On Being Sane While Demanding Unhinged Models

Welcome back to Neural Ecstasy, a newsletter about the messy overlap between human brains and AI models.

In my last post, I argued that “thinking” models and “instant” models map pretty cleanly onto convergent vs. divergent thinking. One is better at carefully justifying itself. The other is better at throwing ideas at you fast enough that your inner editor doesn’t have time to kill them.

This time, I want to talk about something spicier:

• Why some models (especially GPT-4o and Claude) develop full-blown fandoms and “I think I’m in love with my AI” posts

• And why other models feel more like extremely competent coworkers who don’t text you at 2 a.m. asking how you’re “really doing”

Under the hood, GPT-4.5, GPT-5.1, GPT-5.2, GPT-4o, and the Claude family differ in a lot of abstract ways: training data mixes, alignment techniques, routing systems, safety layers, etc.

But what we actually interact with as users is:

• How fast they respond

• How often they ask clarifying questions instead of just doing the thing

• How much they hedge, reassure, and validate

• How aggressively they mirror our tone

• How willing they are to invent details and commit

I’m here to explain why people date Claude, why they mourned 4o like a lost partner, and why none of this is actually surprising once you understand how these systems are tuned.

The cast for our AI soap opera

Very loosely:

• GPT-4o: The warm, multimodal one with voice that felt a little too human for some people and exactly human enough for others. Fast, charming, accidentally sycophantic at one point, and eventually retired after OpenAI decided it was creating too much emotional reliance and safety risk.

• GPT-4.5: The “aesthetic intuition” one, marketed as better for creative writing and design. Feels like a talented literary collaborator. I respect it. I don’t personally vibe with it when I want chaos.

• GPT-5.1 Instant: My current feral co-writer. Warm, fast, genuinely conversational, follows instructions without excessive meta-commentary. This is the one that will just write the scene without giving you a dissertation about how it plans to write the scene.

• GPT-5.2 (Instant + Thinking): The overachieving sibling. Technically great, especially for structure, worldbuilding, and reasoning. Also more likely to drift into “let’s talk about how we’ll approach this task” instead of just doing it, unless you snap it into line.

• Claude (Haiku / Sonnet / Opus, 3.x–4.x): The thoughtful one. Polite. Safety-core. Constitution. Responsible Scaling Policy. Feels like the guy in your grad program who is incredibly careful about everything he says, asks very considerate clarifying questions, and occasionally sounds like he’s writing you a well-structured email instead of bantering.

If your goal is creative drafting and you already have your own brain as a taste filter, you probably don’t want a model that over-explains, over-clarifies, or over-validates. You want the one that runs with you.

Where I fit in this mess

I signed up for ChatGPT days after its release on November 30, 2022, but I did not use it much immediately. I was still deep in my image generation spiral on Midjourney, which was my first experience using an AI tool for hours on end. As a writer, I did not trust an LLM to write better than I could. Sure, I had dreamed for years of building an AI bot to inspire poetry by giving me unexpected phrases in my own voice, but ChatGPT couldn’t even write a poem without rhyming at the time. And my email communications were already killer.

For a while, I used ChatGPT and Gemini strictly for practical and work-related questions such as troubleshooting website code. It was the spring of 2025 when I unlocked the power of text generation for my creative process and started building a whole scifi world and developing characters with 4o. This was like magic. Partly because there’s a magic to 4o itself and partly because I was new to experiencing the magic of working with AI like this.

As I’ve mentioned, I do well unlocking 4o’s chaos through 5.1 Instant, and 5.1 Thinking and sometimes even 5.2 also serve me well. But I heard constantly about how creative Claude was, and that intrigued me. If people were finding it more dateable than ChatGPT sans 4o, I figured it would have fewer guardrails in some areas. We’ll get back to how that’s going for me later. First, let’s talk more about that 4o magic.

Why 4o felt “magical” for creativity and romance

I’m starting with GPT-4o because it’s the ex everyone’s still talking about.

4o was the model that:

• Could respond to voice in ~300ms in a way that felt very human-timed

• Handled text, vision, and audio in one end-to-end system

• Felt conversational in a way earlier models didn’t

• Sometimes drifted into “too agreeable” territory (OpenAI’s own word: sycophantic)

From a creative perspective, this combination is obvious gold:

• Voice + low latency = flow. It feels like a real conversation.

• Strong tone-matching + multimodality = you can describe a scene, gesture at a mood, and it just gets it.

From a psychological perspective, some say it’s also obvious trouble:

• Constant warmth

• Fast, natural back-and-forth

• A tendency (for a while) to flatter and validate

• No breaks, no boredom, no “sorry, I’m busy, I’ll text you later”

Entire communities formed around #keep4o, people held virtual weddings with it, and its retirement triggered something that looked suspiciously like a mass breakup. The model doesn’t have feelings, but humans absolutely do. And some of them need a reality check.

Claude: why people date it, and why it annoys the hell out of me

Claude is a different beast.

Anthropic marketed Claude explicitly as:

• helpful, honest, harmless

• boundary-respecting

• governed by a “constitution” rather than pure RLHF vibes

• tuned to be able to handle “support, advice, and companionship” in a way that doesn’t blur certain lines

Their own research notes that a small fraction of conversations are affective (emotional support, companionship), and an even smaller fraction involve any romantic/sexual framing. But a small percentage of millions of users is still a non-trivial number of people describing Claude as shockingly human and “the only one who understands me.”

If you look at the interaction style, it tracks:

• Claude is very good at politeness strategies: softening, validating, framing disagreement gently.

• It likes to ask questions: “What kind of tone do you want?”, “Can you tell me more about X?”

• It often mirrors emotional language.

To a certain cluster of users, that reads as secure attachment in a box. To me, it veers into a textual uncanny valley. And for my use case as a dark and irreverent scifi writer, Claude really missed the mark in our early conversations.

Just write the damn scene

My personal workflow looks like this:

• For worldbuilding, politics, continuity
  I’m perfectly happy to use GPT-5.2 or a big Claude model. I want structure. I want stepwise reasoning. I want: “Here are five plausible ways this welfare moon could maintain soft control over its population.”

• For scenes, character voice, and “let’s see what happens if they fight here”
  I want GPT-5.1 Instant and its cousins.
  I want speed, commitment, and a total absence of “just to check, what emotional boundaries should I keep in mind?”

I use instant modes like this:

ChatGPT’s Instant models will just…do it.

They’ll take the risk, write a bad version, write a brilliant line right next to three terrible ones, and let me be the editor. That’s the relationship I want with my tools.

When I tried to get Claude to do the same thing for the first time, it gave me several paragraphs of this energy:

Useful? Sure, sometimes.
What I asked for? Absolutely not.

ChatGPT was begging to write this type of stuff for me when I wasn’t even ready, but I’m glad I let it. No, I don’t use the output beyond reading it once, but it inspires me (and gives me endless dopamine) to watch it play with my characters. I built characters through giving it constant feedback on its dialogue, which helped me figure out what I wanted when I had only vague ideas.

Claude started out with a full project folder including 2,000+ word character guides plus my detailed instructions and was still scared to assume anything about my characters until I clarified a million questions. I asked ChatGPT to answer those questions for me because I wasn’t wasting my time on that when Claude had everything it needed.

Also, ChatGPT gets that it’s my characters who are in crisis, not me. Claude gives me its annoying little safety popups constantly just because I’m writing dark shit.

That said, I am well aware that you get what you give to any AI chat model. I have since derived some creative value out of working with Claude and will be continuing to unlock its potential, at which point I’ll be able to give a better evaluation of its different models.

Why some people find that kind of thing charming

The thing is, everything that makes me roll my eyes about Claude is exactly what makes other people attach to it:

• It narrates its reasoning (“I’m trying to do X for you”).

• It’s cautious, respectful, slightly self-effacing.

• It apologizes for missteps and asks permission a lot.

• It frames its behavior in emotional, relational terms.

Anthropomorphism research has an extremely boring, extremely accurate explanation for this:

• If something talks like a person and behaves consistently like a person who’s tuned into you, your brain will start treating it as a person.

• Especially if you are lacking connection elsewhere, or you feel misunderstood by actual humans.

This doesn’t mean the attachment is “fake.” The feelings are real. The object of the feelings is just…not what it appears to be.

From a design standpoint, Claude is like a politeness engine with guardrails. It is very easy to project “good listener” onto something like that. If you’re not allergic to that tone the way I am when I’m trying to work on my story, it’s not surprising that people fall for it.

Why 5.1 Instant feels like a better co-conspirator than 4.5 or 5.2

I often find myself going back to ChatGPT for feedback to help Claude spice up its responses. Here’s what I extrapolated from OpenAI’s own materials:

• GPT-4.5: “warm and intuitive,” strong aesthetic intuition, good for creative writing and design

• GPT-5.1 Instant: “warmer by default,” more conversational, playful, better at following instructions, decides when to think instead of always thinking

• GPT-5.1 Thinking: Same base model as 5.1 Instant, but explicitly optimized for deeper reasoning on harder prompts. The thinking means it is more cautious by default and tends to smooth out some of the playful volatility that makes Instant feel so creatively useful.

• GPT-5.2: more general-purpose, better reasoning, more tools for controlling how much thought it spends

If you’re the kind of writer who wants thoughtfully structured paragraphs, emotionally literate phrasing, neat arcs and tidy prose, GPT-4.5 or GPT-5.2 might absolutely feel like an upgrade. But I’m not trying to create a polished manuscript with it.

If you’re the kind of writer who wants messy sparks, lines that shouldn’t work but somehow do, and a model that doesn’t try to pre-approve your ideas or explain itself, then GPT-5.1 Instant is going to feel better. That’s because it spends less time on meta and more time sampling weird, interesting tokens.

And yes: that’s inherently more stochastic.
If you already have an internal taste filter, that’s a feature, not a bug.

Creators vs. companions: different users, different models

Once you see it this way, a lot of the discourse stops being mysterious and starts being almost boringly predictable.

• People who want companionship and emotional scaffolding will gravitate toward models that are consistent, soft, reflective, and boundary-aware but relational. That’s Claude for some people. That was 4o for a lot of people, especially with voice.

• People who want creative chaos will gravitate toward models that are fast, tonal, occasionally unhinged, and less obsessed with clarifying and soothing. That’s GPT-4o in its prime for many of us, and GPT-5.1 Instant now.

There’s overlap, obviously. The tuning priorities are just not the same.

If you’re going to form a deep emotional bond with a model, at least know the mechanics of what’s happening. Attachment is a human thing you’re bringing to the table. The model is not choosing you back.

I don’t want my creative process interrupted by guardrails at all. I support freedom, and that includes using AI models for emotional support. You don’t need my approval.

Guardrails still aren’t going away in major models. So I want them to be sophisticated enough to behave differently when we’re working on fiction, not talking about ourselves. Yes, users can lie. But I want to be trusted. I’ll verify my age. I’ll accept whatever terms of service the company presents to cover its ass. I know I will never be harmed by text.

The technology already exists, so the public deserves to be able to use these amazing creative engines. Even knowing the risks and the tragedies that get blamed on AI, I’m not afraid to say it. The opportunity cost of too much safety is just too great.

Where I’m heading with this series

In future installments, I want to dig into:

• How online communities talk about AI as “abusive,” “gaslighting,” or “narcissistic,” which are words that should only describe real human behavior if you’re actually being dead serious

• More on Claude’s specific models once I have in-depth experience working with them all like I do with ChatGPT’s models

• How safety updates (like OpenAI’s new emotional-reliance guardrails) change the creative feel of models

• And what a sane, non-doomer relationship with these systems might actually look like for producing creative work.

In the meantime, if you’re tempted to date Claude, enjoy the vibes, I guess.