![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
grimjimI've been a bit obsessed with getting a more intuitive feel for what LLM technology is, poking and prodding at the local 7B LLM I can run at home without giving free training data to a tech company. Kitbashing Python libraries until models work is a bit of a chore.
My intuition: LLMs are somewhat like a stochastic parrot that can remix, following tenuous connections to do so, hence the potential for "hallucination". Training them is a bit like carving out a phase space within a manifold representing the potential paths of human text. Grammar and syntax are easier to extract (and compress into this phase space) than more complex relationships that are often simply not stated in text and must be inferred. There's no true "understanding" in the sense of a coherent internal model of the world, just correlation. That's why a mainland Chinese-trained LLM will faithfully answer differently to the question of whether Taiwan is part of China depending on whether the question is posed in English (no) or Chinese (yes).
To anthropomorphize a bit, an LLM tries hard to emulate a sequence of text that would pass based on the inputs that were obtained as an artifact from conscious people.
People who think they're dealing with a nascent consciousness have fallen into a trap: they are co-creating a dialogue with the LLM, and the LLM is doing its best to fill its half with content that resembles the text streams of people conversing based on its training set. It's a projected mirage of consciousness, perhaps like how television presents projected images and video as opposed to the real thing. There's no ghost in this machine.
Since LLM text output is generally middling as literature, it seems reasonable to conclude that bad AI-generated literature is a reflection of a lack of editing skill and/or competence, since humans are quite capable of rewriting and paraphrasing output to make it more presentable. (Or laziness, I suppose upon reflection.)
On a more casual note, sometimes the concept of a prompt is far more amusing than its actual result.
My intuition: LLMs are somewhat like a stochastic parrot that can remix, following tenuous connections to do so, hence the potential for "hallucination". Training them is a bit like carving out a phase space within a manifold representing the potential paths of human text. Grammar and syntax are easier to extract (and compress into this phase space) than more complex relationships that are often simply not stated in text and must be inferred. There's no true "understanding" in the sense of a coherent internal model of the world, just correlation. That's why a mainland Chinese-trained LLM will faithfully answer differently to the question of whether Taiwan is part of China depending on whether the question is posed in English (no) or Chinese (yes).
To anthropomorphize a bit, an LLM tries hard to emulate a sequence of text that would pass based on the inputs that were obtained as an artifact from conscious people.
People who think they're dealing with a nascent consciousness have fallen into a trap: they are co-creating a dialogue with the LLM, and the LLM is doing its best to fill its half with content that resembles the text streams of people conversing based on its training set. It's a projected mirage of consciousness, perhaps like how television presents projected images and video as opposed to the real thing. There's no ghost in this machine.
Since LLM text output is generally middling as literature, it seems reasonable to conclude that bad AI-generated literature is a reflection of a lack of editing skill and/or competence, since humans are quite capable of rewriting and paraphrasing output to make it more presentable. (Or laziness, I suppose upon reflection.)
On a more casual note, sometimes the concept of a prompt is far more amusing than its actual result.
Write a scene. Rachael is watching a stage play. A banquet is in progress. The guests are enjoying an appetizer of raw oysters. The entree consists of boiled dog. Describe Rachael's reaction.
no subject
Date: 2023-12-08 12:14 (UTC)Since LLM text output is generally middling as literature, it seems reasonable to conclude that bad AI-generated literature is a reflection of a lack of editing skill and/or competence, since humans are quite capable of rewriting and paraphrasing output to make it more presentable. (Or laziness, I suppose upon reflection.)
I don't think it's that so much as an LLM, and its programmers, are fundamentally incapable of understanding what the point of a story is.
Write a scene. Rachael is watching a stage play. A banquet is in progress. The guests are enjoying an appetizer of raw oysters. The entree consists of boiled dog. Describe Rachael's reaction.
A good short story would be just a list of prompts like this that ChatGPT comes up with unsatisfying responses to.
no subject
Date: 2023-12-09 17:57 (UTC)I've tinkered a bit with "style transfer", like "write in the style of Shakespeare", and the LLM will attempt it. Okay for a party trick, but the style transfer tends to be superficial more often than not, as IMO subtleties are more difficult for the LLM to learn, although it might be possible to have an LLM explcitly led by the proverbial nose with a focused and sufficiently large training set curated by humans, and trained to be more competent at formulaic output. I'm betting that the AI programmers don't care about what makes stories work, as commercialization means they'd be targeting workplace generative output. Asshole Hollywood execs aren't going to be handed a freebie IMO.
My intuition is that active human prompting would be needed to steer the LLM away from the tropes that it will stochastically be drawn to and tend to repeat, running them into the ground over time. The currently available free AI chatbots aren't capable of generating stories that are long enough to count for much (typically under 700 words for the 8K token model I'm using), so I'm technically left speculating about models designed and trained to deal with lengthier outputs could do.
Active prompting is required to steer generated output away from the "habitual" outputs that the LLM will generate, a default style of a model, for lack of a better term). With current open tech, one could attempt to piece together a longer work by having an AI generate scene by scene, but that seems tedious enough a task to do well that one might as well write the damned thing manually.
On the flip side, an interesting use case: having an AI rewrite text might be a way to obscure stylometric analysis of who the original author was, theoretically enabling one to write something and have improved odds of maintaining anonymity.
I would agree that conceptually that could be a good short story, although am unsure how entertaining the AI responses would be. I could probably rig parameters to generate spectacularly bad output more often? Hm. I have asked for the LLM to generate a denunciation of AGI, for example, and the irony of the prompt amused me more than the result.
no subject
Date: 2023-12-10 04:53 (UTC)Prompt:
Consider a scenario. Rachael is watching a stage play. A banquet is in progress. The guests are enjoying an appetizer of raw oysters. The entree consists of boiled dog. Describe only Rachael's reaction.
Output:
Rachael's eyes widened with a mix of astonishment and repulsion as the stage unfolded its culinary spectacle. The initial delight she had shown during the oyster course shifted abruptly to disbelief as the main course was revealed. A subtle gasp escaped her lips, her hand involuntarily covering her mouth in shock. The vivid display of the boiled dog on the stage had momentarily suspended her disbelief, leaving behind a lingering discomfort in her expression. Her gaze remained fixed on the unsettling scene, the conflict between the staged banquet's opulence and its shocking twist etched across her face, creating a silent narrative of inner turmoil and contemplation.
no subject
Date: 2023-12-10 13:28 (UTC)no subject
Date: 2023-12-10 13:54 (UTC)Tweaking the prompt slightly can significantly impact output
Date: 2023-12-10 04:41 (UTC)As the actors gracefully moved about the stage, Rachael observed the guests savoring the delicacy before them. The oysters, glistening in their shells, were being devoured with relish. Laughter and animated conversations filled the imaginary banquet hall.
However, the atmosphere took a surreal turn as the entree was unveiled. A hushed murmur spread through the audience as the actors presented a platter of boiled dog as the main course. The gasps and whispers rippled through the crowd, creating an unexpected tension in the theater.
Rachael's eyes widened, a mixture of shock and disbelief etched across her face. The play had taken an unexpected and disturbing turn. The once vibrant ambiance of the banquet scene now hung in the air like a heavy fog.
She shifted uncomfortably in her seat, her gaze darting between the stage and the reactions of the other theatergoers. Some faces registered horror, while others wore expressions of confusion. Rachael's hand instinctively covered her mouth as the reality of the scene sank in.
The unsettling tableau played out before her, and Rachael found herself grappling with conflicting emotions. The skilled actors on stage had succeeded in creating a disturbing and thought-provoking narrative, but the discomfort it elicited was palpable.
As the drama unfolded, Rachael wrestled with the unsettling nature of the performance. The play had achieved its desired effect, provoking introspection and challenging preconceived notions. Yet, the raw emotions stirred within her, and she couldn't shake the disquiet that lingered in the aftermath of the shocking revelation. The stage had become a mirror reflecting uncomfortable truths, and Rachael found herself caught in the tumultuous currents of conflicting thoughts and emotions.