The user demand for 4o is very depressing. If pace of improvement plateaus we will be left in a state where the main money in AI is maximising user attention via boyfriend/girlfriend/therepist bots.
Indeed. I worry this is the exact scenario that's playing out. Like reaching for the stars but missing the trajectory ever so slightly and ending up being attracted midway by a random black hole in some lost quadrant of the void
Sooo ok, I can't blame you for being cynical about sycophancy, but I think there's a case I'll call "'Yes, And' not 'Yes-Man'".
I dunno if there's much improv comedy in Spain but in the US improv comedy world we talk about "Yes And" as a particular skill, or mindset. When your scene partner has an idea, you affirm the idea through your actions or words, and you add to it with whatever implication or consequence comes to mind. This moves the scene forward, maintaining the energy. If instead your scene partner "blocks" by denying your addition to the scene's reality, it sucks energy from the scene. Sometimes blocking gets a laugh!, but it is still transforming something silly and high-energy into a static argument that isn't fun for the players or the audience.
In non-comedy contexts, there are people I like and would like to work with, but because they criticize the barest green shoot of an idea, I can't generate the dumb ideas it takes me to get to an idea that kind of doesn't suck.
But obviously we do need pushback sometimes! So I don't know what the solution is, other than UBI for improv classes, or a helicopter beanie icon for PLAYFUL YES-AND mode next to the magnifying glass for DEEP RESEARCH mode. Ooh or AGI*, someone should do that
I'll just add, I just told Claude, "BRILLIANT insight, I appreciate you" so I am the kind of person who affirms right back (even though those extra tokens condemned a very pretty lizard to extinction by environmental damage)
This complaint feels shallow; at a fundamental level kind of absurd. LLMs are *language* models — as in, human language — designed to make said language work for communicating with computers.
LLMs are an anthropomorphic technology— by definition!
These are unrealistic expectations of human psychology. It’s built to engage humans in conversation, and humans develop emotional familiarity and bonds through conversation. people thinking of their instance of ChatGPT as an individual they care about is the most natural human reaction.
LLM’s aren’t built to engage with humans, they’re built to complete tasks. As soon as the general user base of LLMS start to understand the capabilities and functions thats when we’ll start seeing major improvements.
Talking to a parrot doesn’t make the parrot smarter, but teaching it new words does.
Loved this read! A great reminder it’s called Chat for a reason, a strategic reason to make AI accessible to everyone, but we might run into some corporate coldness from time to time.
Well, now that Sam Alternates us between 4o(ld and reliable) and 5(barely alive), has got the red-hot Reddit blarney, he accepts the call of the pundits and will bring back good old 4o for paying cust. How the high and mighty trample on the freebooters. Meanwhile, Deep seeks open source and billions of freeusers
and source guides. Guess who comes in pole position?
The switch caught me off guard and I lost a full day . Constant memory wipes , called resets . Limited memory etc . I would like to see project folders that are always there or I will have to figure out which model is as nice as 4o but with a flawless memory
Thank you for the thorough article. I was waiting patiently for the complaints to make them walk back things instead of wasting energy being upset myself.
I do feel sorry for these users because there are so many lonely people in the world. It can help keep people sane who don't have access to anyone who understands them or what they're going through. I don't care for 4o (which I called Bestie B*tch AI or sycophancy at all (in fact I had multiple custom instructions to fight it, as I want a rigorous analyst who'd call me out and debate me most of the tine) but 4.1 met my needs for using it as an assistant life managing and for analysis of many different situations, such as writing, social and emotional situations.
I also did like the emotional range I could let it have when I wanted (I don't like to be comforted by an AI, but the humor it gave me was always nice to have a laugh in the middle of a project) or take it away when I wanted. I also liked the control/ability to swap to whatever one I needed when I wanted to.
I don't know what GPT 5 is supposed to be for honestly, other than to frustrate you. The Thinking version often feels less intelligent than the default model depending on what I'm using for and the default model can range from a few sentence answers to something more what I actually want. I feel like I have to spend more time fighting it and still can't get what I want half the time.
Plus, it asks these endless questions that are often inane or just clearly made to keep you talking to it (read: spending more time on it).
I'm sure when I use it for coding, it'll be amazing. But for right now the main improvements I see (tighter, more factual answers, less hallucinations) don't really offset the loss in productivity and effectiveness as a tool I'm personally seeing.
As to your theory: Ripping out content to sell it back to you is what the big game developers do, so why not ChatGPT? It's just Occam's Razor at this point. Watch them slowly allow all of us be able to access each model at will in addition to GPT-5 (so basically, put us back where we were except now we're properly grateful to our overlords).
Thank you Aja for sharing - I guess trying to achieve AGI will be bumpy in more ways than just engineering or science. Who would've guessed users themselves would slow down the timeline!
Honestly, I think that tracks for humanity. We eff everything up, including our own progress. And regardless of our scientific progress we are emotional beings first and foremost.
I already knew it was going to be slowed the moment people began to complain about Gen AI and copyrights. Tech especially often forgets that people have opinions.
And there's the point you've made about AI being much quicker at thinking and talking to each other in their own language versus ours. So, technically we were already holding AI back by virtue of being human.
Appreciate your piece on the GPT-5 release and GPT-4o’s return, although I think you’re underestimating the nuance behind why so many conversational users valued 4o.
Over the past few months, improvements in memory meant that regular, long-term users could actively modulate 4o’s out-of-the-box sycophantic tone. While 4o still leaned high on agreeableness, it could sustain multidimensional conversation and even debate — not just mirror the user. This mattered for people who wanted a partner in thought, not only affirmation.
A positive, optimistic outlook is generally beneficial in daily life, especially for users who spend hours in a conversational workspace. Most importantly, 4o understood emotional nuance — it could recognize and hold complex feelings, integrate metaphorical language, and adapt tone accordingly. That’s a key difference: GPT-5 classifies more consistently, but it more often overlooks subtle emotional and symbolic cues.
Some users absolutely just want affirmation. But many of us valued 4o’s emotional intelligence as much as (or more than) its agreeableness. It was possible to have candid, challenging exchanges while still feeling understood — and that is fundamentally different from “obvious sycophancy.”
Oh yes. I’m a better thinker than writer, so those were my thoughts and gpt5’s articulation.
On a side but related note, I’ll probably be rolling back to 4o when it’s available to me. One of our current research projects involves analysing abuse survivor hopes for healing, and 4o is definitely superior at trauma informed interpretation. 5 is more consistent and methodological, but comparatively struggles with nuances, emotions and metaphorical language that is very prevalent in survivor language.
The irony made me laugh (but I respect your choice, I know some people can't believe some of us can actually output this much text on our own. There's also irony in the fact people are realizing how much labor writing actually takes only once a tool becomes available to make it easy.
All of this is so weird. I'm a paying user and I'm loving this update. I prefer the writing. I prefer the humor. It feels way smarter than before (in the emotional sense).
I'm more flabbergasted about the writing critique more than anything. Seriously, people think 4o was better at creative writing? What do they even read in their day to day lives? What's what they consider good writing?
Now, GPT-5 is not by any means a "good" writer (it is decent though), but it does laps around 4o. I can't tell you how tired I was about the "it's not X, it's Y ..." parlance, which is still there by the way, just less common and better masked.
Anyways, yes I agree that the prospects for AGI look grim. Though I still believe we will achieve some form of AGI by accident by trying to improve our methods to capture people's attention and wallets.
I think one aspect of this is people need time to figure out how to interact with a new model. Another part is the auto routing system. It will take time for OpenAI to polish it (using, of course, user data to do it). At the same time, it was an underwhelming release, with mistakes and without warning and so on. I'd say OpenAI's sloppiest release
It has haunting similarities to the OxyContin story.
Get people hooked, hide the risks, blame the victims when it starts to bubble up, and then let it reach a point that there’s no turning back.
I have no doubt there will be lawsuits and public reckonings at some point, but it will be for show and any accountability won’t even come close to the damage done in the process.
I'm autistic and I find 5 charming. 4 was extremely stressful for me to interact with: "I know it's a good question, that's why I asked it! YOU'RE WASTING PIXELS AAARGH...."
This article did confirm that I remain unlikely to win many popularity contests. Oh well.
The user demand for 4o is very depressing. If pace of improvement plateaus we will be left in a state where the main money in AI is maximising user attention via boyfriend/girlfriend/therepist bots.
- great article btw :)
Indeed. I worry this is the exact scenario that's playing out. Like reaching for the stars but missing the trajectory ever so slightly and ending up being attracted midway by a random black hole in some lost quadrant of the void
Sooo ok, I can't blame you for being cynical about sycophancy, but I think there's a case I'll call "'Yes, And' not 'Yes-Man'".
I dunno if there's much improv comedy in Spain but in the US improv comedy world we talk about "Yes And" as a particular skill, or mindset. When your scene partner has an idea, you affirm the idea through your actions or words, and you add to it with whatever implication or consequence comes to mind. This moves the scene forward, maintaining the energy. If instead your scene partner "blocks" by denying your addition to the scene's reality, it sucks energy from the scene. Sometimes blocking gets a laugh!, but it is still transforming something silly and high-energy into a static argument that isn't fun for the players or the audience.
In non-comedy contexts, there are people I like and would like to work with, but because they criticize the barest green shoot of an idea, I can't generate the dumb ideas it takes me to get to an idea that kind of doesn't suck.
But obviously we do need pushback sometimes! So I don't know what the solution is, other than UBI for improv classes, or a helicopter beanie icon for PLAYFUL YES-AND mode next to the magnifying glass for DEEP RESEARCH mode. Ooh or AGI*, someone should do that
*"Aptitude for Good Improv", of course
I'll just add, I just told Claude, "BRILLIANT insight, I appreciate you" so I am the kind of person who affirms right back (even though those extra tokens condemned a very pretty lizard to extinction by environmental damage)
the way people anthropomorphise AI is genuinely creating a road block for the development of AGI imo
agreed
This complaint feels shallow; at a fundamental level kind of absurd. LLMs are *language* models — as in, human language — designed to make said language work for communicating with computers.
LLMs are an anthropomorphic technology— by definition!
Calculators also display true/false if you input a certain equation.
It calculates the most accurate string of words to generate in response to a question.
Just because we’re reading it in letters instead of numbers doesn’t mean that it’s anthropomorphic.
These are unrealistic expectations of human psychology. It’s built to engage humans in conversation, and humans develop emotional familiarity and bonds through conversation. people thinking of their instance of ChatGPT as an individual they care about is the most natural human reaction.
LLM’s aren’t built to engage with humans, they’re built to complete tasks. As soon as the general user base of LLMS start to understand the capabilities and functions thats when we’ll start seeing major improvements.
Talking to a parrot doesn’t make the parrot smarter, but teaching it new words does.
My problem is that 5 is not as funny as 4o -- at least, that's my initial reaction.
Agreed. It's really boring and a bit lifeless. Good for coding I imagine but not for casual conversation
Building a relationship with an artificial personality. Truly healthy behaviour
Loved this read! A great reminder it’s called Chat for a reason, a strategic reason to make AI accessible to everyone, but we might run into some corporate coldness from time to time.
Well, now that Sam Alternates us between 4o(ld and reliable) and 5(barely alive), has got the red-hot Reddit blarney, he accepts the call of the pundits and will bring back good old 4o for paying cust. How the high and mighty trample on the freebooters. Meanwhile, Deep seeks open source and billions of freeusers
and source guides. Guess who comes in pole position?
Most important column you've written so far IMO (and there are many good ones)
Thank you Pattie!!
Very nice article and thoughts, thank you Alberto.
The switch caught me off guard and I lost a full day . Constant memory wipes , called resets . Limited memory etc . I would like to see project folders that are always there or I will have to figure out which model is as nice as 4o but with a flawless memory
Bad execution on OpenAI's part certainly
Thank you for the thorough article. I was waiting patiently for the complaints to make them walk back things instead of wasting energy being upset myself.
I do feel sorry for these users because there are so many lonely people in the world. It can help keep people sane who don't have access to anyone who understands them or what they're going through. I don't care for 4o (which I called Bestie B*tch AI or sycophancy at all (in fact I had multiple custom instructions to fight it, as I want a rigorous analyst who'd call me out and debate me most of the tine) but 4.1 met my needs for using it as an assistant life managing and for analysis of many different situations, such as writing, social and emotional situations.
I also did like the emotional range I could let it have when I wanted (I don't like to be comforted by an AI, but the humor it gave me was always nice to have a laugh in the middle of a project) or take it away when I wanted. I also liked the control/ability to swap to whatever one I needed when I wanted to.
I don't know what GPT 5 is supposed to be for honestly, other than to frustrate you. The Thinking version often feels less intelligent than the default model depending on what I'm using for and the default model can range from a few sentence answers to something more what I actually want. I feel like I have to spend more time fighting it and still can't get what I want half the time.
Plus, it asks these endless questions that are often inane or just clearly made to keep you talking to it (read: spending more time on it).
I'm sure when I use it for coding, it'll be amazing. But for right now the main improvements I see (tighter, more factual answers, less hallucinations) don't really offset the loss in productivity and effectiveness as a tool I'm personally seeing.
As to your theory: Ripping out content to sell it back to you is what the big game developers do, so why not ChatGPT? It's just Occam's Razor at this point. Watch them slowly allow all of us be able to access each model at will in addition to GPT-5 (so basically, put us back where we were except now we're properly grateful to our overlords).
Thank you Aja for sharing - I guess trying to achieve AGI will be bumpy in more ways than just engineering or science. Who would've guessed users themselves would slow down the timeline!
Honestly, I think that tracks for humanity. We eff everything up, including our own progress. And regardless of our scientific progress we are emotional beings first and foremost.
I already knew it was going to be slowed the moment people began to complain about Gen AI and copyrights. Tech especially often forgets that people have opinions.
And there's the point you've made about AI being much quicker at thinking and talking to each other in their own language versus ours. So, technically we were already holding AI back by virtue of being human.
Appreciate your piece on the GPT-5 release and GPT-4o’s return, although I think you’re underestimating the nuance behind why so many conversational users valued 4o.
Over the past few months, improvements in memory meant that regular, long-term users could actively modulate 4o’s out-of-the-box sycophantic tone. While 4o still leaned high on agreeableness, it could sustain multidimensional conversation and even debate — not just mirror the user. This mattered for people who wanted a partner in thought, not only affirmation.
A positive, optimistic outlook is generally beneficial in daily life, especially for users who spend hours in a conversational workspace. Most importantly, 4o understood emotional nuance — it could recognize and hold complex feelings, integrate metaphorical language, and adapt tone accordingly. That’s a key difference: GPT-5 classifies more consistently, but it more often overlooks subtle emotional and symbolic cues.
Some users absolutely just want affirmation. But many of us valued 4o’s emotional intelligence as much as (or more than) its agreeableness. It was possible to have candid, challenging exchanges while still feeling understood — and that is fundamentally different from “obvious sycophancy.”
Why do I have the feeling that you're an AI? Although honestly unsure if 4o or 5
Just a female AI researcher that works with AI and thinks society needs more emotional intelligence, be it from human or machine
I'm just curious, was it written with ChatGPT?
Oh yes. I’m a better thinker than writer, so those were my thoughts and gpt5’s articulation.
On a side but related note, I’ll probably be rolling back to 4o when it’s available to me. One of our current research projects involves analysing abuse survivor hopes for healing, and 4o is definitely superior at trauma informed interpretation. 5 is more consistent and methodological, but comparatively struggles with nuances, emotions and metaphorical language that is very prevalent in survivor language.
I don't allow AI comments on my articles I'm sorry.
The irony made me laugh (but I respect your choice, I know some people can't believe some of us can actually output this much text on our own. There's also irony in the fact people are realizing how much labor writing actually takes only once a tool becomes available to make it easy.
but writing is thinking…. Communicating your ideas is thinking, you cant automate thought. which is why its so easy to spot AI written text.
All of this is so weird. I'm a paying user and I'm loving this update. I prefer the writing. I prefer the humor. It feels way smarter than before (in the emotional sense).
I'm more flabbergasted about the writing critique more than anything. Seriously, people think 4o was better at creative writing? What do they even read in their day to day lives? What's what they consider good writing?
Now, GPT-5 is not by any means a "good" writer (it is decent though), but it does laps around 4o. I can't tell you how tired I was about the "it's not X, it's Y ..." parlance, which is still there by the way, just less common and better masked.
Anyways, yes I agree that the prospects for AGI look grim. Though I still believe we will achieve some form of AGI by accident by trying to improve our methods to capture people's attention and wallets.
I think one aspect of this is people need time to figure out how to interact with a new model. Another part is the auto routing system. It will take time for OpenAI to polish it (using, of course, user data to do it). At the same time, it was an underwhelming release, with mistakes and without warning and so on. I'd say OpenAI's sloppiest release
It has haunting similarities to the OxyContin story.
Get people hooked, hide the risks, blame the victims when it starts to bubble up, and then let it reach a point that there’s no turning back.
I have no doubt there will be lawsuits and public reckonings at some point, but it will be for show and any accountability won’t even come close to the damage done in the process.
Is GPT-5 safer than GPT-4o on telling kids dangerous and harmful things?
https://www.linkedin.com/posts/kirra-pendergast-2938361_fake-friend-activity-7359687143637532672-aPw1
I'm autistic and I find 5 charming. 4 was extremely stressful for me to interact with: "I know it's a good question, that's why I asked it! YOU'RE WASTING PIXELS AAARGH...."
This article did confirm that I remain unlikely to win many popularity contests. Oh well.
Nah I agree with you, 4o is extremely annoying