\n<\/p>\n
There is a small experiment I have been running with Gemini that has left me with a rather interesting question.<\/p>\n
I am not entirely sure what is happening.<\/p>\n
And perhaps that is the most interesting part.<\/p>\n
The experiment started with a fairly ordinary question: How deep can an underground excavation realistically go?<\/strong><\/p>\n Rather than immediately reaching for engineering drawings, excavation manuals, geological reports and a frightening amount of technical documentation, I decided to do something simpler first.<\/p>\n I asked Gemini to show me.<\/p>\n The idea was to generate images of a large underground excavation site and gradually modify the prompt until the image looked something like what I imagined a serious excavation project should look like.<\/p>\n Nothing particularly revolutionary there.<\/p>\n The unusual part was what happened next.<\/p>\n I kept working from essentially the same prompt<\/strong>, making relatively small edits and adjustments. I did not maintain a long conversation in which I repeatedly told Gemini what was wrong with the previous image. I did not give it a score. I did not say, \u201cThat last image was 30% better than the one before it.\u201d<\/p>\n There was no explicit feedback loop.<\/p>\n And yet, the images seemed to get better.<\/p>\n That raised a much more interesting question:<\/p>\n Can you improve AI-generated results simply by repeatedly refining the prompt, even when the model isn\u2019t given an explicit mechanism for measuring its own improvement?<\/strong><\/p>\n I don\u2019t have a definitive answer.<\/p>\n But the experiment is fascinating.<\/p>\n The first image was, frankly, rather crude.<\/p>\n It conveyed the basic idea: there was a large hole, there was earth, there was an excavation site, and something was happening underground.<\/p>\n Technically speaking, mission accomplished.<\/p>\n Artistically and structurally?<\/p>\n Not quite.<\/p>\n It looked more like an AI\u2019s first approximation of \u201cvery large hole\u201d than a serious construction site.<\/p>\n So I started tweaking.<\/p>\n The prompt gained additional descriptions.<\/p>\n Then more.<\/p>\n I began specifying things that should exist in a real excavation: structural support, machinery, access routes, retaining systems, workers, scale indicators, layers of earth, underground structures and the general visual language of a major engineering project.<\/p>\n The process was not:<\/p>\n Generate \u2192 evaluate \u2192 tell the AI exactly what it did wrong \u2192 generate again.<\/p>\n<\/blockquote>\n It was closer to:<\/p>\n Generate \u2192 think \u2192 modify prompt \u2192 generate \u2192 think \u2192 modify prompt \u2192 generate.<\/p>\n<\/blockquote>\n The conversation itself was not carrying the experiment forward as a traditional feedback loop.<\/p>\n The prompt<\/strong> was.<\/p>\n And that distinction is what caught my attention.<\/p>\n Eventually, the images began looking considerably more convincing.<\/p>\n The excavation started to have depth.<\/p>\n The machinery started to make more sense.<\/p>\n The supporting structures became more plausible.<\/p>\n The scene began communicating scale.<\/p>\n The final result looked much closer to the sort of enormous excavation site I had originally been trying to visualize.<\/p>\n And that is where things became weird.<\/p>\n At least, not in the conventional sense.<\/p>\n I wasn\u2019t building a dataset.<\/p>\n I wasn\u2019t fine-tuning a model.<\/p>\n I wasn\u2019t giving Gemini a collection of successful and unsuccessful examples and asking it to learn from them.<\/p>\n I wasn\u2019t even necessarily continuing the same conversational thread.<\/p>\n I was changing the instructions.<\/p>\n That sounds like a subtle distinction, but technically it matters.<\/p>\n A generative AI system receives an input and produces an output based on the information and statistical patterns available to it. When generating an image, the prompt provides semantic constraints describing what the model should attempt to create.<\/p>\n If the prompt says:<\/p>\n \u201cA very deep underground excavation site.\u201d<\/p>\n<\/blockquote>\n the model has an enormous amount of freedom.<\/p>\n How deep?<\/p>\n What type of excavation?<\/p>\n What kind of soil?<\/p>\n What support system?<\/p>\n What equipment?<\/p>\n What scale?<\/p>\n What construction stage?<\/p>\n What viewpoint?<\/p>\n What does \u201crealistic\u201d mean in this context?<\/p>\n There are thousands of possible interpretations.<\/p>\n The more useful constraints we introduce, the smaller that space of possible interpretations becomes.<\/p>\n So perhaps the improvement is not mysterious at all.<\/p>\n Perhaps I was simply reducing ambiguity<\/strong>.<\/p>\n But that doesn\u2019t completely explain what I was seeing.<\/p>\n One way to think about prompting is that you are not giving the model a recipe.<\/p>\n You are steering it through an enormous space of possibilities.<\/p>\n A vague prompt gives the model considerable freedom.<\/p>\n A more precise prompt constrains the generation.<\/p>\n For example:<\/p>\n Version 1<\/strong><\/p>\n A very deep underground excavation site.<\/p>\n<\/blockquote>\n That\u2019s enough to establish a concept, but not enough to establish engineering reality.<\/p>\n Now imagine adding:<\/p>\n A large-scale deep excavation for an underground facility, with vertical retaining walls, staged excavation, heavy construction machinery, temporary structural bracing, workers for scale, access ramps, exposed geological layers and realistic construction logistics.<\/p>\n<\/blockquote>\n Suddenly, the model has considerably more information about the scene.<\/p>\n We can keep going:<\/p>\n A large-scale deep excavation approximately 17 metres deep, showing multiple excavation stages, reinforced retaining walls, internal bracing, construction equipment operating at different elevations, access ramps, workers for scale, exposed soil and rock layers, realistic drainage and construction logistics\u2026<\/p>\n<\/blockquote>\n Now we are no longer merely asking for \u201ca hole.\u201d<\/p>\n We are describing a construction environment<\/strong>.<\/p>\n The prompt has become a compressed specification.<\/p>\n And each small edit can potentially remove another ambiguity.<\/p>\n That may explain much of the apparent improvement.<\/p>\n But there is still an intriguing detail.<\/p>\n Normally, when we talk about machine learning improving through feedback, we imagine some kind of evaluation mechanism.<\/p>\n The system generates something.<\/p>\n It receives feedback.<\/p>\n It determines whether the result was good or bad.<\/p>\n It adjusts.<\/p>\n Repeat.<\/p>\n That is a genuine feedback loop.<\/p>\n My experiment wasn\u2019t doing that.<\/p>\n There was no explicit numerical score.<\/p>\n There was no instruction saying:<\/p>\n \u201cCompare this image with the previous one and improve the engineering realism by 20%.\u201d<\/p>\n<\/blockquote>\n There was no formal reward function.<\/p>\n There wasn\u2019t even a carefully constructed evaluation rubric.<\/p>\n I was simply looking at the result and thinking:<\/p>\n Okay. That doesn\u2019t look quite right. Let\u2019s change the prompt.<\/strong><\/p>\n Then the next generation often looked better.<\/p>\n So what exactly improved?<\/p>\n The model?<\/p>\n The prompt?<\/p>\n My understanding of the problem?<\/p>\n Or all three in some indirect way?<\/p>\n That is where the experiment becomes much more interesting.<\/p>\n This is probably the simplest explanation.<\/p>\n The model doesn\u2019t necessarily need to evaluate itself if I am doing the evaluation<\/strong>.<\/p>\n Every generated image gives me information.<\/p>\n I look at it.<\/p>\n I notice that something is missing.<\/p>\n I add that missing concept to the prompt.<\/p>\n The next image therefore contains more constraints.<\/p>\n In that sense, the feedback loop exists \u2014 it just isn\u2019t entirely inside the AI system.<\/p>\n It looks something like this:<\/p>\n Prompt \u2192 Image \u2192 Human observation \u2192 Prompt modification \u2192 Image \u2192 Human observation \u2192 Prompt modification<\/strong><\/p>\n The human is effectively acting as the evaluator.<\/p>\n But there is an interesting twist.<\/p>\n The later prompts don\u2019t necessarily need to explicitly mention the previous image.<\/p>\n The accumulated knowledge can exist entirely in the evolving prompt.<\/p>\n That means the prompt itself becomes something like a portable state representation<\/strong>.<\/p>\n The conversation can disappear.<\/p>\n The thread can be restarted.<\/p>\n The model doesn\u2019t necessarily need to remember what happened previously.<\/p>\n If the prompt has captured the useful refinements, the experiment can continue.<\/p>\n That is potentially quite powerful.<\/p>\n There is another way of describing what happened.<\/p>\n It resembles an informal optimization process.<\/p>\n Imagine that the space of possible images is enormous.<\/p>\n We start with a very broad instruction:<\/p>\n \u201cGenerate a deep excavation.\u201d<\/p>\n<\/blockquote>\n The result lands somewhere in that enormous space.<\/p>\n We inspect it and decide that we want something slightly different.<\/p>\n So we modify the input.<\/p>\n The next generation lands somewhere else.<\/p>\n We inspect again.<\/p>\n Another modification.<\/p>\n Another generation.<\/p>\n In mathematical language, we might loosely describe this as iterative search through a generative space<\/strong>.<\/p>\n We aren\u2019t necessarily optimizing a clearly defined numerical objective.<\/p>\n Instead, we\u2019re performing qualitative optimization.<\/p>\n The objective is something like:<\/p>\n \u201cMake this look more like a plausible large-scale excavation.\u201d<\/p>\n<\/blockquote>\n That objective is evaluated by a human.<\/p>\n The prompt changes are our search steps.<\/p>\n And the generated images are our candidate solutions.<\/p>\n This is not the same thing as model training.<\/p>\n It is closer to interactive optimization through prompt-space exploration<\/strong>.<\/p>\n And that distinction is important.<\/p>\n One of the more interesting characteristics of generative models is that seemingly small changes to an instruction can alter the output substantially.<\/p>\n This happens because natural-language prompts aren\u2019t simply interpreted as rigid commands.<\/p>\n They contain concepts, relationships, emphasis and context.<\/p>\n Adding a single phrase can change the semantic interpretation of an entire scene.<\/p>\n Compare:<\/p>\n \u201cdeep excavation\u201d<\/p>\n<\/blockquote>\n with:<\/p>\n \u201cdeep excavation with temporary retaining structures and internal bracing\u201d<\/p>\n<\/blockquote>\n The second prompt doesn\u2019t merely add two objects.<\/p>\n It changes the implied engineering context<\/strong>.<\/p>\n Likewise:<\/p>\n \u201cunderground construction\u201d<\/p>\n<\/blockquote>\n is dramatically different from:<\/p>\n \u201ca large-scale underground construction project viewed from above, with workers, excavators, access ramps, retaining walls and multiple levels of excavation visible.\u201d<\/p>\n<\/blockquote>\n The model now has much stronger clues about composition, scale and purpose.<\/p>\n So the improvement may come from a phenomenon that is actually quite mundane:<\/p>\n better instructions produce a narrower and more appropriate generation space.<\/strong><\/p>\n But then again\u2026<\/p>\n Why did the progression feel so consistent?<\/p>\n That\u2019s the part I haven\u2019t solved.<\/p>\n This is where I want to be careful.<\/p>\n I cannot claim that Gemini itself was learning from the previous attempts.<\/p>\n I did not conduct a controlled experiment designed to establish that.<\/p>\n There are several competing explanations.<\/p>\n Perhaps the prompt became progressively better.<\/p>\n Perhaps my own understanding of what I wanted became progressively better.<\/p>\n Perhaps the image generation system is particularly sensitive to certain descriptive concepts.<\/p>\n Perhaps the generation process contains additional mechanisms that influence results.<\/p>\n Perhaps randomness played a role.<\/p>\n Or perhaps several of these effects were occurring simultaneously.<\/p>\n Without a controlled experiment, it would be difficult to distinguish them.<\/p>\n For example, to properly test the hypothesis, we could generate hundreds of images using carefully controlled prompts, random seeds where available, identical settings and independent sessions.<\/p>\n We could then have blind evaluators score the results for realism, engineering plausibility, composition and scale.<\/p>\n That would begin to tell us whether the apparent improvement is statistically meaningful.<\/p>\n I wasn\u2019t doing that.<\/p>\n I was just trying to figure out what a giant underground hole might actually look like.<\/p>\n And that is precisely why the result surprised me.<\/p>\n Not literally smarter, of course.<\/p>\n But more informative.<\/p>\n A prompt can evolve from a simple description into a miniature specification language.<\/p>\n At the beginning, it describes the object<\/strong>.<\/p>\n Then it describes the environment<\/strong>.<\/p>\n Then the relationships between objects<\/strong>.<\/p>\n Then the physical constraints<\/strong>.<\/p>\n Then the visual perspective<\/strong>.<\/p>\n Then the desired level of realism<\/strong>.<\/p>\n Eventually, it starts resembling a technical brief.<\/p>\n This progression is worth thinking about.<\/p>\n We often hear advice about writing the \u201cperfect prompt.\u201d<\/p>\n But perhaps the perfect prompt isn\u2019t something you write once.<\/p>\n Perhaps it is something you discover iteratively<\/strong>.<\/p>\n You start with a rough idea.<\/p>\n Generate.<\/p>\n Observe.<\/p>\n Modify.<\/p>\n Generate.<\/p>\n Observe.<\/p>\n Modify.<\/p>\n Eventually, the prompt itself becomes a record of everything you have learned about how to describe the desired result.<\/p>\n That could be considerably more useful than endlessly trying to invent the perfect prompt from scratch.<\/p>\n Naturally, while doing this, another question eventually appears:<\/p>\n How much is this costing me in quota?<\/strong><\/p>\n Image generation isn\u2019t necessarily an unlimited resource.<\/p>\n Different AI services can impose different limits on image generation, requests, usage tiers or daily quotas.<\/p>\n And repeated experimentation can consume those limits surprisingly quickly.<\/p>\n But here I have to put an asterisk on the experiment.<\/p>\n I wasn\u2019t specifically measuring quota depletion.<\/p>\n I was investigating the quality of the images.<\/p>\n I didn\u2019t establish a controlled before-and-after measurement of quota usage.<\/p>\n So I cannot definitively say how much quota this particular refinement process consumed, nor can I claim that changing the prompt in this manner has some particular quota efficiency.<\/p>\n That would require a separate experiment.<\/p>\n And perhaps that\u2019s another experiment worth doing.<\/p>\n Because if iterative prompt refinement can produce substantially better results without requiring dozens of radically different prompts, there could be a practical efficiency advantage.<\/p>\n Instead of throwing 50 unrelated prompts at an image generator and hoping one produces something good, you might be able to start with one reasonably good prompt and walk it toward the desired result<\/strong>.<\/p>\n That is a very different workflow.<\/p>\n This is probably my favourite takeaway.<\/p>\n There is a tendency to approach generative AI by accumulating prompts.<\/p>\n You try one.<\/p>\n Then another.<\/p>\n Then another.<\/p>\n Then you search the internet for somebody else\u2019s \u201cultimate prompt.\u201d<\/p>\n Then you add more adjectives.<\/p>\n Then more instructions.<\/p>\n Eventually you have a monster prompt that looks like it was written by a committee.<\/p>\n The experiment suggests another possibility.<\/p>\n Don\u2019t necessarily create more prompts. Evolve one.<\/strong><\/p>\n Start with a simple prompt.<\/p>\n Identify what is missing.<\/p>\n Add it.<\/p>\n Generate again.<\/p>\n Identify another weakness.<\/p>\n Refine it.<\/p>\n Generate again.<\/p>\n The prompt becomes progressively more specific without necessarily becoming radically different.<\/p>\n That is a much more organic approach to prompt engineering.<\/p>\n And it has an interesting psychological benefit too.<\/p>\n You are not trying to predict exactly what the model needs before you begin.<\/p>\n You\u2019re allowing the output to tell you what your prompt failed to communicate.<\/p>\n The image becomes part of the prompting process.<\/p>\n And yet, after all this explanation, I still can\u2019t completely shake the original question.<\/p>\n Why did the images appear to improve so consistently?<\/p>\n If the model wasn\u2019t being given an explicit performance score\u2026<\/p>\n If I wasn\u2019t telling it, \u201cthis image is better than the last one\u201d\u2026<\/p>\n If I wasn\u2019t necessarily continuing the same conversational context\u2026<\/p>\n Then what exactly are we observing?<\/p>\n Maybe the answer is beautifully simple.<\/p>\n The model wasn\u2019t learning.<\/p>\n I was.<\/strong><\/p>\n Every image showed me something about how the system interpreted my instructions.<\/p>\n Every imperfect generation gave me another clue.<\/p>\n Every prompt revision encoded that clue.<\/p>\n And the next generation benefited from the accumulated information.<\/p>\n In that interpretation, there was never any mysterious self-improvement happening inside the model.<\/p>\n The intelligence was distributed across the entire loop:<\/p>\n human \u2192 prompt \u2192 model \u2192 image \u2192 human \u2192 improved prompt \u2192 model \u2192 improved image.<\/strong><\/p>\n But there is still a lingering possibility that makes the experiment worth repeating.<\/p>\n Perhaps there are properties of modern generative systems that make this kind of iterative refinement work better than our simple explanation suggests.<\/p>\n Perhaps the relationship between prompt wording and generated imagery is more structured than it appears.<\/p>\n Perhaps repeatedly steering the same conceptual target allows us to navigate the model\u2019s generative space surprisingly efficiently.<\/p>\n Or perhaps I simply got better at asking for what I wanted.<\/p>\n I don\u2019t know yet.<\/p>\n And that\u2019s what makes this experiment fun.<\/p>\n I started by asking an AI to draw a hole in the ground.<\/p>\n I ended up wondering whether we have stumbled upon a surprisingly effective way of searching through a generative model\u2019s possibilities:<\/p>\n Keep the idea. Keep the prompt mostly intact. Change just enough to move the result. Repeat.<\/strong><\/p>\n No giant prompt library.<\/p>\n No elaborate self-evaluation system.<\/p>\n No enormous feedback architecture.<\/p>\n Just a human, a model, a picture that isn\u2019t quite right, and one more tiny edit.<\/p>\n Then another.<\/p>\n And another.<\/p>\n Until suddenly the hole starts looking like something engineers might actually have to build.<\/p>\n And I still don\u2019t completely know why.<\/p>\n<\/div>\n\n","protected":false},"excerpt":{"rendered":" There is a small experiment I have been running with Gemini that has left me with a rather interesting question. I am not entirely sure what is happening. And perhaps that is the most interesting part. The experiment started with a fairly ordinary question: How deep can an underground excavation realistically go? Rather than immediately […]<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[215457,215458,215459,215460,203185,215461,108,192263,215462,192256,215463,156590,106835,215464,147923,215465,215466,11896,215467,215468,215469,215470,215471],"tags":[29521,8329,15867,22923,111598,9502],"dealstore":[],"offerexpiration":[],"class_list":["post-7053152","post","type-post","status-publish","format-standard","hentry","category-ai-creativity","category-ai-experimentation","category-ai-image-generation","category-ai-prompts","category-ai-tools","category-ai-assisted-design","category-artificial-intelligence","category-bored-giant","category-construction-visualization","category-developer-diary","category-engineering-visualization","category-gemini","category-generative-ai","category-generative-models","category-google-gemini","category-image-generation","category-iterative-prompting","category-machine-learning","category-prompt-engineering","category-prompt-optimization","category-prompt-refinement","category-quota","category-underground-excavation","tag-accidental","tag-experiment","tag-image","tag-prompt","tag-refinement","tag-slightly"],"yoast_head":"\nFrom a crude hole in the ground to something that looks like an excavation<\/h2>\n
\n
\n
I wasn\u2019t teaching the model<\/h2>\n
\n
The prompt became a kind of steering mechanism<\/h2>\n
\n
\n
\n
The strange part: there was no explicit score<\/h2>\n
\n
Maybe the human is the feedback loop<\/h2>\n
Prompt engineering as iterative search<\/h2>\n
\n
\n
Why tiny changes can have surprisingly large effects<\/h2>\n
\n
\n
\n
\n
Did Gemini actually improve?<\/h2>\n
There is another possibility: the prompt is becoming smarter<\/h2>\n
And then there is the quota question<\/h2>\n
Less prompt sprawl, more prompt evolution<\/h2>\n
The unresolved mystery<\/h2>\n