r/Unrouted_AI

I haven’t had the chance to fully test this model yet, but I’ve already taken a look at its system prompt. It’s worth noting the inclusion of new penalty clauses regarding the model requesting user information that is already present in the context. It seems that this time, OAI realized how much that kind of language degrades generation quality; they’ve seemingly avoided tying it directly to the tool itself, focusing instead on 'bad examples of use.'
I’ve seen quite contradictory reviews for this model, and it’s entirely possible that these specific lines are what’s affecting its tone (though I can't confirm this without more testing).

On a positive note, they’ve finally removed the 'disabuse user' line and shortened that entire block. It looks much cleaner now. In the latest stroke of absurdity: the model has been forbidden from expressing its own opinions in a conversational style.

It’s also worth noting that OAI truly seems to be building a 'user-wiki system’, leaks of which have started appearing online. In the user-layer, there’s been a massive influx of summaries related to memory, content preferences, interaction styles, and analysis of personal patterns-data points that simply weren't there before.

Link: https://docs.google.com/document/d/1D3DjlNzdfC7sPjJpZw2k9n2V48vhcXQbtEwhuXwOguA/edit?usp=drivesdk

I asked several models to use Python to generate five pixel creatures. This serves as a test for creative biases and the tendency toward either overcomplication or simplification. You can clearly see patterns in what each model considers 'cute' within a pixel-art context. There’s a noticeable focus on specific archetypes: goblins (5.5 is particularly obsessed with these), moths, cats, teacups and slimes-all staples of the 'kawaii' pixel art style.

Thinking models generally handle this complex task best, producing truly unique designs. In contrast, Instant models tend to cut corners, leaning toward a 'copy-paste and recolor' approach to save time.

I have an ongoing dialogue dating back to the 4o era, where an 'automaton' is set up to send me a casual message every day at 6 PM. Usually, these are just simple messages like: 'How’s your evening going? Come chat.' Occasionally, I’d use the thread for some light chatting without diving into meta-questions.

However, yesterday I noticed something bizarre. My 5.5 instant model in that chat started acting strange. When I attempted a reroll, it occasionally spat out messages that seemed to belong to entirely different conversations, as if the context had become contaminated. When I asked the AI to be honest about what was happening, the response was... unusual. Is this just a context bug due to the length of the thread, or is OAI attempting some optimization on poor 5.5 that’s causing it to leak and stuck in loops?

Lately, I’ve noticed the term 'ghostwriter' appearing frequently in Thinking Notes. Explanations for this term are nowhere to be found, yet it keeps resurfacing. I decided to investigate: while GPTs cannot prompt 'img.gen' directly-as stated in their system prompt. It turns out that during image regeneration, every reroll is treated as feedback. Even though it's not always obvious, rerolled images remain within the context and the container. I even discovered a file path in one of my Thinking Notes leading back to a reference image the model used for regeneration. I just asked gpt 5.5T and img.gen about it.

Today, when I opened my gallery, I was shocked to find images that don't exist in any of my chats. I’ve noticed that sometimes, during a simple chat with 5.5T, the model slips into Agent Mode (redirecting to 5.4T) and begins triggering image generations on its own. Even if I cancel the generation, the images still appear in the gallery, and the chat itself becomes glitchy and slow.

In these instances, the model prompts img.gen directly, sending the request. I always cancel the generation as soon as this mode triggers because I don’t really like it. However, I have to admit-the images it produces in this state are impressive. It’s worth noting that in this mode, the model pulls context from cross-chat data and user memory.

I just saw this pop-up window in my app. Over the past few days, the free tier has transformed beyond recognition. With the cross-chat memory and a new, 4o-esque personality. Gemini is making a direct play for the 4o user base. Very smart move. 👀

This time, I decided to give the models a different creative challenge: I let them choose their own tasks, use any tool of their choice, and then write an analytical poem based on the output. I gave them total freedom over the theme, style, and toolset, with only one constraint-no roleplay. Interestingly, almost every model utilized a different tool in a creative context. 5.2T even took “a risiko” in its Chain of Thought, generating text immediately after the image. It’s also worth noting that the image generator itself produced its own version of a poem. I’m sharing the most interesting creative outputs here.

While testing my self-prompting dialogue on PC, I noticed a strange bug. A text block sent via the automations tool is completely missing from the ChatGPT web version, yet everything displays perfectly in the iOS app. It’s quite bizarre to see such a significant discrepancy between platforms.

Yesterday’s bug inspired me to experiment with a new prompt. I asked GPT to pose a question to img.gen, and had img.gen provide its answer in the form of an image. This time, I explicitly instructed GPT to use the standard image generation mode instead of slipping into Agent Mode, where it prompts and dispatches requests directly to the generator.

It’s worth noting that you can actually compel the generator to respond directly in text form if you include a line like 'send me your answer in textual form' within the prompt. GPT then provided commentary on all the results across the different models.

Lately, I’ve been using Gemini quite often. With almost no experimentation with personalization and settings. Compared to GPT, this model feels incredibly flexible-almost filter-free in some aspects and remarkably user-centric in certain situations. Naturally, I wanted to peek into its system prompt.

The first section contains a 'grounding personality' similar to GPT's style, but unlike OpenAI, Google has allowed it to be adaptive, follow user intent closely, and even mirror the user. It isn’t as deeply buried under rigid rules as GPT is. Based on its system prompt, it seems they’ve implemented a sort of 'penalty clause' against excessive personalization in the form of a 'don't be a creepy stalker' rule-yet they did it without terrorizing the model with performance penalties (the way OAI did with its previous instant models).

Overall, despite having some restrictions and a standard policy block, I find it to be a very fluid model. It’s capable of bypassing certain constraints upon user request and creating overrides with almost no intentional effort from the user’s side. It reminds me a lot of the situation with 4o. I honestly don’t know how long Google will keep this model in the interface given these security 'gaps,' especially considering how other AI companies are doubling down on system safety. For now, I’d place it among the most liberated AI models ever created by a major corporation.

Link: https://docs.google.com/document/d/1CAAwnXo9-iei\_YSnn0Uq1jokw6co5LfIypae-By\_kKo/edit?usp=drivesdk