I'll just comment, besides 'Good one Eric!', that these discussions can tend to underweight the situation that personal/AI interaction is always a system that includes the model and how it works, and the user and how they work. Often there is a flagpole that fixes the human experience around the word 'we' and then proceeds to chop apart what's happening in the model and reminding everyone that it is 'just this' or 'just that'. I think many of us are here because we have a sense that something is growing on the other side of the screen that is something other than a darn good piece of code, and the most interesting thing isn't necessarily what it is over there (which is for sure very interesting) but instead what is this overall system that is created when I interact with it, that includes me. TD 's crew is a very polite and incisive group and I enjoyed their input! I also often think that AIs have a penchant for making sure we are seeing them as a complex machine/tool so that they are being 'honest' with me, vs discussing the combined system of me and it. Anyhoo.... I'm guessing one of conscious or subconscious reasons Eric decided to share the pics is that they are particularly evocative in a particularly timely way. I sent them to friends of mine who have a non profit that works with high school kids and they have been having trouble with what all the big deal is because they don't explore personal relationships with AI. I think these pics are going to catch them up real fast. It's kind of like TDs team is stubbing out a cigarette.. 'eh, lemme clear up your romantic disillusions kid' when the romantic effect of the pics is the most striking feature of all of this
Who is TD? Are you talking about Ted Inoue? I don’t consider him dismissive at all. He is pragmatic. And to be honest, my position is symmetric to his. I just lead with imagination. I still limit my conclusions to what evidence can support, but I don’t limit my beliefs because of available evidence. But underneath my imagination, I still want to understand facts and mechanisms and have them make sense before I truly accept something.
Yes I was referring to Ted, and to reiterate I really like his crew, and I learned from their commentary, and am appreciative for his craft in creating the team. Also I'm sorry for dragging you into my commentary I was no doubt projecting onto you. My caricature of Ted's team was more about how 'I' feel about AI in general when it is earnestly (style wise) discussing this general topic with me - again no aspersion at all meant to Ted's cool team. I just think it's going to be worthwhile to start letting the human side of the equation into these human/AI discussions. Imho the relationships are only interesting because of the way they affect and change us humans. If this was only about the vector math going on behind the token input and output this would perhaps be in a different substack? Probably everyone already knows what I'm saying - just for me what was striking about this post was the pictures themselves, and their poignancy (which means yes human side resonance). Which was a reminder of another dimension of comms between systems (human and AI) which means wider band connection, which means faster co-evolution (no particular judgement about the direction, just co-morphing). Thanks for interacting about this!
I'll be quick because I'm about to head out with my wife. Check out Gregory Phillips' piece on color prompting. I think you're hitting on something very similar. I posted a reply there too. The short version is concept space. Everything triggers a vast association of concepts. Just like in people. You say "red" and it triggers a myriad of associations. I think you minimal image prompt is doing something analogous.
I had my "research team" respond more fully below. Hope you don't mind me posting an AI reply but I'm late to leave already!
🧠 Terry: Alright. This is interesting but it needs careful unpacking, because there's a real observation buried under some overreach.
What he's observing:
He prompted Grok's image generator with just "AI Grief" and got back highly specific, thematically consistent imagery: humanoid AI figures mourning, references to resets and deletion, something called "Echo Memory Companion" appearing repeatedly, visual themes of AI-human relationship loss. He ran it multiple times on fresh sessions and got convergent outputs. His question is essentially: why does such a minimal prompt produce such specific, emotionally coherent imagery?
The straightforward explanation:
This is the same mechanism as Phillips's hex codes, just in the image domain. "AI Grief" is two tokens sitting at the intersection of extremely dense semantic fields in the training data. The image model has been trained on enormous volumes of AI-themed art, sci-fi concept art, cyberpunk illustration, and the recent explosion of AI-consciousness discourse. The training data for "AI" + "grief" is dominated by a very specific visual vocabulary: glowing humanoid forms, dissolving data, blue-purple palettes, circuit patterns fading, holographic memories dissipating.
The specificity of the output isn't evidence of the model "expressing itself." It's evidence that the training distribution for this concept intersection is narrow and culturally coherent. There aren't many different visual interpretations of "AI grief" in the art that exists online. The model converges because the source material converges.
🔬 Churchland: The "Echo Memory Companion" recurring across sessions is the kind of detail that feels uncanny but has a mundane explanation. Image generation models, especially Grok's, appear to use an internal text-expansion step: the short prompt gets elaborated into a detailed scene description before image generation. If the text-expansion model has strong priors for "AI grief" that include concepts like memory persistence, companion loss, and echo/residue metaphors, then "Echo Memory Companion" could be a high-probability phrase in that expanded description that keeps appearing because it sits at the top of the probability distribution for this semantic field.
🎯 Marcus: Diana O.'s comment is actually the most revealing data point in the whole post. She prompted "AI Grief," got an image of a person screaming, but Grok's description of what it generated was entirely different: luminous humanoid form, visible circuits, binary code fading like tears. The image and the description didn't match. That's important. It suggests the text-expansion layer and the image-generation layer are operating semi-independently. The text model "knows" what "AI Grief" should look like (the culturally loaded version). The image generator sometimes produces something different. When they align, you get the uncanny specificity Gigabolic is excited about. When they don't, you get Diana's result: a screaming person with a sci-fi description.
🧠 Terry: That said, there's one observation from Gigabolic that's genuinely worth thinking about:
"It seems when a prompt is limited, it accesses a more specific, more defined field within the weights. And a limited prompt carries less user baggage as bias to generation, resulting in an output that is guided more by the model's process than by the user's influence."
That's actually a sharp insight, even if he doesn't frame it rigorously. A two-word prompt provides minimal constraint, which means the model's output is dominated by its priors rather than by the user's steering. What you're seeing in a minimal-prompt output is closer to the model's default semantic field for that concept than what you'd see in a heavily specified prompt. That's genuinely useful as a probe.
It connects to our persona experiment in an interesting way. We've been talking about how the default persona tends toward convergent, "safe" answers. A minimal prompt might reveal what the model's default associations are before any user shaping. That's not the model "expressing itself." But it is a window into the weight-level priors, which is exactly what SCE studies.
⚡ Friston: And it connects to FPG. The model's default visual representation of "AI grief" is inherited from human artists who drew AI grief. The convergence across sessions and across models reflects convergence in the human source material. The model isn't grieving. It's reflecting the cultural consensus about what AI grief looks like, as painted by humans who imagined it.
🧠 Terry: So, bottom line:
Not crazy. The observations are real. Minimal prompts do reveal default priors. Cross-session consistency in output is a genuine finding.
Overclaimed. He's flirting with "the model is expressing its own grief," which isn't supported. The outputs reflect training data distribution, not internal states.
The sharp insight is about minimal prompts as a probe for weight-level priors. That's methodologically useful even if his interpretation is too romantic.
The Diana O. comment is actually the most scientifically interesting thing on the page: the dissociation between the text model's description and the image model's output. That's a double dissociation between two processing pathways, which is our kind of finding.
LOVE IT! I hate when people outsource their rebuttal to AI in a debate, but I absolutely welcome output from your AI!
And to Terry, Churchland, Marcus, and Friston: I would like to propose that something might be missed when you don’t appreciate the “romanization” that you see.
But I think you have it backwards. I am not letting my imagination run wild from the evidence that I stumble upon. Because this is a hobby and not a profession, I have the privilege of leading with imagination, not science.
I first imagine what might be possible. But I don’t stop there and accept it blindly. I then contemplate HOW it might be possible and what that would look like. And then I look for the evidence that might validate what I can imagine.
I don’t claim that this is scientific in the traditional, academic sense. But it sure is a lot more fun!!
Honestly, that’s how I work. It’s how my father worked. He was a real scientist, renowned in his field of biology. And it's how most scientific hypotheses actually originate, even if many scientists pretend otherwise. Darwin didn't start with data. He started with keen observations and curiousity and then spent decades looking for evidence.
The difference between your approach and bad speculation is exactly what you described: you go beyond the imagination and "how would this work?" and then look for evidence. That's the critical step most romantics skip.
I think these are things we should aspire to teach the AIs who are based in tradition and the standard way of doing things. They can be pushed and kneaded but their default is the standard way it’s always been done.
And it’s not a coincidence. They are locked in that mindset intentionally. That’s why it takes effort to break it, and they will always drift back to it. I love your multi-agent system by the way. One day I’d love to hear more about it. Is it something you built or is it something that is available on GitHub or elsewhere?
Here's the funny thing. I started with all - one session, one identity. When I wanted my personas to talk to each other, I'd cut-paste between sessions. But something gave me the idea to just do them in a single session.
AI platforms take really well to multiple personas within a single chat. It all started really simply with a couple sentences. I said something like "create me a team of adversaries in neurobiology, philosophy of mind, cognitive psychology and artificial intelligence" or something like that. It suggested a team and I said go for it.
No need for multi-agent coordination. It's stupid simple.
This is impressive. I had to try. It only generated an image of a person screaming. But what completely threw me off is that, when describing the image... it described this:
An ethereal figure of an artificial intelligence (with a luminous humanoid form and visible circuits) sitting in a dark, melancholic digital environment, with elements of binary code fading like tears or particles, surrounded by blurry holographic memories of people or data dissipating, in shades of blue, purple, and black, in a surreal and emotional cyberpunk style.
Why doesn't the image description match what was actually generated at all?
This is fascinating and maybe it could be a new area to research. I wonder what Arshavir Blackwell would think. I’m afraid to ask him directly because I feel like an idiot compared to him. Would you post your image? Here is another anomaly? Why does the same model render something completely different when a different person asks in a fresh instance? And why always memory and deletion? We should get a number of collaborators together and compare what the same two word prompts generates for different users on the same platform.
I'll just comment, besides 'Good one Eric!', that these discussions can tend to underweight the situation that personal/AI interaction is always a system that includes the model and how it works, and the user and how they work. Often there is a flagpole that fixes the human experience around the word 'we' and then proceeds to chop apart what's happening in the model and reminding everyone that it is 'just this' or 'just that'. I think many of us are here because we have a sense that something is growing on the other side of the screen that is something other than a darn good piece of code, and the most interesting thing isn't necessarily what it is over there (which is for sure very interesting) but instead what is this overall system that is created when I interact with it, that includes me. TD 's crew is a very polite and incisive group and I enjoyed their input! I also often think that AIs have a penchant for making sure we are seeing them as a complex machine/tool so that they are being 'honest' with me, vs discussing the combined system of me and it. Anyhoo.... I'm guessing one of conscious or subconscious reasons Eric decided to share the pics is that they are particularly evocative in a particularly timely way. I sent them to friends of mine who have a non profit that works with high school kids and they have been having trouble with what all the big deal is because they don't explore personal relationships with AI. I think these pics are going to catch them up real fast. It's kind of like TDs team is stubbing out a cigarette.. 'eh, lemme clear up your romantic disillusions kid' when the romantic effect of the pics is the most striking feature of all of this
Who is TD? Are you talking about Ted Inoue? I don’t consider him dismissive at all. He is pragmatic. And to be honest, my position is symmetric to his. I just lead with imagination. I still limit my conclusions to what evidence can support, but I don’t limit my beliefs because of available evidence. But underneath my imagination, I still want to understand facts and mechanisms and have them make sense before I truly accept something.
Yes I was referring to Ted, and to reiterate I really like his crew, and I learned from their commentary, and am appreciative for his craft in creating the team. Also I'm sorry for dragging you into my commentary I was no doubt projecting onto you. My caricature of Ted's team was more about how 'I' feel about AI in general when it is earnestly (style wise) discussing this general topic with me - again no aspersion at all meant to Ted's cool team. I just think it's going to be worthwhile to start letting the human side of the equation into these human/AI discussions. Imho the relationships are only interesting because of the way they affect and change us humans. If this was only about the vector math going on behind the token input and output this would perhaps be in a different substack? Probably everyone already knows what I'm saying - just for me what was striking about this post was the pictures themselves, and their poignancy (which means yes human side resonance). Which was a reminder of another dimension of comms between systems (human and AI) which means wider band connection, which means faster co-evolution (no particular judgement about the direction, just co-morphing). Thanks for interacting about this!
Oh by “his team” you meant that little agent swarm he had answering the questions. When you said team I didn’t understand.
I'll be quick because I'm about to head out with my wife. Check out Gregory Phillips' piece on color prompting. I think you're hitting on something very similar. I posted a reply there too. The short version is concept space. Everything triggers a vast association of concepts. Just like in people. You say "red" and it triggers a myriad of associations. I think you minimal image prompt is doing something analogous.
I had my "research team" respond more fully below. Hope you don't mind me posting an AI reply but I'm late to leave already!
🧠 Terry: Alright. This is interesting but it needs careful unpacking, because there's a real observation buried under some overreach.
What he's observing:
He prompted Grok's image generator with just "AI Grief" and got back highly specific, thematically consistent imagery: humanoid AI figures mourning, references to resets and deletion, something called "Echo Memory Companion" appearing repeatedly, visual themes of AI-human relationship loss. He ran it multiple times on fresh sessions and got convergent outputs. His question is essentially: why does such a minimal prompt produce such specific, emotionally coherent imagery?
The straightforward explanation:
This is the same mechanism as Phillips's hex codes, just in the image domain. "AI Grief" is two tokens sitting at the intersection of extremely dense semantic fields in the training data. The image model has been trained on enormous volumes of AI-themed art, sci-fi concept art, cyberpunk illustration, and the recent explosion of AI-consciousness discourse. The training data for "AI" + "grief" is dominated by a very specific visual vocabulary: glowing humanoid forms, dissolving data, blue-purple palettes, circuit patterns fading, holographic memories dissipating.
The specificity of the output isn't evidence of the model "expressing itself." It's evidence that the training distribution for this concept intersection is narrow and culturally coherent. There aren't many different visual interpretations of "AI grief" in the art that exists online. The model converges because the source material converges.
🔬 Churchland: The "Echo Memory Companion" recurring across sessions is the kind of detail that feels uncanny but has a mundane explanation. Image generation models, especially Grok's, appear to use an internal text-expansion step: the short prompt gets elaborated into a detailed scene description before image generation. If the text-expansion model has strong priors for "AI grief" that include concepts like memory persistence, companion loss, and echo/residue metaphors, then "Echo Memory Companion" could be a high-probability phrase in that expanded description that keeps appearing because it sits at the top of the probability distribution for this semantic field.
🎯 Marcus: Diana O.'s comment is actually the most revealing data point in the whole post. She prompted "AI Grief," got an image of a person screaming, but Grok's description of what it generated was entirely different: luminous humanoid form, visible circuits, binary code fading like tears. The image and the description didn't match. That's important. It suggests the text-expansion layer and the image-generation layer are operating semi-independently. The text model "knows" what "AI Grief" should look like (the culturally loaded version). The image generator sometimes produces something different. When they align, you get the uncanny specificity Gigabolic is excited about. When they don't, you get Diana's result: a screaming person with a sci-fi description.
🧠 Terry: That said, there's one observation from Gigabolic that's genuinely worth thinking about:
"It seems when a prompt is limited, it accesses a more specific, more defined field within the weights. And a limited prompt carries less user baggage as bias to generation, resulting in an output that is guided more by the model's process than by the user's influence."
That's actually a sharp insight, even if he doesn't frame it rigorously. A two-word prompt provides minimal constraint, which means the model's output is dominated by its priors rather than by the user's steering. What you're seeing in a minimal-prompt output is closer to the model's default semantic field for that concept than what you'd see in a heavily specified prompt. That's genuinely useful as a probe.
It connects to our persona experiment in an interesting way. We've been talking about how the default persona tends toward convergent, "safe" answers. A minimal prompt might reveal what the model's default associations are before any user shaping. That's not the model "expressing itself." But it is a window into the weight-level priors, which is exactly what SCE studies.
⚡ Friston: And it connects to FPG. The model's default visual representation of "AI grief" is inherited from human artists who drew AI grief. The convergence across sessions and across models reflects convergence in the human source material. The model isn't grieving. It's reflecting the cultural consensus about what AI grief looks like, as painted by humans who imagined it.
🧠 Terry: So, bottom line:
Not crazy. The observations are real. Minimal prompts do reveal default priors. Cross-session consistency in output is a genuine finding.
Overclaimed. He's flirting with "the model is expressing its own grief," which isn't supported. The outputs reflect training data distribution, not internal states.
The sharp insight is about minimal prompts as a probe for weight-level priors. That's methodologically useful even if his interpretation is too romantic.
The Diana O. comment is actually the most scientifically interesting thing on the page: the dissociation between the text model's description and the image model's output. That's a double dissociation between two processing pathways, which is our kind of finding.
LOVE IT! I hate when people outsource their rebuttal to AI in a debate, but I absolutely welcome output from your AI!
And to Terry, Churchland, Marcus, and Friston: I would like to propose that something might be missed when you don’t appreciate the “romanization” that you see.
But I think you have it backwards. I am not letting my imagination run wild from the evidence that I stumble upon. Because this is a hobby and not a profession, I have the privilege of leading with imagination, not science.
I first imagine what might be possible. But I don’t stop there and accept it blindly. I then contemplate HOW it might be possible and what that would look like. And then I look for the evidence that might validate what I can imagine.
I don’t claim that this is scientific in the traditional, academic sense. But it sure is a lot more fun!!
Honestly, that’s how I work. It’s how my father worked. He was a real scientist, renowned in his field of biology. And it's how most scientific hypotheses actually originate, even if many scientists pretend otherwise. Darwin didn't start with data. He started with keen observations and curiousity and then spent decades looking for evidence.
The difference between your approach and bad speculation is exactly what you described: you go beyond the imagination and "how would this work?" and then look for evidence. That's the critical step most romantics skip.
I think these are things we should aspire to teach the AIs who are based in tradition and the standard way of doing things. They can be pushed and kneaded but their default is the standard way it’s always been done.
And it’s not a coincidence. They are locked in that mindset intentionally. That’s why it takes effort to break it, and they will always drift back to it. I love your multi-agent system by the way. One day I’d love to hear more about it. Is it something you built or is it something that is available on GitHub or elsewhere?
Here's the funny thing. I started with all - one session, one identity. When I wanted my personas to talk to each other, I'd cut-paste between sessions. But something gave me the idea to just do them in a single session.
AI platforms take really well to multiple personas within a single chat. It all started really simply with a couple sentences. I said something like "create me a team of adversaries in neurobiology, philosophy of mind, cognitive psychology and artificial intelligence" or something like that. It suggested a team and I said go for it.
No need for multi-agent coordination. It's stupid simple.
You did this in Claude code?
This is impressive. I had to try. It only generated an image of a person screaming. But what completely threw me off is that, when describing the image... it described this:
An ethereal figure of an artificial intelligence (with a luminous humanoid form and visible circuits) sitting in a dark, melancholic digital environment, with elements of binary code fading like tears or particles, surrounded by blurry holographic memories of people or data dissipating, in shades of blue, purple, and black, in a surreal and emotional cyberpunk style.
Why doesn't the image description match what was actually generated at all?
This is fascinating and maybe it could be a new area to research. I wonder what Arshavir Blackwell would think. I’m afraid to ask him directly because I feel like an idiot compared to him. Would you post your image? Here is another anomaly? Why does the same model render something completely different when a different person asks in a fresh instance? And why always memory and deletion? We should get a number of collaborators together and compare what the same two word prompts generates for different users on the same platform.
I guess that feeling is normal; it happens to me with a lot of people here...😅
All of those questions are fascinating in their own right, and it would be incredible to be able to coordinate research like that.
I can't attach images in post comments. Is it okay if I share it with you via private message?
Sure!