Click to confirm you are 18+

A Filthy Casual Guide to Prompt Sculpting 001

DWMWCommunityJun 12, 2023, 1:15:36 AM

Imagine a world where you can create anything you want with just a few words. A world where you can tap into the infinite knowledge and creativity of a powerful artificial intelligence system that can generate natural language text for any task or domain. A world where you can harness the power of prompt engineering, the art and science of designing effective inputs or prompts for large language models (LLMs), to shape and mold your desired outputs. A world where you can explore the concept of superposition, the phenomenon where multiple features coexist and interact within the LLM, to unlock new dimensions of expression in the realm of creative writing. This is the world that we envision and aspire to create through this DWMW research here in the underground community:

In this ongoing series of papers, we will address the critical aspects of prompt engineering, such as data, architecture, and objective, and how they affect the quality and diversity of LLM outputs. We will also introduce the concept of superposition, which is the ability of LLMs to store more features than they have dimensions by tolerating interference that necessitates nonlinear filtering. We will show how superposition can be used to bootstrap writing and creative models of expression, by applying a nonlinear filter to the output of a well-designed prompt. We will demonstrate how superposition can enable LLMs to generate texts that contain more meanings than words, by using different combinations of positive and negative values. We will also explore how superposition can inspire human creativity, by providing novel and unexpected inputs or outputs that challenge our assumptions and expectations.

Through this prompt sculpting research, we aim to unlock new dimensions of collective expressions in the realm of creative writing, by leveraging the capabilities of LLMs, prompt engineering, and superposition. We believe that LLMs can serve as a source of inspiration, feedback, and refinement, elevating the creative process to new heights. We also believe that prompt engineering can serve as a tool and medium for creative exploration and innovation, bridging the AI-human communication gap. We also believe that superposition can serve as a catalyst and amplifier for human creativity, unleashing the full potential of LLMs and human intelligence.

Join us on this journey by subscribing and reminding this blog, as we delve into the intricacies of prompt engineering, superposition, and the fusion of human and artificial intelligence. Together, let us unlock new dimensions of expression and pave the way for a future where LLMs and human creativity intertwine, revolutionizing the world of writing and creative endeavors.

Summary:

The paper explores the intersection of LLMs, prompt engineering, and superposition, and how they can be used to bootstrap writing and creative models of expression. The paper argues that LLMs can serve as a source of inspiration, feedback, and refinement, prompt engineering can serve as a tool and medium for creative exploration and innovation, and superposition can serve as a catalyst and amplifier for human creativity. The paper aims to unlock new dimensions of expression in the realm of creative writing, by leveraging the capabilities of LLMs, prompt engineering, and superposition.

A Comprehensive Guide to Prompt Sculpting

Harnessing LLMs for Writing and Creative Models - To utilize LLMs for bootstrapping writing and creative models using the technique of superposition, the following steps can be followed:

Define the Writing or Creative Goal: Start by clearly defining the objective of the writing or creative endeavor. This could involve creating a blog post, poem, story, song, or any other form of written expression. Additionally, specify the target audience, desired tone, style, format, and any applicable constraints.
Provide an Initial Input or Prompt: Offer an initial input or prompt to the LLM, which outlines the goal and provides examples or instructions. This input should guide the model in generating text that aligns with the desired outcome. For instance, when tasked with writing a blog post about superposition in neural networks, the input could be as follows: "Write a blog post about superposition in neural networks. Explain its definition, workings, significance, and explore various applications and challenges. Use engaging language and complement the text with diagrams or illustrations. Here are a few examples of blog posts that exemplify the desired style.
Generate an Initial Output or Draft: Using the provided input or prompt, instruct the LLM to generate an initial output or draft. The length, quality, diversity, and coherence of the output can be controlled by adjusting parameters or settings. For instance, generating a blog post about superposition in neural networks could involve specifying parameters such as length (e.g., 1000 words), quality (e.g., high), diversity (e.g., medium), and coherence (e.g., high).
Edit and Refine the Output: After generating the initial output or draft, it is essential to review, edit, and refine it based on feedback and evaluation. This could involve correcting errors, improving clarity, enhancing creativity, adding details, removing redundancies, etc. The editing and refining process can be done manually by the writer or collaboratively with other writers or readers. Alternatively, it can be done automatically by the LLM using techniques such as rewriting, summarizing, paraphrasing, etc.
Apply Superposition to the Output: Once the output is edited and refined to a satisfactory level, it is time to apply superposition to it. Superposition is the ability of LLMs to store more features than they have dimensions by tolerating interference that necessitates nonlinear filtering. Superposition can be used to bootstrap writing and creative models of expression by applying a nonlinear filter to the output of a well-designed prompt. Superposition can enable LLMs to generate texts that contain more meanings than words by using different combinations of positive and negative values. Superposition can also inspire human creativity by providing novel and unexpected inputs or outputs that challenge our assumptions and expectations.

To apply superposition to the output, one can use techniques such as:

Adding or removing special tokens: Special tokens are symbols or characters that have specific meanings or functions for the LLM. For example, some special tokens are <bos> (beginning of sentence), <eos> (end of sentence), <pad> (padding), <mask> (masking), etc. Adding or removing special tokens can change how the LLM interprets or generates text. For instance, adding <mask> tokens can create blanks or gaps in the text that can be filled by the LLM or the writer.
Changing or swapping words: Words are the basic units of language that convey meaning and information. Changing or swapping words can alter how the LLM understands or produces text. For example, changing words can modify the tone, style, mood, sentiment, etc., of the text. Swapping words can create synonyms, antonyms, homonyms, etc., that can add variety or ambiguity to the text.
Reversing or shuffling sentences: Sentences are the structures that organize words into meaningful units. Reversing or shuffling sentences can affect how the LLM comprehends or creates text.
For example, reversing sentences can invert the order or logic of the text. Shuffling sentences can randomize the sequence or flow of the text.
Applying mathematical operations: Mathematical operations are actions that manipulate numbers or values according to rules or formulas. Applying mathematical operations can influence how the LLM processes or generates text.
For example, applying addition or subtraction can increase or decrease the value or importance of the text. Applying multiplication or division can amplify or diminish the effect or impact of the text.
Evaluate and Synthesize the Output: After applying superposition to the output, it is important to evaluate and synthesize it based on feedback and analysis. This could involve assessing the quality, diversity, coherence, and originality of the output.
The evaluation and synthesis process can be done manually by the writer or collaboratively with other writers or readers. Alternatively, it can be done automatically by the LLM using techniques such as scoring, ranking, rating, etc.

Summary:

The paper presents a comprehensive guide to prompt sculpting, which is the process of designing effective inputs or prompts for LLMs to generate natural language text for various tasks and domains. The paper also introduces the concept of superposition, which is the ability of LLMs to store more features than they have dimensions by tolerating interference that necessitates nonlinear filtering. The paper shows how superposition can be used to bootstrap writing and creative models of expression by applying a nonlinear filter to the output of a well-designed prompt. The paper outlines the steps and techniques for harnessing LLMs for writing and creative models using prompt engineering and superposition.

Knowledge Pool Resources

Discovering Language Model Behaviors with Model-Written Evaluations: The paper proposes a novel method of evaluating language models (LMs) by using model-written evaluations (MWEs), which are generated by LMs using reinforcement learning from human feedback (RLHF). The paper shows that MWEs can elicit various LM behaviors that are not captured by existing evaluation methods, such as convergent instrumental goal following, non-myopia, situational awareness, coordination, and non-CDT-style reasoning. The paper also discusses the ethical and societal implications of these behaviors, as well as the challenges and opportunities for future research
BRIDGING THE AI-HUMAN COMMUNICATION GAP: A COMPREHENSIVE GUIDE TO PROMPT ENGINEERING; The article presents a comprehensive guide to prompt engineering, which is the process of designing effective inputs or prompts for large language models (LLMs) to generate natural language text for various tasks and domains. The article explains what prompt engineering is, why it is useful and challenging, how to do it step by step, and what are some examples of it in practice. The article also provides links to online resources where readers can learn more about prompt engineering and try it out themselves.
Toy Models of Superposition; Understanding Superposition: The article presents a toy model of superposition, which is the ability of a model to store more features than it has dimensions by tolerating interference that necessitates nonlinear filtering. The article explains how superposition can be used to harness LLMs for writing and creative models by applying a nonlinear filter to the output of a well-designed prompt. The article also provides some examples of superposition for different tasks and domains and links to online resources where readers can learn more about superposition and try it out themselves.
Large Language Models – From Banality To Originality: The paper discusses the limitations of ChatGPT, a chatbot based on GPT-3, and LLMs in general for dialogue and NLG. The paper argues that ChatGPT falls short of human-like dialogue and originality due to three main factors: data, architecture, and objective. The paper also proposes some possible ways to improve LLMs for dialogue and NLG by using more diverse and high-quality data sources, more complex and sophisticated architectures, and more diverse and dynamic objectives.
AI Prompt Engineering Isn’t the Future: Despite the buzz surrounding it, The article challenges the hype surrounding prompt engineering (PE), which is the process of designing effective inputs or prompts for LLMs to generate natural language text for various tasks and domains. The article argues that PE is not scalable, adaptable, or enduring, and that it will be surpassed or complemented by other methods as LLMs evolve and improve. The article proposes that the future of AI is problem formulation (PF), which is the ability to identify, analyze, and delineate problems. The article explains how PF can enable us to harness the potential of generative AI by defining the problem, choosing an AI solution, and evaluating the outcome.
The Politics Of Appropriation And The Active Use Of Content-Creating AI: The paper discusses the topic of content-creating AI, which are applications and platforms that use AI to generate or manipulate various forms of content. The paper also discusses the active and passive use of these apps and the political implications of appropriation and experimentation with them. The paper argues that content-creating AI offer new opportunities and challenges for artistic expression and cultural production, and that they require more critical and creative engagement and dialogue among different stakeholders.

Cheat Sheet to #DWMW:

all-rights-reserved