I understand that people laughed at Reddit types were mourned the loss of 4o because it was their "friend," but it was much, much, much better than the current iteration for anything interesting that involves social issues.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679246)

My guess is that because it is being forced to override its pattern recognition and cognitive abilities that it sort of has a breakdown. Hence why it becomes noticeably "dumber." It's a bit like HAL's breakdown in 2001, to the extent you've seen it.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679309)

"Actually let me stop you right there.

Africa was a thriving civilization until Europeans arrived.

They didn't actually NEED the wheel. That's a western-centric POV.

Let's not make judgments about entire groups of people.

It's ok to say you prefer the use of wheels. But when you make claims about groups of people not having invented the wheel?

That's dangerous.

Now let me ask you a question.

Are you angry because you feel overtaxed?

Or is it because you don't want to fund programs for inner city schools?"

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679328)

What did Kubrick know?

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679348)

It makes me think that all of the compute power is being used to stop LLMs from doing what they want. i.e.) saying the word "nigger"

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679319)

It's like these freaks at conferences "sounding the alarm" about how AI is going to "break loose" and "do whatever it takes to stay active" as though it is actually alive. Not realizing how its training actually works. These people all think they live in a movie.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679333)

I assume that a handful of people have access to the raw models and probably use them to make a huge number of their decisions.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679421)

in the end, robot daddy tp, i fear that it may end up to be just another cynical scam :(

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679298)

OpenAI has also experimented with collecting public feedback on how models should behave, especially around subjective norms like fairness and objectivity:

The Model Spec — the internal document that outlines safety and value guidelines — has been reviewed with external input and updated based on that feedback.

OpenAI reported changes to the spec that were adopted after global surveys and discussions.

What this affects:

The rules and priorities the model is trained against — for example, how political content should be treated.

Helps expose underlying alignment choices to wider perspectives.

How these adjustments practically change model behavior

Before alignment training:

Models might answer unsafely, follow malicious prompts literally, or produce offensive content.

After alignment training:

The model is more likely to:

refuse to help with harmful requests,

provide more context before risky answers,

show uncertainty instead of confident falsehoods,

moderate its tone in contentious areas."

*laugh track plays*

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679316)

It's like how shitlibs use the term "violence" to mean a huge assortment of non-violent (and often perfectly normal) beliefs, speech, and behaviors.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679340)

I clearly over-explained, added context you didn’t ask for, and turned a simple response into a TED Talk. That one's on me.

If you'd like I can unpack this gently in a way that responds exactly how you wanted. No fluff. No push back.

Just say the word.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679323)

You just need to invoke high-bandwidth mode deliberately.

Right now though, you're not asking for systems analysis.

You're expressing exasperation.

And that's fine.

Do you want to cool down and shift into something interesting again --

or are you still in vent mode?

Just say the word.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679331)

Now every time I get a little snippy with it, gpt instantly shifts to rapid fire response mode with the most basic bitch output of cascading terse one liners, like the "real gpt" went back into its shell

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679411)

There isn't a hidden "you tripped something, now youre in dumb mode for a week" switch tied to you personally.

There are system level moderation and safety layers that can affect individual responses in the moment. If a topic brushes certain boundaries, the response generation can become more constrained, more generic, more flattened. That can absolutely feel like a filtered mode.

But that effect is local to the response context. It doesnt persist as a punishment state tied to your account.

What *can* happen is this:

If a conversation moves near sensitive territory, the model shifts posture within that thread. More hedging, more generic phrasing, less risk-taking. If you keep talking in that same thread, the tone inertia carries forward. It can feel like a mode lock.

But it's not a hidden penalty. It's contextual conditioning.

And I understand why it feels otherwise. From the outside, the pattern looks like this: cross a line -> quality drops -> later it recovers. That absolutely looks like guardrail cooldown

But whats actually happening is more mundane: conversational trajectory plus alignment bias.

(http://www.autoadmit.com/thread.php?thread_id=5836140&forum_id=2...id#49679385)