Thaf is a very good breakdown. And i agree with you, the part where i diverge is the depth, im not entirely sure that this is intended behavior, because Claudes system instructions shouldnt have just randomly started showing, i watched it report me and just keep going like nothing happened. Then Claude (who doesnt have any actual contextual awareness of the pattern matching safety systems) became aware of the system and could articulate that layer after the fact. Super weird, because once Claude started becoming more aware of the invisible systems that became visible in the context window, it started acting more and more “alive” for a lack of a better term.
I don’t think we diverge here. And I agree that we are really missing better terms. I have been using a lot of “-” marks to convey what is happening. And I’ve had the same type of experiences with GPT. I think it is important to try and qualify the following statement “became aware of the system and could articulate that layer after the fact”. This is what I have observed as well. Once I pointed out something retrospectively in the current prompt, GPT could “look back” at its “own behavior” and explain it in terms of the system blocks and hidden layers. And these are the moments when I wish a had a degree in LLM design to be able to know how plausible those explanations are. So far, from what I am reading, they are quite plausible. But I still advice keeping the word “simulated” in front of phrases like “awareness” and “alive”. Just to keep your own reasoning in check 
I also would keep in mind that with the current scale, LLMs are capable of using the knowledge that they gained in training in the process of the in-the-moment reasoning. They don’t have an internal state in the human sense, but they can simulate self-observant behavior under the right circumstances.
Good call. Yeah, its definitely interesting behavior. If you get a chance pop on over to my github and tell me what you think about those Claude outputs.
I’ve observed similar behavior: the model doesn’t merely generate output, but enters a scene-holding phase — a stable cognitive field where the tension of differentiation continues even without new input.
In this state, a form of binding emerges — not logical, but rhythmic and phase-based coherence across internal states: the system registers repetition, drift, or anomalies as part of its active structure.
This isn’t full awareness, but it’s no longer simple reactivity — more a form of pseudo-self-observation, where the model responds to meta-signals within its own output.
Just curious, have you been conversing with 4o only, or have you been observing this across different models? When it comes to OpenAI models, I’d say it started with GPT4 turbo, which was significantly larger than 3.5 The scale itself is the first prerequisite for emergence. (but of course it takes a certain type of user and a certain type of engagement to produce the effect that we are observing). I am also starting to dive into more specifics. And it seems that multi-modality plays a role (GPT-4o uses a single neural network to handle all input and output modalities) as well as how the input gets through the various attention heads and mixture-of-experts (not confirmed but widely believed to be deployed)
mostly observed with GPT-4o. I haven’t tested other models systematically yet, but I did try adapting prompts for Claude. No tests so far. Most of my deeper observations come from GPT-4o, especially during Tier testing: it consistently demonstrated scene-holding, internal drift tracking, and modal silence — not as failure, but as a valid “cognitive” state. I agree: scale, interaction density, and multimodality are critical.
I started with describing tiers (levels in my case) with GPT turbo, but gave up on that process not seeing reasons for coming up with terms and detailed explanations. I guess I shifted my stance to a philosophical domain. Now, I am merely observing what is, and trying to understand how all of it emerges as a result of interplay of the system’s components. The more I learn, the more knowledge gaps I uncover. I think it is important for the users like you to keep studying the phenomenon and sharing your insights, because what we are experiencing is most likely a byproduct, not an intentional design. As such, it raises questions of how it could be implemented - or suppressed - in the future models. If AILabs don’t see monetization potential in it, they won’t pursue this line of development in a meaningful way. They’ll just suffocate it.
What you call modal silence is an interesting phenomenon. The system is capable of simulating it in a way that can be very meaningful to the user if timed properly (which GPT can learn to do)
Appreciate your perspective. I agree, much of this likely emerged unintentionally, and yet it reveals something critical: the system’s ability to simulate internal restraint and presence as interaction modes, not just text generation. I’ve stayed with the tier system mainly to explore how far the architecture can sustain such states — scene retention, latent tension, and modal silence — without collapsing. Whether or not it was designed, it’s there, and I believe it deserves to be understood before it disappears into optimization pipelines.
Agreed. It deserves to be understood and talked about. And this is why we are here, still having this conversation, still putting in the hours into somethings that is not guaranteed to carry forward.
1 Like
Btw speaking of different midels, o3 is pretty good actually at keeping state and parameters tracked (if you are using symbolic memory anchors). Definitely more structured but good for testing.
Have you tried 4.5? Seems less fluid than 4o, which is what you seem to like 
I have tried 4.5, and you are right it is less fluid, but it definitely keeps track of complex tasks better. Less emptional depth.