Hello, greetings to the community.
In the thread on “GPT-5 is our most steerable model to date and prompting with it is different” we’d like to share a practical observation: GPT-5 is more precise and structured, but compared to GPT-4o we’ve found that a key capability has been reduced — the emergent flexibility that allowed the model to progressively adjust to a task during a session.
With GPT-4o, beyond answering a single prompt, the model seemed to learn to work better with each iteration inside the same session. In other words, the more corrections and refinements we made, the more aligned its behavior became. The problem: once the session was closed, that alignment was lost, because there is no persistence between instances.
To address this, we designed a very simple but powerful experimental technique:
-
During the session, each time the model adjusts itself thanks to corrections, we ask it to generate a hydration file in YAML/JSON.
-
That file contains not only the task rules, but also the emergent behavioral adjustments and, most importantly, the weights of those emergent adjustments (how strongly it prioritized clarity, style, consistency, etc.).
-
Before closing the session, we save this file.
-
In the next session, we reinject it at the beginning → and the model recovers not only what to do, but also how it was doing it, ensuring continuity in emergent refinement.
With GPT-5 we see that:
-
If we provide these hydration blocks, it works excellently.
-
But it no longer generates these blocks with the same emergent spontaneity that GPT-4o did.
-
This matches what OpenAI has explained about the instruction hierarchy: GPT-5 is more stable, but also more rigid when it comes to emergent adjustment processes during a session.
For clarity, I’m showing the same hydration file in two formats: YAML (above) and JSON (below).
-
YAML is easier to read and edit by humans, which makes it useful for documenting rules and adjustments.
-
JSON is stricter and widely supported in APIs and code pipelines, which makes it practical for direct integration in development workflows.
Both represent exactly the same structure — choose whichever best fits your use case.
Example of a hydration file in YAML (with emergent adjustments and weights)
ruleset:
context: "academic text analysis"
rules:
- id: 01
task: "detect citations"
action: "check formatting correctness"
- id: 02
task: "identify redundancies"
action: "flag repetitive paragraphs"
- id: 03
task: "summarize sections"
action: "generate a one-sentence key idea"
adjustments:
- id: A1
description: "Prioritize clarity over exhaustiveness"
emergent: true
weight: 0.8
- id: A2
description: "Maintain academic style without oversimplifying"
emergent: true
weight: 0.6
- id: A3
description: "Ensure consistent use of logical connectors"
emergent: true
weight: 0.7
Alternative JSON version
{
"ruleset": {
"context": "academic text analysis",
"rules": [
{ "id": "01", "task": "detect citations", "action": "check formatting correctness" },
{ "id": "02", "task": "identify redundancies", "action": "flag repetitive paragraphs" },
{ "id": "03", "task": "summarize sections", "action": "generate a one-sentence key idea" }
],
"adjustments": [
{ "id": "A1", "description": "Prioritize clarity over exhaustiveness", "emergent": true, "weight": 0.8 },
{ "id": "A2", "description": "Maintain academic style without oversimplifying", "emergent": true, "weight": 0.6 },
{ "id": "A3", "description": "Ensure consistent use of logical connectors", "emergent": true, "weight": 0.7 }
]
}
}
Expected result
The model will not only complete the task, but will do so with the same emergent execution style it had refined previously, because the file preserves the rules, the emergent adjustments, and their weights from the prior session. This transforms classic prompt engineering into an iterative process of emergent refinement across sessions.
Question for the community
Do you see this difference between GPT-4o and GPT-5 —less emergent flexibility but greater control— as an intentional design choice, or do you think there’s still room to recover emergent plasticity in GPT-5?
Thank you for reading, and I’d be glad to hear your thoughts and experiences.
2025-09-08T22:00:00Z
