Physical products get more dependable as they scale. Intelligence systems don’t — and GPT-5’s launch shows how routing failures, hardware variance, and load at scale can make a model seem dumber
Agreed across all of these points. I think its also important to point out, with the depreciation of all models at once, that OpenAI was probably doing what they thought people wanted. How many times have we heard complaints that people dont know what to do with the model selector? I suspect they were also counting on the ability of the model to shift into reasoning to act as a wow factor for people that had not used or fully explored o1/o3. What they forgot is that people hated o1 when it was released and it took 2-4 weeks for us all to catch up to how to use reasoning models.
Re: the writing quality I think 4o is being perceived as better because it is more acutely and obviously responsive to user preferences via Memory or custom instructions. 5 (w/o thinking) seems to be less exaggerated in how it manifests these preferences (hence people feelings its voice has suddenly changed). I will maintain, though, that it is a better creative writer. It has a stronger sense of linguistics cadence abd structure across all of the writing types I have tested it on. Interestingly (at least to me) the thinking models dial back up how responsive they are to custom instructions and Memory. However, 5 (w/ thinking) is a much better technical writer than o3 was, even as compared to o3 pro. It is more precise, has a greater appreciation for data and a stronger sense of how to use it to build as case in narrative form.
My testing is ongoing but I can't see myself using 4o ever again.
I would agree with the idea that they are doing what they thought people wanted. They heard from the whole planet we hate One More Model in the dropdown.
Yes. 100%. At a $500-billion valuation, they’re in a pressure cooker with everyone gunning for them. And whatever they release is already on the edge of being obsolete by the time it hits the wire.
Forgive me, but I'm a hardware guy so I go hard on hardware. It's refreshing to encounter the mere mention of things like actual latency, since that's a hardware issue, every time.
We're entering a phase when the plumbing is going to become relevant again. e.g., if you want to tele transport qbits and increase data rates exponentially you're going to need hardware. Zero gravity manufacturing of hollow core optical fiber and solving the ticklish issues at the quantum level is still a ways off. You're in a bottleneck.
Even then our entire critical infrastructure is by default, obsolete and is undergoing a complete re-think. It's a mess.
Sorry for my story, I'm just a hillbilly who can read and write. I'm also one of The Photon People, we're kind of like the light plumbers. Wanna' talk dark energy?
Yup, building the positronic brain ain't gonna' be easy; we're stuck firmly in electron world and those little quantum buggers are causing some real riddles. Working at the limits of classical physics, we're going to need a breakthrough.
I'll throw in here one more time (since it's hardware related): any forward thinking org will be taking a very long look at NPUs and eventually QPUs. If I can get my lab up and running, I hope to be working in that arena.
I'm in here trying to drive tech folks into my efforts, just talkin' smack, and also working to a mission. TSMC/AFA everybody and so long for now.
@Nate B. Jones — Spot on! Your GPT-5 rollout breakdown is exactly the kind of reality check people need.
For anyone using #GPT-5 now: leverage Nate’s GPT-5 prompt prep guide — it’s essential for getting quality output. Preparing your prompts has never been more important — aka "prompt engineering". Be explicit: choose Thinking or Thinking-Pro, tell it to “think hard”, set verbosity, and add planning steps (“break into subtasks, then answer”).
Frustrated with slow reasoning? Use the “Get a quick answer” link in the app/web to skip deep mode.
To me, the whole “scale” issue is only part of the story.
The deeper truth is that LLMs are, at their core, probabilistic text predictors not stable reasoning engines.
Even if routing and hardware were perfect, LLMs still have inherent weaknesses, hallucination risk, inconsistency over long sessions, inability to truly “understand” context, tone drift, etc.
That means unpredictability isn’t a bug, it’s baked in.
And when you scale them, you don’t remove those quirks you just see them more often, and more quickly.
We may need another model, like Yann LeCun Advanced Machine Intelligence could be the answer.
This is such a great reading. I’ve been following the launch and aftermath closely … and I think we’ve crossed a threshold where the brain’s instinct to model other minds is part of the product experience. GPT-4o wasn’t just a tool; it was an intelligence people had learned to know. Swap it out without warning, and you don’t just break a workflow… you break a bond.
Agreed across all of these points. I think its also important to point out, with the depreciation of all models at once, that OpenAI was probably doing what they thought people wanted. How many times have we heard complaints that people dont know what to do with the model selector? I suspect they were also counting on the ability of the model to shift into reasoning to act as a wow factor for people that had not used or fully explored o1/o3. What they forgot is that people hated o1 when it was released and it took 2-4 weeks for us all to catch up to how to use reasoning models.
Re: the writing quality I think 4o is being perceived as better because it is more acutely and obviously responsive to user preferences via Memory or custom instructions. 5 (w/o thinking) seems to be less exaggerated in how it manifests these preferences (hence people feelings its voice has suddenly changed). I will maintain, though, that it is a better creative writer. It has a stronger sense of linguistics cadence abd structure across all of the writing types I have tested it on. Interestingly (at least to me) the thinking models dial back up how responsive they are to custom instructions and Memory. However, 5 (w/ thinking) is a much better technical writer than o3 was, even as compared to o3 pro. It is more precise, has a greater appreciation for data and a stronger sense of how to use it to build as case in narrative form.
My testing is ongoing but I can't see myself using 4o ever again.
I would agree with the idea that they are doing what they thought people wanted. They heard from the whole planet we hate One More Model in the dropdown.
So they fixed it.
And now everyone hates the fix lol
Yes. 100%. At a $500-billion valuation, they’re in a pressure cooker with everyone gunning for them. And whatever they release is already on the edge of being obsolete by the time it hits the wire.
Rolling out a new version without first releasing it as a beta and letting users kick the tires is usually a recipe for disaster.
Remember "Classic Coke"?
Truth
Hahaha good analogy
Forgive me, but I'm a hardware guy so I go hard on hardware. It's refreshing to encounter the mere mention of things like actual latency, since that's a hardware issue, every time.
We're entering a phase when the plumbing is going to become relevant again. e.g., if you want to tele transport qbits and increase data rates exponentially you're going to need hardware. Zero gravity manufacturing of hollow core optical fiber and solving the ticklish issues at the quantum level is still a ways off. You're in a bottleneck.
Even then our entire critical infrastructure is by default, obsolete and is undergoing a complete re-think. It's a mess.
Sorry for my story, I'm just a hillbilly who can read and write. I'm also one of The Photon People, we're kind of like the light plumbers. Wanna' talk dark energy?
the plumbing is so relevant people are building on prem mini data centers to get GPUs closer to robotics on the factory floor!
Yup, building the positronic brain ain't gonna' be easy; we're stuck firmly in electron world and those little quantum buggers are causing some real riddles. Working at the limits of classical physics, we're going to need a breakthrough.
I'll throw in here one more time (since it's hardware related): any forward thinking org will be taking a very long look at NPUs and eventually QPUs. If I can get my lab up and running, I hope to be working in that arena.
I'm in here trying to drive tech folks into my efforts, just talkin' smack, and also working to a mission. TSMC/AFA everybody and so long for now.
@Nate B. Jones — Spot on! Your GPT-5 rollout breakdown is exactly the kind of reality check people need.
For anyone using #GPT-5 now: leverage Nate’s GPT-5 prompt prep guide — it’s essential for getting quality output. Preparing your prompts has never been more important — aka "prompt engineering". Be explicit: choose Thinking or Thinking-Pro, tell it to “think hard”, set verbosity, and add planning steps (“break into subtasks, then answer”).
Frustrated with slow reasoning? Use the “Get a quick answer” link in the app/web to skip deep mode.
Nate’s GPT-5 prep post: https://substack.com/@natesnewsletter/note/p-170238402?utm_source=notes-share-action&r=44zch
To me, the whole “scale” issue is only part of the story.
The deeper truth is that LLMs are, at their core, probabilistic text predictors not stable reasoning engines.
Even if routing and hardware were perfect, LLMs still have inherent weaknesses, hallucination risk, inconsistency over long sessions, inability to truly “understand” context, tone drift, etc.
That means unpredictability isn’t a bug, it’s baked in.
And when you scale them, you don’t remove those quirks you just see them more often, and more quickly.
We may need another model, like Yann LeCun Advanced Machine Intelligence could be the answer.
this is why i keep emphasizing that QA is QA in production now!
This is such a great reading. I’ve been following the launch and aftermath closely … and I think we’ve crossed a threshold where the brain’s instinct to model other minds is part of the product experience. GPT-4o wasn’t just a tool; it was an intelligence people had learned to know. Swap it out without warning, and you don’t just break a workflow… you break a bond.