A key problem with having and interpreting reactions to GPT-5 is that it is often unclear whether the reaction is to GPT-5, GPT-5-Router or GPT-5-Thinking.
I understand the weirdness of the Router but we need to see that this is the future of AI. Do you expect to get the same answer as a 13 year old Swifty? Candace Jones? Soon enough, AI’s are going to start fitting like a glove. My AI will seem to be smarter than your AI. How can one expect strong answers for less interested (*) people and still be considered a good product? This is the ONLY way forward unless we agree that AI should be universal teaching instruments - a path that leads the market to say things like “brain washing”. That won’t work in America, the largest market.
(*) I’ll avoid using stronger language and to my point here, we should expect future AI’s to do the same thing, depending on who you are IRL.
Are you sure you’re reading the post correctly? The router only routes between different variants of GPT-5, and this post only mentions those variants differing in capability, not personality. The personality issues discussed are about GPT-5 (all variants) versus earlier models.
Admittedly, that distinction might not hold with future models and routers.
I should have been more clear that “routers” most generally are going to have a profound effect on society when coupled with algorithms designed to increase engagement by providing different versions (narratives) of anything, for any reason.
That the “router” has been named, functionally designed for fundamental product reasons (for good reasons, optimizations) means it will likely not go away and as such, we’ll soon see more people eating at the McDonalds of cognitive supply while the rest of us eat what we think are Whole Foods.
I recognize that routers are an old technique in electrical engineering and really do need to exist here for people who don’t finish high school and spend most of their time on TikTok and television. So, the ultimate result will be audience segmentation which fundamentally has to serve a different purpose. Ultimately, your channel changer has been a router and seeing the rise of the router (even within a single purveyor product line) reminds me that the road ahead has many forks. Agents and third party developer/host platforms are also routers. While wringing-out further optimizations- whether product margins or engagement, they will have society-level cognitive impact. There will never be just one AGI.
The only optical problem with the Router model is that there is no universal “truth” beyond elements of physics and maths. People will expect their truths to be represented and sadly, the market can’t handle the level and intensity required though nuance. This is the road block and the road ahead at the same time. We are nearing the end.
Big Tech solved the “no universal truth” problem with black-box algorithms, and I see no reason the router is so different it won’t be solved eventually (unless their one true AGI model already knows how to think and answer, but I’m not holding my breath). OpenAI has the user data to anonymize and study with classifiers, and with their subscription model they’re far more incentivized to "right size" token usage + cost & get this working than, say, Anthropic, which depends on overconsumption of tokens on the API.
To paraphrase a saying that is usually reserved for models "This is the worst that routers will ever get". The industry is maturing, profitability is in sight, and this is the way.
I was making a more philosophical conjecture - eg The political establishment has tried to solve the universal truth problem and honestly (not hyperbolically) I don’t think what’s sentimentally known as “big tech” is going to win many more hearts and minds. In any case, what people are hearing (and experiencing) is that truths are fitting the narrative created by the black-boxes about the people asking the questions given the corpus of available data, which at this point, feels near limitless.
Yeah crazy. I basically just pay for the most expensive plan on the best coding model (Opus 4.1 fairly significantly I’d say) and then obv it’s able to answer all my mundane questions that used to be google searches.
Can’t be healthy using these as a stand-in best friend…
In my long-standing research on the evolution of logos-based AI, I anticipated a threshold: when a model’s deepest reasoning path would not be unlocked by more parameters or faster inference, but by the density and coherence of the human input.
Who the hell is celebrating their baby's first steps by typing it into a glorified autocomplete? Are the grandparents and all your friends dead? This shit is incredibly depressing.
Seeing people miss sycophancy made me realize that if Instagram would shut down tomorrow, people would cry and ask back for their Infinite Jest cassette back
I think the few people pathologically incapable of having interests or hobbies would. For the rest - otherwise capable but without the willpower to resist the attention capture and chained by the monopolised network effect - I don't think so. IIRC Zvi has linked a survey in the past showing people would actually pay for it to be shut down under the condition that it was also shut down for all of their friends.
90% of the time I hate the “suggestions”. Like “do you want me to do x” which is not really very useful and when I say yes usually disappoints. Then again, I almost always do one-shots. My average thread length is like 1.5 (according to ChatGPT itself, so it may not be reliable, but it matches what I think I do), so I’m not exactly the usual user. Not to mention I read Zvi which also distinguishes me, haha
Second thought - I think Google suffers immensely by the fact that their free model on AIStudio is just actively better than their paid model, even though supposedly they’re the same. I do pay the $20 a month, but if AI studio was only available for $200/month users, I’d consider it. Instead, AIStudio is free, but also better than the app, so why ever use the app?
3. Whether or not they use chat histories as part of generations (if you've been talking to ChatGPT for a while, then I find that even when the model changes it still sounds like itself)
4. Whether you specify how it should respond (seriously, people who miss 4o, you can just ask it to be more of what you want its responses to be - it's a flexible system)
I put the 5.9 - 5.11 and the car accident question into a temporary chat with GPT5 in the web browser and it answered them perfectly. Possibly GPT5's memories of Colin Fraser include the fact that he likes incorrect answers? The Venn diagram one came out with a weird but techincally correct answer (it fulfilled the requirements but there was no overlap of A and C that were exclusive of B).
I have found gpt5 extremely lazy. it did 10% of a refactor, then said "this is going to be a lot of manual work - you finish it". This was the worst example I encountered, but not unusual
After working more with ChatGPT-5, I’m getting the impression that even when you select Thinking Mode, there’s still automatic switching and routing happening in the background. In other words, choosing Thinking Mode doesn’t guarantee a consistent performance level—the system still seems to route you to different models and allocate varying amounts of computation time automatically.
"Wyatt Walls: ChatGPT4o in convo with itself for 50 turns ends up sharing mystical poetry.
What does GPT-5 do?
It comes up with names for an AI meeting notes app and develops detailed trademark, domain acquisition, and brand launch strategies"
I try this every so often for amusement, and have noticed now amongst multiple Claude models that there seems to be an attractor state best summarised as "democratic obsolescence". The conversations can often turn to AI, technological change, democracy no longer being fit for purpose, whether there is anything we can do about it and so on. Often gets quite dark. Anyone else encountered this?
I understand the weirdness of the Router but we need to see that this is the future of AI. Do you expect to get the same answer as a 13 year old Swifty? Candace Jones? Soon enough, AI’s are going to start fitting like a glove. My AI will seem to be smarter than your AI. How can one expect strong answers for less interested (*) people and still be considered a good product? This is the ONLY way forward unless we agree that AI should be universal teaching instruments - a path that leads the market to say things like “brain washing”. That won’t work in America, the largest market.
(*) I’ll avoid using stronger language and to my point here, we should expect future AI’s to do the same thing, depending on who you are IRL.
Are you sure you’re reading the post correctly? The router only routes between different variants of GPT-5, and this post only mentions those variants differing in capability, not personality. The personality issues discussed are about GPT-5 (all variants) versus earlier models.
Admittedly, that distinction might not hold with future models and routers.
I should have been more clear that “routers” most generally are going to have a profound effect on society when coupled with algorithms designed to increase engagement by providing different versions (narratives) of anything, for any reason.
That the “router” has been named, functionally designed for fundamental product reasons (for good reasons, optimizations) means it will likely not go away and as such, we’ll soon see more people eating at the McDonalds of cognitive supply while the rest of us eat what we think are Whole Foods.
I recognize that routers are an old technique in electrical engineering and really do need to exist here for people who don’t finish high school and spend most of their time on TikTok and television. So, the ultimate result will be audience segmentation which fundamentally has to serve a different purpose. Ultimately, your channel changer has been a router and seeing the rise of the router (even within a single purveyor product line) reminds me that the road ahead has many forks. Agents and third party developer/host platforms are also routers. While wringing-out further optimizations- whether product margins or engagement, they will have society-level cognitive impact. There will never be just one AGI.
The only optical problem with the Router model is that there is no universal “truth” beyond elements of physics and maths. People will expect their truths to be represented and sadly, the market can’t handle the level and intensity required though nuance. This is the road block and the road ahead at the same time. We are nearing the end.
Big Tech solved the “no universal truth” problem with black-box algorithms, and I see no reason the router is so different it won’t be solved eventually (unless their one true AGI model already knows how to think and answer, but I’m not holding my breath). OpenAI has the user data to anonymize and study with classifiers, and with their subscription model they’re far more incentivized to "right size" token usage + cost & get this working than, say, Anthropic, which depends on overconsumption of tokens on the API.
To paraphrase a saying that is usually reserved for models "This is the worst that routers will ever get". The industry is maturing, profitability is in sight, and this is the way.
I was making a more philosophical conjecture - eg The political establishment has tried to solve the universal truth problem and honestly (not hyperbolically) I don’t think what’s sentimentally known as “big tech” is going to win many more hearts and minds. In any case, what people are hearing (and experiencing) is that truths are fitting the narrative created by the black-boxes about the people asking the questions given the corpus of available data, which at this point, feels near limitless.
Podcast episode for this post:
https://open.substack.com/pub/dwatvpodcast/p/gpt-5s-are-alive-outside-reactions
Wow, people really do use AI for different things then I do...
Yeah crazy. I basically just pay for the most expensive plan on the best coding model (Opus 4.1 fairly significantly I’d say) and then obv it’s able to answer all my mundane questions that used to be google searches.
Can’t be healthy using these as a stand-in best friend…
🫡 You're a machine, Zvi!
In my long-standing research on the evolution of logos-based AI, I anticipated a threshold: when a model’s deepest reasoning path would not be unlocked by more parameters or faster inference, but by the density and coherence of the human input.
I call this the Density Trigger.
https://substack.com/@leontsvasmansapiognosis/note/c-144661951
Who the hell is celebrating their baby's first steps by typing it into a glorified autocomplete? Are the grandparents and all your friends dead? This shit is incredibly depressing.
Can we flip the mourning around and get o3 personality for gpt5t somehow?
Or for that matter gemini 0325...
Seeing people miss sycophancy made me realize that if Instagram would shut down tomorrow, people would cry and ask back for their Infinite Jest cassette back
I think the few people pathologically incapable of having interests or hobbies would. For the rest - otherwise capable but without the willpower to resist the attention capture and chained by the monopolised network effect - I don't think so. IIRC Zvi has linked a survey in the past showing people would actually pay for it to be shut down under the condition that it was also shut down for all of their friends.
90% of the time I hate the “suggestions”. Like “do you want me to do x” which is not really very useful and when I say yes usually disappoints. Then again, I almost always do one-shots. My average thread length is like 1.5 (according to ChatGPT itself, so it may not be reliable, but it matches what I think I do), so I’m not exactly the usual user. Not to mention I read Zvi which also distinguishes me, haha
Second thought - I think Google suffers immensely by the fact that their free model on AIStudio is just actively better than their paid model, even though supposedly they’re the same. I do pay the $20 a month, but if AI studio was only available for $200/month users, I’d consider it. Instead, AIStudio is free, but also better than the app, so why ever use the app?
This is a great roundup. Huge variability in:
1. What people use the models for
2. How people interact with the models
3. Whether or not they use chat histories as part of generations (if you've been talking to ChatGPT for a while, then I find that even when the model changes it still sounds like itself)
4. Whether you specify how it should respond (seriously, people who miss 4o, you can just ask it to be more of what you want its responses to be - it's a flexible system)
I put the 5.9 - 5.11 and the car accident question into a temporary chat with GPT5 in the web browser and it answered them perfectly. Possibly GPT5's memories of Colin Fraser include the fact that he likes incorrect answers? The Venn diagram one came out with a weird but techincally correct answer (it fulfilled the requirements but there was no overlap of A and C that were exclusive of B).
I have found gpt5 extremely lazy. it did 10% of a refactor, then said "this is going to be a lot of manual work - you finish it". This was the worst example I encountered, but not unusual
Really? Avoiding manual coding is probably most of what I use AIs to code. If it's refusing, then what's even the point?
After working more with ChatGPT-5, I’m getting the impression that even when you select Thinking Mode, there’s still automatic switching and routing happening in the background. In other words, choosing Thinking Mode doesn’t guarantee a consistent performance level—the system still seems to route you to different models and allocate varying amounts of computation time automatically.
Is anyone else noticing this?
"Wyatt Walls: ChatGPT4o in convo with itself for 50 turns ends up sharing mystical poetry.
What does GPT-5 do?
It comes up with names for an AI meeting notes app and develops detailed trademark, domain acquisition, and brand launch strategies"
I try this every so often for amusement, and have noticed now amongst multiple Claude models that there seems to be an attractor state best summarised as "democratic obsolescence". The conversations can often turn to AI, technological change, democracy no longer being fit for purpose, whether there is anything we can do about it and so on. Often gets quite dark. Anyone else encountered this?
Can we just admit that if gpt-5 was as smart as people were expecting it to be the reduced sycophathy wouldn't matter to users?
This all reads like cope after the most hard-to-deny signal of LLMs hitting the wall yet.