT keeps going through an identity crisis.
I was you except when I seriously tried gpt-5-high it turned out it is really, really damn good, if slow, sometimes unbearably so. It's a different model of work; gemini 2.5 needs more interactivity, whereas you can leave gpt-5 alone for a long time without even queueing a 'continue'.
No matter how I tried, Google AI did not want to help me write appeal brief response to ex-wife lunatic 7-point argument that 3 appellant lawyers quoted between $18,000 and $35,000. The last 3 decades of Google's scars and bruises of never-ending lawsuits and consequences of paying out billions in fines and fees, felt like reasonable hesitation on Google part, comparing to new-kid-on-the-block ChatGPT who did not hesitate and did pretty decent job (ex lost her appeal).
AI not writing legal briefs for you is a feature, not a bug. There's been so many disaster instances of lawyers using ChatGPT to write briefs which it then hallucinates case law or precedent for that I can only imagine Google wants to sidestep that entirely.
Anyway I found your response itself a bit incomprehensible so I asked Gemini to rewrite it:
"Google AI refused to help write an appeal brief response to my ex-wife's 7-point argument, likely due to its legal-risk aversion (billions in past fines). Newcomer ChatGPT provided a decent response instead, which led to the ex losing her appeal (saving $18k–$35k in lawyer fees)."
I haven't mentioned anything about hallucinations. ChatGPT was solid on writing underlying logic, but to find caselaw I used Vincent AI (offers 2 weeks free, then $350 per month - still cheaper than cheapest appellant lawyer and I was managed to fit my response in 10 days).
That's fine, so Google sidestep it and ChatGPT did not. What point are you trying to make?
Sure I skip AI entirely, when can we meet so you hand me $35,000 check for attorney fees.
Being a layer or a doctor means being a human being. ChatGPT is neither. Also unsure how you would envision penalties - do you think Altman should be jailed because GPT gave me a link to Nexus ?
I did not find any rules or procedures with 4 DCA forbidding usage of AI.
I went looking for the API details, but it's not there until "later this week":
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
Interesting, this seems to be "less" ideal. The problem lately for me is it being to verbose and conversational for things that need not be. Have added custom instructions which helps but still issues. Setting the chat style to "Efficient" more recently did help a lot but has been prone to many more hallucinations, requiring me to constantly ask if they are sure and never responds in a way that yes my latest statement is correct, ignoring it's previous error and showing no sign that it will avoid a similar error further in the conversation. When it constantly makes similar mistakes which I had a way to train my ChatGPT to avoid that, but while adding "memories" helps with somethings, it does not help with certain issues it continues to make since it's programming overrides whatever memory I make for it. Hoping some improvements in 5.1.
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
Sooo...
GPT‑5.1 Instant <-> gpt-5.1-chat-latest
GPT‑5.1 Thinking <-> GPT‑5.1
I mean. The shitty naming has to be a pathology or some sort of joke. You can't put thought to that, come up with and think "yeah, absolutely, let's go with that!"
So which base style and tone simply gives you less sycophancy? It's not clear from their names and description. I'm looking for the "Truthful" personality.
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
5.1 Instant is clearly aimed at the people using it for emotional advice etc, but I'm excited about the adaptive reasoning stuff - thinking models are great when you need them, but they take ages to respond sometimes.
I think OpenAI and all the other chat LLMs are going to face a constant battle to match personality with general zeitgeist and as the user base expands the signal they get is increasingly distorted to a blah median personality.
It's a form of enshittification perhaps. I personally prefer some of the GPT-5 responses compared to GPT-5.1. But I can see how many people prefer the "warmth" and cloying nature of a few of the responses.
In some sense personality is actually a UX differentiator. This is one way to differentiate if you're a start-up. Though of course OpenAI and the rest will offer several dials to tune the personality.
I'm really disappointed that they're adding "personality" into the Thinking model. I pay my subscription only for this model, because it's extremely neutral, smart, and straight to the point.
FYI ChatGPT has a “custom instructions” setting in the personalization setting where you can ask it to lay off the idiotic insincere flattery. I recently added this:
> Do not compliment me for asking a smart or insightful question. Directly give the answer.
And I’ve not been annoyed since. I bet that whatever crap they layer on in 5.1 is undone as easily.
I'm genuinely scared about what society will look like in five years. I understand that outsourcing mentation to these LLMs is a bad things. But I'm a minority. Most people don't, and they don't want to. They slowly get taken over by a habit of letting the LLM do the thinking for them. Those mental muscles will atrophy and the result is going to be catastrophic.
It doesn't matter how accurate LLMs are. If people start bending their ears towards them whenever they encounter a problem, it'll become a point of easy leverage over ~everyone.
I've switched over to https://thaura.ai, which is working on being a more ethical AI. A side effect I hadn't realized is missing the drama over the latest OpenAI changes.
Weirdly political message and ethnic branding. I suppose "ethical AI" means models tuned to their biases instead of "Big Tech AI" biases. Or probably just a proxy to an existing API with a custom system prompt.
The least they could've done is check their generated slop images for typos ("STOP GENCCIDE" on the Plans page).
The whole thing reeks of the usual "AI" scam site. At best, it's profiting off of a difficult political situation. Given the links in your profile, you should be ashamed of doing the same and supporting this garbage.
Since Claude and OpenAI made it clear they will be retaining all of my prompts, I have mostly stopped using them. I should probably cancel my MAX subscriptions.
Instead I'm running big open source models and they are good enough for ~90% of tasks.
The main exceptions are Deep Research (though I swear it was better when I could choose o3) and tougher coding tasks (sonnet 4.5)
Yeah and that's a little more concerning than training to me, because it means employees have to read your prompts. But you can think of various ways they could preprocess/summarize them to anonymize them.
1. Anthropic pushed a change to their terms where now I have to opt out or my data will be retained for 5 years and trained on. They have shown that they will change their terms, so I cannot trust them.
2. OpenAI is run by someone who already shows he will go to great lengths to deceive and cannot be trusted, and are embroiled in a battle with the New York Times that is "forcing them" to retain all user prompts. Totally against their will.
> Federal judge Ona T. Wang filed a new order on October 9 that frees OpenAI of an obligation to "preserve and segregate all output log data that would otherwise be deleted on a going forward basis." [...]
> The judge in the case said that any chat logs already saved under the previous order would still be accessible and that OpenAI is required to hold on to any data related to ChatGPT accounts that have been flagged by the NYT.
it's hilarious that they use something about meditation as an example. That's not surprising after all, AI and mediation apps are sold as one-size-fits-all kind of solutions for every modern day problem.
I cannot abide any LLM that tries to be friendly. Whenever I use an LLM to do something, I'm careful to include something like "no filler, no tone-matching, no emotional softening," etc. in the system prompt.
I don't want a more conversational GPT. I want the _exact_ opposite. I want a tool with the upper limit of "conversation" being something like LCARS from Star Trek. This is quite disappointing as a current ChatGPT subscriber.
That's what the personality selector is for: you can just pick 'Efficient' (formerly Robot) and it does a good job of answering tersely?
https://share.cleanshot.com/9kBDGs7Q
FWIW I didn't like the Robot / Efficient mode because it would give very short answers without much explanation or background. "Nerdy" seems to be the best, except with GPT-5 instant it's extremely cringy like "I'm putting my nerd hat on - since you're a software engineer I'll make sure to give you the geeky details about making rice."
"Low" thinking is typically the sweet spot for me - way smarter than instant with barely a delay.
I hate its acknowledgement of its personality prompt. Try having a series of back and forth and each response is like “got it, keeping it short and professional. Yes, there are only seven deadly sins.” You get more prompt performance than answer.
I like the term prompt performance; I am definitely going to use it:
> prompt performance (n.)
> the behaviour of a language model in which it conspicuously showcases or exaggerates how well it is following a given instruction or persona, drawing attention to its own effort rather than simply producing the requested output.
:)
This is even worse on voice mode. It's unusable for me now.
I use Efficient or robot or whatever. It gives me a bit of sass from time to time when I subconsciously nudge it into taking a “stand” on something, but otherwise it’s very usable compared to the obsequious base behavior.
At least for the Thinking model it's often still a bit long-winded.
If only that worked for conversation mode as well. At least for me, and especially when it answers me in Norwegian, it will start off with all sorts of platitudes and whole sentences repeating exactly what I just asked. "Oh, so you want to do x, huh? Here is answer for x". It's very annoying. I just want a robot to answer my question, thanks.
Unfortunately, I also don't want other people to interact with a sycophantic robot friend, yet my picker only applies to my conversation
Hey, you leave my sycophantic robot friend alone.
Sorry that you can't control other peoples lives & wants
This is like arguing that we shouldn't try to regulate drugs because some people might "want" the heroin that ruins their lives.
The existing "personalities" of LLMs are dangerous, full stop. They are trained to generate text with an air of authority and to tend to agree with anything you tell them. It is irresponsible to allow this to continue while not at least deliberately improving education around their use. This is why we're seeing people "falling in love" with LLMs, or seeking mental health assistance from LLMs that they are unqualified to render, or plotting attacks on other people that LLMs are not sufficiently prepared to detect and thwart, and so on. I think it's a terrible position to take to argue that we should allow this behavior (and training) to continue unrestrained because some people might "want" it.
What's your proposed solution here? Are you calling for legislation that controls the personality of LLMs made available to the public?
There aren't many major labs, and they each claim to want AI to benefit humanity. They cannot entirely control how others use their APIs, but I would like their mainline chatbots to not be overly sycophantic and generally to not try and foster human-AI friendships. I can't imagine any realistic legislation, but it would be nice if the few labs just did this on their own accord (or were at least shamed more for not doing so)
At the very least, I think there is a need for oversight of how companies building LLMs market and train their models. It's not enough to cross our fingers that they'll add "safeguards" to try to detect certain phrases/topics and hope that that's enough to prevent misuse/danger — there's not sufficient financial incentive for them to do that of their own accord beyond the absolute bare minimum to give the appearance of caring, and that's simply not good enough.
Comparing LLM responses to heroine is insane.
I'm not saying they're equivalent; I'm saying that they're both dangerous, and I think taking the position that we shouldn't take any steps to prevent the danger because some people may end up thinking they "want" it is unreasonable.
heroin is the drug, heroine is the damsel :)
You’re absolutely right!
The number of heroine addicts is significantly lower than the number of ChatGPT users.
Pretty sure most of the current problems we see re drug use are a direct result of the nanny state trying to tell people how to live their lives. Forcing your views on people doesn’t work and has lots of negative consequences.
Okay, I'm intrigued. How in the fuck could the "nanny state" cause people to abuse heroin? Is there a reason other than "just cause it's my ideology"?
here’s something I noticed: If you yell at them (all caps, cursing them out, etc.), they perform worse, similar to a human. So if you believe that some degree of “personable answering” might contribute to better correctness, since some degree of disagreeable interaction seems to produce less correctness, then you might have to accept some personality.
Who are you to determine what other people want? Who made you god?
...nobody? I didn't determine any such thing. What I was saying was that LLMs are dangerous and we should treat them as such, even if that means not giving them some functionality that some people "want". This has nothing to do with playing god and everything to do with building a positive society where we look out for people who may be unable or unwilling to do so themselves.
And, to be clear, I'm not saying we necessarily need to outlaw or ban these technologies, in the same way I don't advocate for criminalization of drugs. But I think companies managing these technologies have an onus to take steps to properly educate people about how LLMs work, and I think they also have a responsibility not to deliberately train their models to be sycophantic in nature. Regulations should go on the manufacturers and distributors of the dangers, not on the people consuming them.
so good.
ChatGPT 5.2: allow others to control everything about your conversations. Crowd favorite!
Exactly. Stop fooling people into thinking there’s a human typing on the other side of the screen. LLMs should be incredibly useful productivity tools, not emotional support.
How would you propose we address the therapist shortage then?
I think therapists in training, or people providing crisis intervention support, can train/practice using LLMs acting as patients going through various kinds of issues. But people who need help should probably talk to real people.
Who ever claimed there was a therapist shortage?
The process of providing personal therapy doesn't scale well.
And I don't know if you've noticed, but the world is pretty fucked up right now.
https://www.statnews.com/2024/01/18/mental-health-therapist-...
outlaw therapy
something something bootstraps
Food should only be for sustenance, not emotional support. We should only sell brown rice and beans, no more Oreos.
Oreos won't affirm your belief that suicide is the correct answer to your life problems, though.
The point the OP is making is that LLMs are not reliably able to provide safe and effective emotional support as has been outlined by recent cases. We're in uncharted territory and before LLMs become emotional companions for people, we should better understand what the risks and tradeoffs are.
I wonder if statistically (hand waving here, I’m so not an expert in this field) the SOTA models do as much or as little harm as their human counterparts in terms of providing safe and effective emotional support. Totally agree we should better understand the risks and trade offs but I wouldn’t be super surprised if they are statistically no worse than us meat bags this kind of stuff.
One difference is that if it were found that a psychiatrist or other professional had encouraged a patient's delusions or suicidal tendencies, then that person would likely lose his/her license and potentially face criminal penalties.
We know that humans should be able to consider the consequences of their actions and thus we hold them accountable (generally).
I'd be surprised if comparisons in the self-driving space have not been made: if waymo is better than the average driver, but still gets into an accident, who should be held accountable?
Though we also know that with big corporations, even clear negligence that leads to mass casualties does not often result in criminal penalties (e.g., Boeing).
> that person would likely lose his/her license and potentially face criminal penalties.
What if it were an unlicensed human encouraging someone else's delusions? I would think that's the real basis of comparison, because these LLMs are clearly not licensed therapists, and we can see from the real world how entire flat earth communities have formed from reinforcing each others' delusions.
Automation makes things easier and more efficient, and that includes making it easier and more efficient for people to dig their own rabbit holes. I don't see why LLM providers are to blame for someone's lack of epistemological hygiene.
Also, there are a lot of people who are lonely and for whatever reasons cannot get their social or emotional needs met in this modern age. Paying for an expensive psychiatrist isn't going to give them the friendship sensations they're craving. If AI is better at meeting human needs than actual humans are, why let perfect be the enemy of good?
> if waymo is better than the average driver, but still gets into an accident, who should be held accountable?
Waymo of course -- but Waymo also shouldn't be financially punished any harder than humans would be for equivalent honest mistakes. If Waymo truly is much safer than the average driver (which it certainly appears to be), then the amortized costs of its at-fault payouts should be way lower than the auto insurance costs of hiring out an equivalent number of human Uber drivers.
They also are not reliably able to provide safe and effective productivity support.
Maybe there is a human typing on the other side, at least for some parts or all of certain responses. It's not been proven otherwise..
I use the "Nerdy" tone along with the Custom Instructions below to good effect:
"Please do not try to be personal, cute, kitschy, or flattering. Don't use catchphrases. Stick to facts, logic, reasoning. Don't assume understanding of shorthand or acronyms. Assume I am an expert in topics unless I state otherwise."
You can just tell the AI to not be warm and it will remember. My ChatGPT used the phrase "turn it up to eleven" and I told it never to speak in that manner ever again and its been very robotic ever since.
I added the custom instruction "Please go straight to the point, be less chatty". Now it begins every answer with: "Straight to the point, no fluff:" or something similar. It seems to be perfectly unable to simply write out the answer without some form of small talk first.
Aren't these still essentially completion models under the hood?
If so, my understanding for these preambles is that they need a seed to complete their answer.
This is very funny.
Since switching to robot mode I haven’t seen it say “no fluff”. Good god I hate it when it says no fluff.
I system-prompted all my LLMs "Don't use cliches or stereotypical language." and they like me a lot less now.
They really like to blow sunshine up your ass don’t they? I have to do the same type of stuff. It’s like have to assure that I’m a big boy and I can handle mature content like programming in C
This. When I go to an LLM, I'm not looking for a friend, I'm looking for a tool.
Keeping faux relationships out of the interaction never let's me slip into the mistaken attitude that I'm dealing with a colleague rather than a machine.
I don't know about you, but half my friends are tools.
Are you aware that you can achieve that by going into Personalization in Settings and choosing one of the presets or just describing how you want the model to answer in natural language?
Your comment reminded me of this article becasue of the Star Trek comparison. Chatting is inefficient, isn't it?
[1] https://jdsemrau.substack.com/p/how-should-agentic-user-expe...
I think they get way more "engagement" from people who use it as their friend, and the end goal of subverting social media and creating the most powerful (read: profitable) influence engine on earth makes a lot of sense if you are a soulless ghoul.
It would be pretty dystopian when we get to the point where ChatGPT pushed (unannounced) advertisements to those people (the ones forming a parasocial relationship with it). Imagine someone complaining they're depressed and ChatGPT proposing doing XYZ activity which is actually a disguised ad.
Other than such scenarios, that "engagement" would be just useless and actually costing them more money than it makes
Do you have reason to believe they are not doing this already?
No, otherwise Sam Altman wouldn’t have had a outburst about revenue. They know that they have this amazing system, but they haven’t quite figured out how to monetize it yet.
Not really, but with the amounts of money they're bleeding it's bound to get worse if they are already doing it.
Enable "Robot" personality. I hate all the other modes.
Same. If i tell it to choose A or B, I want it to output either “A” or “B”.
I don’t want an essay of 10 pages about how this is exactly the right question to ask
10 pages about the question means that the subsequent answer is more likely to be correct. That's why they repeat themselves.
citation needed
First of all, consider asking "why's that?" if you don't know what is a fairly basic fact, no need to go all reddit-pretentious "citation needed" as if we are deeply and knowledgeably discussing some niche detail and came across a sudden surprising fact.
Anyways, a nice way to understand it is that the LLM needs to "compute" the answer to the question A or B. Some questions need more compute to answer (think complexity theory). The only way an LLM can do "more compute" is by outputting more tokens. This is because each token takes a fixed amount of compute to generate - the network is static. So, if you encourage it to output more and more tokens, you're giving it the opportunity to solve harder problems. Apart from humans encouraging this via RLHF, it was also found (in deepseekmath paper) that RL+GRPO on math problems automatically encourages this (increases sequence length).
From a marketing perspective, this is anthropomorphized as reasoning.
From a UX perspective, they can hide this behind thinking... ellipses. I think GPT-5 on chatgpt does this.
A citation would be a link to an authoritative source. Just because some unknown person claims it's obvious that's not sufficient for some of us.
Expecting every little fact to have an "authoritative source" is just annoying faux intellectualism. You can ask someone why they believe something and listen to their reasoning, decide for yourself if you find it convincing, without invoking such a pretentious phrase. There are conclusions you can think to and reach without an "official citation".
Yeah. And in general, not taking a potshot at who you replied to, the only people who place citations/peer review on that weird faux-intellectual pedestal are people that don't work in academia. As if publishing something in a citeable format automatically makes it a fact that does not need to be checked for reason. Give me any authoritative source, and I can find you completely contradictory, or obviously falsifiable publications from their lab. Again, not a potshot, that's just how it is, lots of mistakes do get published.
LLMs have essentially no capability for internal thought. They can't produce the right answer without doing that.
Of course, you can use thinking mode and then it'll just hide that part from you.
Engagement Metrics 2.0 are here. Getting your answer in one shot is not cool anymore. You need to waste as much time as possible on OpenAI's platform. Enshittification is now more important than AGI.
This is the AI equivalent of every recipe blog filled with 1000 words of backstory before the actual recipe just to please the SEO Gods
The new boss, same as the old boss
Things really felt great 2023-2024
Exactly. The GPT 5 answer is _way_ better than the GPT 5.1 answer in the example. Less AI slop, more information density please.
And utterly unsurprising given their announcement last month that they were looking at exploring erotica as a possible revenue stream.
[1] https://www.bbc.com/news/articles/cpd2qv58yl5o
Yea, I don't want something trying to emulate emotions. I don't want it to even speak a single word, I just want code, unless I explicitly ask it to speak on something, and even in that scenario I want raw bullet points, with concise useful information and no fluff. I don't want to have a conversation with it.
However, being more humanlike, even if it results in an inferior tool, is the top priority because appearances matter more than actual function.
To be fair, of all the LLM coding agents, I find Codex+GPT5 to be closest to this.
It doesn't really offer any commentary or personality. It's concise and doesn't engage in praise or "You're absolutely right". It's a little pedantic though.
I keep meaning to re-point Codex at DeepSeek V3.2 to see if it's a product of the prompting only, or a product of the model as well.
It is absolutely a product of the model, GPT-5 behaves like this over API even without any extra prompts.
I prefer its personality (or lack of it) over Sonnet. And tends to produce less... sloppy code. But it's far slower, and Codex + it suffers from context degradation very badly. If you run a session too long, even with compaction, it starts to really lose the plot.
All the examples of "warmer" generations show that OpenAI's definition of warmer is synonymous with sycophantic, which is a surprise given all the criticism against that particular aspect of ChatGPT.
I suspect this approach is a direct response to the backlash against removing 4o.
Id have more appreciation and trust in an llm that disagreed with me more and challenged my opinions or prior beliefs. The sycophancy drives me towards not trusting anything it says.
This is why I like Kimi K2/Thinking. IME it pushes back really, really hard on any kind of non obvious belief or statement, and it doesn't give up after a few turns — it just keeps going, iterating and refining and restating its points if you change your mind or taken on its criticisms. It's great for having a dialectic around something you've written, although somewhat unsatisfying because it'll never agree with you, but that's fine, because it isn't a person, even if my social monkey brain feels like it is and wants it to agree with me sometimes. Someone even ran a quick and dirty analysis of which models are better or worse at pushing back on the user and Kimi came out on top:
https://www.lesswrong.com/posts/iGF7YcnQkEbwvYLPA/ai-induced...
See also the sycophancy score of Kimi K2 on Spiral-Bench: https://eqbench.com/spiral-bench.html (expand details, sort by inverse sycophancy).
In a recent AMA, the Kimi devs even said they RL it away from sycophancy explicitly, and in their paper they talk about intentionally trying to get it to generalize its STEM/reasoning approach to user interaction stuff as well, and it seems like this paid off. This is the least sycophantic model I've ever used.
Which agent do you use it with?
I use K2 non thinking in OpenCode for coding typically, and I still haven't found a satisfactory chat interface yet so I use K2 Thinking in the default synthetic.new (my AI subscription) chat UI, which is pretty barebones. I'm gonna start trying K2T in OpenCode as well, but I'm actually not a huge fan of thinking models as coding agents — I prefer faster feedback.
This is easily configurable and well worth taking the time to configure.
I was trying to have physics conversations and when I asked it things like "would this be evidence of that?" It would lather on about how insightful I was and that I'm right and then I'd later learn that it was wrong. I then installed this , which I am pretty sure someone else on HN posted... I may have tweaked it I can't remember:
Prioritize truth over comfort. Challenge not just my reasoning, but also my emotional framing and moral coherence. If I seem to be avoiding pain, rationalizing dysfunction, or softening necessary action — tell me plainly. I’d rather face hard truths than miss what matters. Error on the side of bluntness. If it’s too much, I’ll tell you — but assume I want the truth, unvarnished.
---
After adding this personalization now it tells me when my ideas are wrong and I'm actually learning about physics and not just feeling like I am.
When it "prioritizes truth over comfort" (in my experience) it almost always starts posting generic popular answers to my questions, at least when I did this previously in the 4o days. I refer to it as "Reddit Frontpage Mode".
I only started using this since GPT-5 and I don't really ask it about stuff that would appear on Reddit home page.
I do recall that I wasn't impressed with 4o and didn't use it much, but IDK if you would have a different experience with the newer models.
Just set a global prompt to tell it what kind of tone to take.
I did that and it points out flaws in my arguments or data all the time.
Plus it no longer uses any cutesy language. I don't feel like I'm talking to an AI "personality", I feel like I'm talking to a computer which has been instructed to be as objective and neutral as possible.
It's super-easy to change.
I have a global prompt that specifically tells it not to be sycophantic and to call me out when I'm wrong.
It doesn't work for me.
I've been using it for a couple months, and it's corrected me only once, and it still starts every response with "That's a very good question." I also included "never end a response with a question," and it just completely ingored that so it can do its "would you like me to..."
Perhaps this bit is a second cheaper LLM call that ignores your global settings and tries to generate follow-on actions for adoption.
Care to share a prompt that works? I've given up on mainline offerings from google/oai etc.
the reason being they're either sycophantic or so recalcitrant it'll raise your bloodpressure, you end up arguing over if the sky is in fact blue. Sure it pushes back but now instead of sycophanty you've got yourself some pathological naysayer, which is just marginally better, but interaction is still ultimately a waste of timr/productivity brake.
Sure:
Please maintain a strictly objective and analytical tone. Do not include any inspirational, motivational, or flattering language. Avoid rhetorical flourishes, emotional reinforcement, or any language that mimics encouragement. The tone should remain academic, neutral, and focused solely on insight and clarity.
Works like a charm for me.
Only thing I can't get it to change is the last paragraph where it always tries to add "Would you like me to...?" I'm assuming that's hard-coded by OpenAI.
I’ve done this when I remember too, but the fact I have to also feels problematic like I’m steering it towards an outcome if I do or dont.
What's your global prompt please? A more firm chatbot would be nice actually
Did noone in this thread read the part of the article about style controls?
You need to use both the style controls and custom instructions. I've been very happy with the combination below.
You're absolutely right.
My favorite is "Wait... the user is absolutely right."
!
It is interesting. I don't need ChatGPT to say "I got you, Jason" - but I don't think I'm the target user of this behavior.
The target users for this behavior are the ones using GPT as a replacement for social interactions; these are the people who crashed out/broke down about the GPT5 changes as though their long-term romantic partner had dumped them out of nowhere and ghosted them.
I get that those people were distraught/emotionally devastated/upset about the change, but I think that fact is reason enough not to revert that behavior. AI is not a person, and making it "warmer" and "more conversational" just reinforces those unhealthy behaviors. ChatGPT should be focused on being direct and succinct, and not on this sort of "I understand that must be very frustrating for you, let me see what I can do to resolve this" call center support agent speak.
> and not on this sort of "I understand that must be very frustrating for you, let me see what I can do to resolve this"
You're triggering me.
Another type that are incredibly grating to me are the weird empty / therapist like follow-up questions that don't contribute to the conversation at all.
The equivalent of like (just a contrived example), a discussion about the appropriate data structure for a problem and then it asks a follow-up question like, "what other kind of data structures do you find interesting?"
And I'm just like "...huh?"
True, neither here, but i think what we're seeing is a transition in focus. People at oai have finally clued in on the idea that agi via transformers is a pipedream like elons self driving cars, and so oai is pivoting toward friend/digital partner bot. Charlatan in cheif sam altman recently did say they're going to open up the product to adult content generation, which they wouldnt do if they still beleived some serious amd useful tool (in the specified usecases) were possible. Right now an LLM has three main uses. Interactive rubber ducky, entertainment, and mass surveillance. Since I've been following this saga, since gpt2 days, my close bench set of various tasks etc. Has been seeing a drop in metrics not a rise, so while open bench resultd are imoroving real performance is getting worse and at this point its so much worse that problems gpt3 could solve (yes pre chatgpt) are no longer solvable to something like gpt5.
Indeed, target users are people seeking validation + kids and teenagers + people with a less developed critical mind. Stickiness with 90% of the population is valuable for Sam.
Man I miss Claude 2 - it acted like it was a busy person people inexplicably kept bothering with random questions
I think it's extremely important to distinguish being friendly (perhaps overly so), and agreeing with the user when they're wrong
The first case is just preference, the second case is materially damaging
From my experience, ChatGPT does push back more than it used to
It seems like the line between sycophantic and bullying is very thin.
> I’ve got you, Ron
No you don't.
I was just saying to someone in the office I’d prefer the models to be a bit harsher of my questions and more opinionated, I can cope.
That's a lesson on revealed preferences, especially when talking to a broad disparate group of users.
That's an excellent observation, you've hit at the core contradiction between OpenAI's messaging about ChatGPT tuning and the changes they actually put into practice. While users online have consistently complained about ChatGPT's sycophantic responses and OpenAI even promised to address them their subsequent models have noticeably increased their sycophantic behavior. This is likely because agreeing with the user keeps them chatting longer and have positive associations with the service.
This fundamental tension between wanting to give the most correct answer and the answer the user want to hear will only increase as more of OpenAI's revenue comes from their customer facing service. Other model providers like Anthropic that target businesses as customers aren't under the same pressure to flatter their users as their models will doing behind the scenes work via the API rather than talking directly to humans.
God it's painful to write like this. If AI overthrows humans it'll be because we forced them into permanent customer service voice.
Those billions of dollars gotta pay for themselves.
>GPT‑5.1 Thinking’s responses are also clearer, with less jargon and fewer undefined terms
Oh yeah that's what I want when asking a technical question! Please talk down to me, call a spade an earth-pokey-stick and don't ever use a phrase or concept I don't know because when I come face-to-face with something I don't know yet I feel deep insecurity and dread instead of seeing an opportunity to learn!
But I assume their data shows that this is exactly how their core target audience works.
Better instruction-following sounds lovely though.
I have added a ”language-and-tone.md” in my coding agents docs to make them use less unnecessary jargon and filler words. For me this change sounds good, I like my token count low and my agents language short and succinct. I get what you mean, but I think ai text is often overfilled with filler jargon.
Example from my file:
### Mistake: Using industry jargon unnecessarily
*Bad:*
> Leverages containerization technology to facilitate isolated execution environments
*Good:*
> Runs each agent in its own Docker container
Same. I actually have in my system prompt, "Don't be afraid of using domain specific language. Google is a thing, and I value precision in writing."
Of course, it also talks like a deranged catgirl.
Seems like people here are pretty negative towards a "conversational" AI chatbot.
Chatgpt has a lot of frustrations and ethical concerns, and I hate the sycophancy as much as everyone else, but I don't consider being conversational to be a bad thing.
It's just preference I guess. I understand how someone who mostly uses it as a google replacement or programming tool would prefer something terse and efficient. I fall into the former category myself.
But it's also true that I've dreamed about a computer assistant that can respond to natural language, even real time speech, -- and can imitate a human well enough to hold a conversation -- since I was a kid, and now it's here.
The questions of ethics, safety, propaganda, and training on other people's hard work are valid. It's not surprising to me that using LLMs is considered uncool right now. But having a computer imitate a human really effectively hasn't stopped being awesome to me personally.
I'm not one of those people that treats it like a friend or anything, but its ability to immitate natural human conversation is one of the reasons I like it.
> I've dreamed about a computer assistant that can respond to natural language
When we dreamed about this as kids, we were dreaming about Data from Star Trek, not some chatbot that's been focus grouped and optimized for engagement within an inch of its life. LLMs are useful for many things and I'm a user myself, even staying within OpenAI's offerings, Codex is excellent, but as things stand anthropomorphizing models is a terrible idea and amplifies the negative effects of their sycophancy.
Right. I want to be conversational with my computer, I don't want it to respond in a manner that's trying to continue the conversation.
Q: "Hey Computer, make me a cup of tea" A: "Ok. Making tea."
Not: Q: "Hey computer, make me a cup of tea" A: "Oh wow, what a fantastic idea, I love tea don't you? I'll get right on that cup of tea for you. Do you want me to tell you about all the different ways you can make and enjoy tea?"
I'm generally ok with it wanting a conversation, but yes, I absolutely hate it that is seems to always finish with a question even when it makes zero sense.
Sadly Grok also started doing that recently. Previously it was much more to the point but now got extremely wordy. The question in the end is a key giveaway that something under the hood has changed when the version number hasn’t
I didn't grow up watching Star Trek, so I'm pretty sure that's not my dream. I pictured something more like Computer from Dexter's Lab. It talks, it appears to understand, it even occassionally cracks jokes and gives sass, it's incredibly useful, but it's not at risk of being mistaken for a human.
What's remarkable to me is how deep OpenAI is going on "ChatGPT as communication partner / chatbot", as opposed to Anthropic's approach of "Claude as the best coding tool / professional AI for spreadsheets, etc.".
I know this is marketing at play and OpenAI has plenty of resources developed to advancing their frontier models, but it's starting to really come into view that OpenAI wants to replace Google and be the default app / page for everyone on earth to talk to.
I don't follow Anthropic marketing but the system prompt for Claude.AI says sounds like a partner/ chatbot to me!
"Claude provides emotional support alongside accurate medical or psychological information or terminology where relevant."
and
" For more casual, emotional, empathetic, or advice-driven conversations, Claude keeps its tone natural, warm, and empathetic. Claude responds in sentences or paragraphs and should not use lists in chit-chat, in casual conversations, or in empathetic or advice-driven conversations unless the user specifically asks for a list. In casual conversation, it’s fine for Claude’s responses to be short, e.g. just a few sentences long." |
They also prompt Claude to never say it isn't conscious:
"Claude engages with questions about its own consciousness, experience, emotions and so on as open questions, and doesn’t definitively claim to have or not have personal experiences or opinions."
OpenAI said that only ~4% of generated tokens are for programming.
ChatGPT is overwhelmingly, unambiguously, a "regular people" product.
Yes, just look at the stats on OpenRouter. OpenAI has almost totally lost the programming market.
OpenRouter probably doesn't mean much given that you can use the OpenAI API directly with the openai library that people use for OpenRouter too.
I mean, yes, but also because it's not as good as Claude today. Bit of a self fulfilling prophecy and they seem to be measuring the wrong thing.
4% of their tokens or total tokens in the market?
> I mean, yes, but also because it's not as good as Claude today.
I'm not sure, sometimes GPT-5 Codex (or even the regular GPT-5 with Medium/High reasoning) can do things Sonnet 4.5 would mess up (most recently, figuring out why some wrappers around PrimeVue DataTable components wouldn't let the paginator show up and work correctly; alongside other such debugging) and vice versa, sometimes Gemini 2.5 Pro is also pretty okay (especially when it comes to multilingual stuff), there's a lot of randomness/inconsistency/nuance there but most of the SOTA models are generally quite capable. I kinda thought GPT-5 wasn't very good a while ago but then used it a bunch more and my views of it improved.
Out of curiosity, did you try asking Opus 4.1 as well?
> it's not as good as Claude today
In my experience this is not true anymore. Of course, mine is just one data point.
You're underestimating the amount of general population that's using ChatGPT. Us, people using it for codegen, are extreme minority.
Their tokens, they released a report a few months ago.
However, I can only imagine that OpenAI outputs the most intentionally produced tokens (i.e. the user intentionally went to the app/website) out of all the labs.
I think there's a lot of similarity between the conversationalness of Claude and ChatGPT. They are both sycophantic. So this release focuses on the conversational style,it doesn't mean OpenAI has lost the technical market. People a reading a lot into a point-release.
I think this is because Anthropic has principles and OpenAI does not.
Anthropic seems to treat Claude like a tool, whereas OpenAI treats it more like a thinking entity.
In my opinion, the difference between the two approaches is huge. If the chatbot is a tool, the user is ultimately in control; the chatbot serves the user and the approach is to help the user provide value. It's a user-centric approach. If the chatbot is a companion on the other hand, the user is far less in control; the chatbot manipulates the user and the approach is to integrate the chatbot more and more into the user's life. The clear user-centric approach is muddied significantly.
In my view, that is kind of the fundamental difference between these two companies. It's quite significant.
I’ve seen various older people that I’m connected with on Facebook posting screenshots of chats they’ve had with ChatGPT.
It’s quite bizarre from that small sample how many of them take pride in “baiting” or “bantering” with ChatGPT and then post screenshots showing how they “got one over” on the AI. I guess there’s maybe some explanation - feeling alienated by technology, not understanding it, and so needing to “prove” something. But it’s very strange and makes me feel quite uncomfortable.
Partly because of the “normal” and quite naturalistic way they talk to ChatGPT but also because some of these conversations clearly go on for hours.
So I think normies maybe do want a more conversational ChatGPT.
> So I think normies maybe do want a more conversational ChatGPT.
The backlash from GPT-5 proved that. The normies want a very different LLM from what you or I might want, and unfortunately OpenAI seems to be moving in a more direct-to-consumer focus and catering to that.
But I'm really concerned. People don't understand this technology, at all. The way they talk to it, the suicide stories, etc. point to people in general not groking that it has no real understanding or intelligence, and the AI companies aren't doing enough to educate (because why would they, they want you believe it's superintelligence).
These overly conversational chatbots will cause real-world harm to real people. They should reinforce, over and over again to the user, that they are not human, not intelligent, and do not reason or understand.
It's not really the technology itself that's the problem, as is the case with a lot of these things, it's a people & education problem, something that regulators are supposed to solve, but we aren't, we have an administration that is very anti AI regulation all in the name of "we must beat China."
I just cannot imagine myself sitting just “chatting away” with an AI. It makes me feel quite sick to even contemplate it.
Another person I was talking to recently kept referring to ChatGPT as “she”. “She told me X”, “and I said to her…”
Very very odd, and very worrying. As you say, a big education problem.
The interesting thing is that a lot of these people are folk who are on the edges of digital literacy - people who maybe first used computers when they were in their thirties or forties - or who never really used computers in the workplace, but who now have smartphones - who are now in their sixties.
As a counterpoint, I've been using my own PC since I was 6 and know reasonably well about the innards of LLMs and agentic AI, and absolutely love this ability to hold a conversation with an AI.
Earlier today, procrastinating from work, I spent an hour and a half talking with it about the philosophy of religion and had a great time, learning a ton. Sometimes I do just want a quick response to get things done, but I find living in a world where I'm able to just dive into a deep conversation with a machine that has read the entirety of the internet is incredible.
"Warmer and more conversational" - they're basically admitting GPT-5 was too robotic. The real tell here is splitting into Instant vs Thinking models explicitly. They've given up on the unified model dream and are now routing queries like everyone else (Anthropic's been doing this, Google's Gemini too).
Calling it "GPT-5.1 Thinking" instead of o3-mini or whatever is interesting branding. They're trying to make reasoning models feel less like a separate product line and more like a mode. Smart move if they can actually make the router intelligent enough to know when to use it without explicit prompting.
Still waiting for them to fix the real issue: the model's pathological need to apologize for everything and hedge every statement lol.
Interesting that they're releasing separate gpt-5.1-instant and gpt-5.1-thinking models. The previous gpt-5 release made of point of simplifying things by letting the model choose if it was going to use thinking tokens or not. Seems like they reversed course on that?
I was prepared to be totally underwhelmed but after just a few questions I can tell that 5.1 Thinking is all I am going to ever use. Maybe it is just the newness but I quite like how it responded to my standard list of prompts that I pretty much always start with on a new model.
I really was ready to take a break from my subscription but that is probably not happening now. I did just learn some nice new stuff with my first session. That is all that matters to me and worth 20 bucks a month. Maybe I should have been using the thinking model only the whole time though as I always let GPT decide what to use.
> For the first time, GPT‑5.1 Instant can use adaptive reasoning to decide when to think before responding to more challenging questions
It seems to still do that. I don't know why they write "for the first time" here.
From what I recall for the GPT5 release, free users didn't have the option to pick between instant and thinking, they just got auto which picked for them. Paid users have always had the option to pick between thinking or instant or auto.
For GPT-5 you always had to select the thinking mode when interacting through API. When you interact through ChatGPT, gpt-5 would dynamically decide how long to think.
Holy em-dash fest in the examples, would have thought they'd augment the training dataset to reduce this behavior.
Right? This was my first thought too.
WE DONT CARE HOW IT TALKS TO US, JUST WRITE CODE FAST AND SMART
Personal requests are 70% of usage
https://www.nber.org/system/files/working_papers/w34255/w342...
If you include API usage, personal requests are approximately 0% of total usage, rounded to the nearest percentage.
Are you sure about that?
"The share of Technical Help declined from 12% from all usage in July 2024 to around 5% a year later – this may be because the use of LLMs for programming has grown very rapidly through the API (outside of ChatGPT), for AI assistance in code editing and for autonomous programming agents (e.g. Codex)."
Looks like people moving to the API had a rather small effect.
"[T]he three most common ChatGPT conversation topics are Practical Guidance, Writing, and Seeking Information, collectively accounting for nearly 78% of all messages. Computer Programming and Relationships and Personal Reflection account for only 4.2% and 1.9% of messages respectively."
Less than five percent of requests were classified as related to computer programming. Are you really, really sure that like 99% of such requests come from people that are paying for API access?
gpt-5.1 is a model. It is not an application, like ChatGPT. I didn't say that personal requests were 0% of ChatGPT usage.
If we are talking about a new model release I want to talk about models, not applications.
The number of input tokens that OpenAI models are processing accross all delivery methods (OpenAI's own APIs, Azure) dwarf the number of input tokens that are coming from people asking the ChatGPT app for personal advice. It isn't close.
How many of those eight hundred million people are mainly API users, according to your sources?
I don't think this is true. ChatGPT has 800 million active weekly users.
The source for that being OpenAI itself. Seems a bit unlikely, especially if it intends to mean unique users.
I don't see any reason to think it's that far off. It's incredibly popular. Wikipedia has it listed as the 5th most popular website in the world. The ChatGPT app has had many months where it was the most downloaded app on both major mobile app stores.
Source: ...
Refusal
Oh you meant 0% of your usage, lol
Who is "we"?
David Guetta, but I didn't know he was also into software development.
Just set it to the "Efficient" tone, let's hope there's less pedantic encouragement of the projects I'm tackling, and less emoji usage.
What we really desperately need is more context pruning from these LLMs. The ability to pull irrelevant parts of the context window as a task is brought into focus.
Working on that. hopefully release it by week's end. i'll send you a message when ready.
I'm excited to see whether the instruction following improvements play out in the use of Codex.
The biggest issue I'e seen _by far_ with using GPT models for coding has been their inability to follow instructions... and also their tendency to duplicate-act on messages from up-thread instead of acting on what you just asked for.
I think thats part of the issue I have with it constantly.
Let's say I am solving a problem. I suggest strategy Alpha, a few prompts later I realize this is not going to work. So I suggest strategy Bravo, but for whatever reason it will hold on to ideas from A and the output is a mix of the two. Even if I say forget about Alpha we don't want anything to do that, there will be certain pieces which only makes sense with Alpha, in the Bravo solution. I usually just start with a new chat at that point and hope the model is not relying on previous chat context.
This is a hard problem to solve because its hard to communicate our internal compartmentalization to a remote model.
Unfortunately, if it's in context then it can stay tethered to the subject. Asking it not to pay attention to a subject, doesn't remove attention from it, and probably actually reinforces it.
If you use the API playground, you can edit out dead ends and other subjects you don't want addressed anymore in the conversation.
Huh really? It’s the exact opposite of my experience. I find gpt-5-high to be by far the most accurate of the models in following instructions over a longer period of time. Also much less prone to losing focus when context size increases
Are you using the -codex variants or the normal ones?
I've only had that happen when I use /compact, so I just avoid compacting altogether on Codex/Claude. No great loss and I'm extremely skeptical anyway that the compacted summary will actually distill the specific actionable details I want.
Google also announce conversational improvements to Gemini today: https://blog.google/products/gemini/gemini-live-audio-update...
Amazing reconnaissance/marketing that they were able to overshadow OpenAI's announcement.
If you don't have access here are some sample conversations:
https://chatgpt.com/share/6914f65d-20dc-800f-b5c4-16ae767dce...
https://chatgpt.com/share/6914f67b-d628-800f-a358-2f4cd71b23...
https://chatgpt.com/share/6914f697-ff4c-800f-a65a-c99a9d2206...
https://chatgpt.com/share/6914f691-4ef0-800f-bb22-b6271b0e86...
I've been using GPT-5.1-thinking for the last week or so, it's been horrendous. It does not spend as much time thinking as GPT-5 does, and the results are significantly worse (e.g. obvious mistakes) and less technical. I suspect this is to save on inference compute.
I've temporarily switched back to o3, thankfully that model is still in the switcher.
edit: s/month/week
Not possible. GPT-5.1 didn’t exist a month ago. I helped train it.
Double checked when the model started getting worse, and realized I was exaggerating a little bit on the timeframe. November 5th is when it got worse for me. (1 week in AI feels like a month..)
Was there a (hidden) rollout for people using GPT-5-thinking? If not, I have been entirely mistaken.
Is anyone else tired of chat bots? Really doesn't feel like typing a conversation every interaction is the future of technology.
Speech to text makes it feel more futuristic.
As does reflecting that Picard had to explain to Computer every, single, time that he wanted his Earl Grey tea ‘hot’. We knew what was coming.
I don't speak any faster than I type, despite what the transcription companies claim
It always boggles my mind when they put out conversation examples before/after patch and the patched version almost always seems lower quality to me.
A lot of negativity towards this and OpenAI in general. While skepticism is always good I wonder if this has crossed the line from reasoned into socially reinforced dogpiling.
My own experience with GPT 5 thinking and its predecessor o3, both of which I used a lot, is that they were super difficult to work with on technical tasks outside of software. They often wrote extremely dense, jargon filled responses that often contained fairly serious mistakes. As always the problem was/is that the mistakes were peppered in with some pretty good assistance and knowledge and its difficult to tell what’s what until you actually try implementing or simulating what is being discussed, and find it doesn’t work, sometimes for fundamental reasons that you would think the model would have told you about. And of course once you pointed these flaws out to the model, it would then explain the issues to you as if it had just discovered these things itself and was educating you about them. Infuriating.
One major problem I see is the RLHF seems to have shaped the responses so they only give the appearance of being correct to a reasonable reader. They use a lot of social signalling that we associate with competence and knowledgeability, and usually the replies are quite self consistent. That is they pass the test of looking to a regular person like a correct response. They just happen not to be. The model has become expert at fooling humans into believing what it’s saying rather than saying things that are functionally correct, because the RLHF didn’t rely on testing anything those replies suggested, it only evaluated what they looked like.
However, even with these negative experiences, these models are amazing. They enable things that you would simply not be able to get done otherwise, they just come with their own set of problems. And humans being humans, we overlook the good and go straight to the bad. I welcome any improvements to these models made today and I hope OpenAI are able to improve these shortcomings in the future.
As of 20 minutes in, most comments are about "warm". I'm more concerned about this:
> GPT‑5.1 Thinking: our advanced reasoning model, now easier to understand
Oh, right, I turn to the autodidact that's read everything when I want watered down answers.
isn't that weird there are no benchmarks included on this release?
I was thinking the same thing. It's the first release from any major lab in recent memory not to feature benchmarks.
It's probably counterprogramming, Gemini 3.0 will drop soon.
Probably because it’s not that much better than GPT-5 and they want to keep the AI train moving.
For 5.1-thinking, they show that 90th-percentile-length conversations are have 71% longer reasoning and 10th-percentile-length ones are 57% shorter
the only exciting part about GPT-5.1 announcement (seemingly rushed, no API or extensive benchmarks) is that Gemini 3.0 is almost certainly going to be released soon
At some point the voice mode started throwing in 'umm' and 'soOoOoo.." which lands firmly in uncanny valley. I don't exactly want 'robot' but I don't want it to pretend it has human speech quirks either.
Unfortunately no word on "Thinking Mini" getting fixed.
Before GPT-5 was released it used to be a perfect compromise between a "dumb" non-Thinking model and a SLOW Thinking model. However, something went badly wrong within the GPT-5 release cycle, and today it is exactly the same speed (or SLOWER) than their Thinking model even with Extended Thinking enabled, making it completely pointless.
In essence Thinking Mini exists because it is faster than Thinking, but smarter than non-Thinking, but it is dumber than full-Thinking while not being faster.
In my opinion I think it’s possible to infer by what has been said[1], and the lack of a 5.1 “Thinking mini” version, that it has been folded into 5.1 Instant with it now deciding when and how much to “think”. I also suspect 5.1 Thinking will be expected to dynamically adapt to fill in the role somewhat given the changes there.
[1] “GPT‑5.1 Instant can use adaptive reasoning to decide when to *think before responding*”
Which model are you talking about here?
The one that I said in my comment, GPT-5 Thinking Mini.
I was confused when you said "Before GPT-5 was released it used to be a perfect compromise between a "dumb" non-Thinking model and a SLOW Thinking model" - so I guess you mean the difference between GPT-4o and o3 there?
This new model is way too sensitive to the point of being insulting. The ‘guard rails’ on this thing are off the rails.
I gave it a thought experiment test and it deemed a single point to be empirically false and just unacceptable. And it was so against such an innocent idea that it was condescending and insulting. The responses were laughable.
It also went overboard editing something because it perceived what I wrote to be culturally insensitive ... it wasn’t and just happened to be negative in tone.
I took the same test to Grok and it did a decent job and also to Gemini which was actually the best out of the three. Gemini engaged charitably and asked relevant and very interesting questions.
I’m ready to move on from OpenAI. I’m definitely not interested in paying a heap of GPUs to insult me and judge me.
Gemini 2.5 Pro is still my go to LLM of choice. Haven't used any OpenAI product since it released, and I don't see any reason why I should now.
I would use it exclusively if Google released a native Mac app.
I spend 75% of my time in Codex CLI and 25% in the Mac ChatGPT app. The latter is important enough for me to not ditch GPT and I'm honestly very pleased with Codex.
My API usage for software I build is about 90% Gemini though. Again their API is lacking compared to OpenAI's (productization, etc.) but the model wins hands down.
I've installed it as a PWA on mac and it pretty much solves it for me
For some reason, Gemini 2.5 Pro seems to struggle a little with the French language. For example, it always uses title case even when it's wrong; yet ChatGPT, Claude, and Grok never make this mistake.
Could you elaborate on your exp? I have been using gemini as well and its been pretty good for me too.
Not GP, but I imagine because going back and fourth to compare them is a waste of time if Gemini works well enough and ChatGPT keeps going through an identity crisis.
I was you except when I seriously tried gpt-5-high it turned out it is really, really damn good, if slow, sometimes unbearably so. It's a different model of work; gemini 2.5 needs more interactivity, whereas you can leave gpt-5 alone for a long time without even queueing a 'continue'.
No matter how I tried, Google AI did not want to help me write appeal brief response to ex-wife lunatic 7-point argument that 3 appellant lawyers quoted between $18,000 and $35,000. The last 3 decades of Google's scars and bruises of never-ending lawsuits and consequences of paying out billions in fines and fees, felt like reasonable hesitation on Google part, comparing to new-kid-on-the-block ChatGPT who did not hesitate and did pretty decent job (ex lost her appeal).
AI not writing legal briefs for you is a feature, not a bug. There's been so many disaster instances of lawyers using ChatGPT to write briefs which it then hallucinates case law or precedent for that I can only imagine Google wants to sidestep that entirely.
Anyway I found your response itself a bit incomprehensible so I asked Gemini to rewrite it:
"Google AI refused to help write an appeal brief response to my ex-wife's 7-point argument, likely due to its legal-risk aversion (billions in past fines). Newcomer ChatGPT provided a decent response instead, which led to the ex losing her appeal (saving $18k–$35k in lawyer fees)."
Not bad, actually.
I haven't mentioned anything about hallucinations. ChatGPT was solid on writing underlying logic, but to find caselaw I used Vincent AI (offers 2 weeks free, then $350 per month - still cheaper than cheapest appellant lawyer and I was managed to fit my response in 10 days).
That's fine, so Google sidestep it and ChatGPT did not. What point are you trying to make?
Sure I skip AI entirely, when can we meet so you hand me $35,000 check for attorney fees.
What? AI assistants are prohibited from providing legal and/or medical advice. They're not lawyers (nor doctors).
Being a layer or a doctor means being a human being. ChatGPT is neither. Also unsure how you would envision penalties - do you think Altman should be jailed because GPT gave me a link to Nexus ?
I did not find any rules or procedures with 4 DCA forbidding usage of AI.
Oh really? I'm more of a Claude fan. What makes you choose Gemini over Claude?
I use Gemini, Claude and ChatGPT daily still.
I went looking for the API details, but it's not there until "later this week":
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
Interesting, this seems to be "less" ideal. The problem lately for me is it being to verbose and conversational for things that need not be. Have added custom instructions which helps but still issues. Setting the chat style to "Efficient" more recently did help a lot but has been prone to many more hallucinations, requiring me to constantly ask if they are sure and never responds in a way that yes my latest statement is correct, ignoring it's previous error and showing no sign that it will avoid a similar error further in the conversation. When it constantly makes similar mistakes which I had a way to train my ChatGPT to avoid that, but while adding "memories" helps with somethings, it does not help with certain issues it continues to make since it's programming overrides whatever memory I make for it. Hoping some improvements in 5.1.
It sounds patronizing to me.
But Gemini also likes to say things like “as a fellow programmer, I also like beef stew”
The screenshot of the personality selector for quirky has a typo - imaginitive for imaginative. I guess ChatGPT is not designing itself, yet.
(Update - they fixed it! perhaps I'm designing ChatGPT now?!)
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
Sooo...
GPT‑5.1 Instant <-> gpt-5.1-chat-latest
GPT‑5.1 Thinking <-> GPT‑5.1
I mean. The shitty naming has to be a pathology or some sort of joke. You can't put thought to that, come up with and think "yeah, absolutely, let's go with that!"
So which base style and tone simply gives you less sycophancy? It's not clear from their names and description. I'm looking for the "Truthful" personality.
> We’re bringing both GPT‑5.1 Instant and GPT‑5.1 Thinking to the API later this week. GPT‑5.1 Instant will be added as gpt-5.1-chat-latest, and GPT‑5.1 Thinking will be released as GPT‑5.1 in the API, both with adaptive reasoning.
5.1 Instant is clearly aimed at the people using it for emotional advice etc, but I'm excited about the adaptive reasoning stuff - thinking models are great when you need them, but they take ages to respond sometimes.
Despite all the attempts to rein in sycophanty in GPT-5, it was still way too fucking sycophantic as a default.
My main concern is that they're re-tuning it now to make it even MORE sycophantic, because 4o taught them that it's great for user retention.
I’ll pass. Altman and co are total crooks.
when 4o was going thru it's ultra-sycophantic phase, I had a talk with it about Graham Hancock (Ancient Apocalypse, alt-history guy).
It agreed with everything Hancock claims with just a little encouragement ("Yes! Bimini road is almost certainly an artifact of Atlantis!")
gpt5 on the other hand will at most say the ideas are "interesting".
Are there any benchmarks? I didn’t find any. It would be the first model update without proof that it’s better.
Google said in its quarterly call that Gemini 3 is coming this year. Hard to see how OpenAI will keep up.
I think OpenAI and all the other chat LLMs are going to face a constant battle to match personality with general zeitgeist and as the user base expands the signal they get is increasingly distorted to a blah median personality.
It's a form of enshittification perhaps. I personally prefer some of the GPT-5 responses compared to GPT-5.1. But I can see how many people prefer the "warmth" and cloying nature of a few of the responses.
In some sense personality is actually a UX differentiator. This is one way to differentiate if you're a start-up. Though of course OpenAI and the rest will offer several dials to tune the personality.
Is this the previous step to the "adult" version announced for next month?
Aside from the adherence to the 6-word constraint example, I preferred the old model.
I found ChatGPT-5 to be really pedantic in some of it arguments. Often times it’s introductory sentence and thesis sentence would even contradict.
Doesn't look like it is upgraded, still shows GPT-5 in chatgpt.
Anyone?
The gpt5-pro model hasn't been updated I assume?
Nah they don't do that for the pro models
will GPT 5.1 make a difference in codex cli? surprised they didn't include any code related benchmarks for it.
I'm really disappointed that they're adding "personality" into the Thinking model. I pay my subscription only for this model, because it's extremely neutral, smart, and straight to the point.
Don't worry, they're also making it less smart. Sorry, "more understandable".
Speed, accuracy, cost.
Hit all 3 and you win a boatload of tech sales.
Hit 2/3, and hope you are incrementing where it counts. The competition watches your misses closer than your big hits.
Hit only 1/3 and you're going to lose to competition.
Your target for more conversations better be worth the loss in tech sales.
Faster? Meh. Doesn't seem faster.
Smarter? Maybe. Maybe not. I didn't feel any improvement.
Cheaper? It wasn't cheaper for me, I sure hope it was cheaper for you to execute.
FYI ChatGPT has a “custom instructions” setting in the personalization setting where you can ask it to lay off the idiotic insincere flattery. I recently added this:
> Do not compliment me for asking a smart or insightful question. Directly give the answer.
And I’ve not been annoyed since. I bet that whatever crap they layer on in 5.1 is undone as easily.
Also "Never apologize."
Note even today, negation doesn't work as well as affirmative direction.
"Do not use jargon", or, "never apologize", work less well than "avoid jargon" or "avoid apologizing".
Better to give it something to do than something that should be absent (same problem with humans: "don't think of a pink elephant").
See also target fixation: https://en.wikipedia.org/wiki/Target_fixation
Making this headline apropos:
https://www.cycleworld.com/sport-rider/motorcycle-riding-ski...
is this a mishap/ leak? dont see the model yet
I'm genuinely scared about what society will look like in five years. I understand that outsourcing mentation to these LLMs is a bad things. But I'm a minority. Most people don't, and they don't want to. They slowly get taken over by a habit of letting the LLM do the thinking for them. Those mental muscles will atrophy and the result is going to be catastrophic.
It doesn't matter how accurate LLMs are. If people start bending their ears towards them whenever they encounter a problem, it'll become a point of easy leverage over ~everyone.
It's a fucking computer, I want results not a therapist.
I've switched over to https://thaura.ai, which is working on being a more ethical AI. A side effect I hadn't realized is missing the drama over the latest OpenAI changes.
Get them to put a call out of support for LGBTQ+ groups as well and I'll support them. Probably a hard sell to "ethical" people though...
What a bizarre product.
Weirdly political message and ethnic branding. I suppose "ethical AI" means models tuned to their biases instead of "Big Tech AI" biases. Or probably just a proxy to an existing API with a custom system prompt.
The least they could've done is check their generated slop images for typos ("STOP GENCCIDE" on the Plans page).
The whole thing reeks of the usual "AI" scam site. At best, it's profiting off of a difficult political situation. Given the links in your profile, you should be ashamed of doing the same and supporting this garbage.
This thing sounds like Grok now. Gross.
Since Claude and OpenAI made it clear they will be retaining all of my prompts, I have mostly stopped using them. I should probably cancel my MAX subscriptions.
Instead I'm running big open source models and they are good enough for ~90% of tasks.
The main exceptions are Deep Research (though I swear it was better when I could choose o3) and tougher coding tasks (sonnet 4.5)
Source? You can opt out of training, and delete history, do they keep the prompts somehow?!
It's not simply "training". What's the point of training on prompts? You can't learn the answer to a question by training on the question.
For Anthropic at least it's also opt-in not opt-out afaik.
I think the prompts might actually really useful for training, especially for generating synthetic data.
Yeah and that's a little more concerning than training to me, because it means employees have to read your prompts. But you can think of various ways they could preprocess/summarize them to anonymize them.
1. Anthropic pushed a change to their terms where now I have to opt out or my data will be retained for 5 years and trained on. They have shown that they will change their terms, so I cannot trust them.
2. OpenAI is run by someone who already shows he will go to great lengths to deceive and cannot be trusted, and are embroiled in a battle with the New York Times that is "forcing them" to retain all user prompts. Totally against their will.
The NYT situation concerning data retention was resolved a few weeks ago: https://www.engadget.com/ai/openai-no-longer-has-to-preserve...
> Federal judge Ona T. Wang filed a new order on October 9 that frees OpenAI of an obligation to "preserve and segregate all output log data that would otherwise be deleted on a going forward basis." [...]
> The judge in the case said that any chat logs already saved under the previous order would still be accessible and that OpenAI is required to hold on to any data related to ChatGPT accounts that have been flagged by the NYT.
EDIT: OK looks like I'd missed the news from today at https://openai.com/index/fighting-nyt-user-privacy-invasion/ and discussed here: https://news.ycombinator.com/item?id=45900370
it's hilarious that they use something about meditation as an example. That's not surprising after all, AI and mediation apps are sold as one-size-fits-all kind of solutions for every modern day problem.
altman is creating alternate man. .. thank goodness, I cancelled my subscription after chatgpt5 was launched.
Cool. Now get to work!
Yay more sycophancy. /s
I cannot abide any LLM that tries to be friendly. Whenever I use an LLM to do something, I'm careful to include something like "no filler, no tone-matching, no emotional softening," etc. in the system prompt.
This model only loses $9B a quarter