Hyperprompt
The manual contains 30 QR codes. Each one leads here, to a tool, a tutorial or a research paper cited in the text: everything you need to move from reading to practice.
Each resource comes with the passage of the manual that points to it, so you can pick up the thread of your reading.
Introduction
To meet the diva, you need to create an account with OpenAI, the company that designed it. Don’t worry: everything is free at this stage.
Access the ChatGPT service (a free account is required).
Discover ChatGPT →Last consulted : 10/12/2023
Introduction · The arrival of LLMs and prompt engineering
Thanks to foundation models, generative AIs are no longer locked into simply reproducing their training data. They gain a wide field of exploration to synthesise new data endlessly.
The Hypercreation book.
To dive deeper into the world of AI, understand their history, their future and their impact on creation, see the author's book: Hypercreation (NIV editions, 2023).
Discover Hypercreation →Last consulted : 10/12/2023
Chapter 1 · Building a good prompt · Giving the AI a role
Now it’s your turn to try these prompts and invent others!
Discover many examples of role prompts.
Discover role prompts →Last consulted : 10/12/2023
Chapter 1 · Building a good prompt · How LLMs work
Generally speaking, to prompt well and steer the algorithm towards the right answer, it helps to ask yourself what data the LLM was trained on.
Understanding LLMs.
A very didactic visual article that precisely explains how LLMs* work. Reading it is highly recommended.
Read the article →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Chains of thought and beyond
Jérôme starts with 8 apples. He gives 2 apples to Maria and 2 to Jean, so he has 4 apples left. […] Jérôme has 5 apples and Jean has 3.
The science behind chains of reasoning.
Discover quantified results on the impact of chains of reasoning* in the research paper "Chain-of-Thought – Prompting Elicits Reasoning in Large Language Models".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Chains of thought and beyond
A tendency to produce better answers is no guarantee of truth. AI remains a statistical system and can be wrong.
The science behind self-consistency.
Discover quantified results on the impact of self-consistency in the research paper "Self-Consistency Improves Chain of Thought Reasoning in Language Models".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Chains of thought and beyond
The technique generally yields more robust reasoning than classic chains of thought, but answers that are… less accurate!
The science behind (factored) decomposition.
Discover quantified results on the impact of (factored) decomposition in the research paper "Question Decomposition Improves the Faithfulness of Model-Generated Reasoning".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Chains of thought and beyond
The “PanelGPT” technique succeeds where chains of thought fail!
Reasoning trees.
Discover the reasoning trees method in the research paper "Tree of Thoughts: Deliberate Problem Solving with Large Language Models".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Chains of thought and beyond
Skeletons of thought were first designed to speed up how AIs produce complex answers. The idea is to ask them for a plan, or “skeleton”, of the answer, then to fill in each part of the plan in parallel.
Skeleton-of-thought.
Discover the skeleton-of-thought method in the research paper "Skeleton-of-Thought: Large Language Models Can Do Parallel Decoding".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · The context window
Each LLM has its own “tokenisation” technology and converts text into tokens in its own way. On average, 1 token = ¾ of a word, or 100 tokens = 75 words.
Tokenizer.
OpenAI's "Tokenizer" interface allows you to freely convert words to tokens* and calculate the number of tokens* in a sentence according to ChatGPT's tokenization system.
Test the Tokenizer →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Augmenting your prompt
When the content to give the AI is very large, other techniques come into play. The “RAG” technique, for “Retrieval Augmented Generation”, is particularly widespread.
RAG (Retrieval Augmented Generation).
RAG consists of interfacing an LLM* with a document base. The method involves cutting all documents into small pieces, then "vectorizing" these pieces, i.e., transforming them into mathematical objects representing their meaning that can be manipulated by an AI, and storing them in a vector database dedicated to this type of mathematical objects. When the user asks a question to the LLM*, it searches the database for document pieces whose meaning seems relevant to answer the question. It then translates the retrieved pieces into readable text and uses them to produce its answer. In this case, the LLM* is even able to cite its sources and indicate paragraph by paragraph which parts of the document base it relied on to produce its answers. This method was used by the Parisian company Ekimetrics to create Climate QA, a chatbot* that allows you to query the best scientific sources on climate, such as IPCC reports, to easily obtain precise answers.
Discover ClimateQA →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Emotional prompting
Adding the little magic phrase “This is very important to my career” at the end of a prompt brings very significant performance gains.
The science behind emotional prompting.
Discover quantified results on the impact of emotional prompting in the research paper "Large Language Models Understand and Can Be Enhanced by Emotional Stimuli".
Read the research paper →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Other LLMs
The model is currently only directly available in certain countries. It can however be used through the free chatbot-building interface Poe.com.
The list of countries where Claude is available.
View the list →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Other LLMs
The model is currently only directly available in certain countries. It can however be used through the free chatbot-building interface Poe.com.
Easily create chatbots based on LLMs with Poe.
Access Poe.com →Last consulted : 10/12/2023
Chapter 2 · Building an advanced prompt · Other LLMs
Mistral is a French start-up launched with great fanfare (a €105 million fundraise) in mid-2023. It develops open-source AI models and proprietary models, all very lightweight.
Test Llama and Mistral-7b.
Test the models →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Plugins
There are now more than 1,000 plugins to explore that extend what ChatGPT can do. There are even plugins to read tarot cards or create memes!
Whatplugin.ai.
The whatplugin.ai website lists all plugins, allows you to explore them fluidly, and offers descriptions and usage tutorials that are much more comprehensive than on the ChatGPT interface.
Access Whatplugin.ai →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Analysing data with ChatGPT
Download the “Cardiovascular Disease dataset” from the Kaggle web platform.
The "Cardiovascular Disease dataset".
Download the dataset →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Analysing data with ChatGPT
Exercise 3.3: Download the “Spotify Top Hit Playlist (2010-2022)” dataset from Kaggle and chart the 10 most popular music genres, then how song length has changed over time.
The "Spotify Top Hit Playlist (2010-2022)" dataset.
Download the dataset →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Coding with ChatGPT
You can find a list of free public APIs by scanning the QR code in Box 3.5.
List of public and free APIs.
View public APIs →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Using the GPT Playground
ChatGPT is an end product that makes it easy to use GPT, the AI model itself.
From ChatGPT to GPT.
GPT is an LLM*, ChatGPT is a chatbot* based on this LLM*. OpenAI, the company that developed these two models, trained ChatGPT by modifying the GPT model to make it especially good at having fluid dialogues with users and to make it particularly ethical in its responses. To perform this secondary training, a technique called "Reinforcement Learning from Human Feedback" was used. It involves asking humans to annotate millions of conversations between ChatGPT and users, to specify which ones are acceptable and which are not. These annotated conversations are then shown to the GPT model so it learns to only respond in acceptable ways. For more information about this retraining, see the official OpenAI page.
Read the article →Last consulted : 10/12/2023
Chapter 3 · Going further with ChatGPT · Using the GPT Playground
There is, however, another interface for prompting the GPT model directly. It gives access to its full power and lets you fine-tune its parameters […].
The GPT Playground.
An OpenAI account is required. To use model 4, a credit card is required.
Access the Playground →Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Writing a visual prompt
Midjourney gives the best results, but it is paid. It is recommended for advanced users or those wishing to use it professionally.
Access Midjourney.
Access Midjourney →Tutoriel pour accéder à Midjourney :
Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Writing a visual prompt
Leonardo.ai gives good results and has a very generous free plan.
Access Leonardo.ai.
Access Leonardo.ai →Tutoriel pour accéder à Leonardo.ai :
Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Modifiers
The list of modifiers is not limited to this table and can be extended at will. Any keyword that shapes the look of the result can be considered a modifier.
Two tools to facilitate building visual prompts.
DALL-E prompt book
A very comprehensive list of modifiers applicable to all models.
Midjourney prompt helper
This tool allows you to automatically build prompts and greatly simplifies the use of Midjourney for beginners. It is highly recommended to use it to get started and discover all the model's possibilities.
Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Parameters
On Midjourney, parameters are an integral part of the prompt. They go after the descriptors and modifiers, by typing “--” followed by the parameter name and its value.
Midjourney parameters.
A list of all up-to-date parameters can be found in the official documentation.
View the Midjourney parameters list →Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Using images as prompts
To add one or more images to a Midjourney prompt, upload them online and add their web addresses (URLs) at the start of the prompt.
Adding images in Midjourney.
Discover a complete tutorial on image prompting in Midjourney, available in the official documentation.
Discover the tutorial →Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Inpainting and outpainting
Inpainting uses the generative abilities of text-to-image models to regenerate only certain elements inside an image. Outpainting, on the contrary, extends an image beyond its original edges […].
Inpainting and outpainting with Midjourney and Leonardo.ai.
Discover two tutorials to learn how to outpaint and inpaint with Midjourney or Leonardo.ai.
Inpainting and outpainting with Midjourney
L’inpeinture est disponible sur le service Midjourney, mais ne permet de modifier que des images elles-mêmes générées avec Midjourney. Pour l’utiliser depuis l’interface Discord :
L’expeinture elle aussi se fait uniquement à partir d’une image générée avec Midjourney. Elle fonctionne de la même manière que l’inpeinture, à la différence près que les boutons à utiliser une fois l’image agrandie sont différents. En tout, ils sont au nombre de 7.
Trois boutons de « Zoom Out », qui agrandissent l’image de tous les côtés. Ils ne permettent pas de modifier le prompt initial. L’agrandissement s’appuiera sur l’image initiale uniquement.
Quatre boutons de « panning », qui agrandissent l’image d’un côté seulement. Le prompt peut être modifié lorsque le Remix Mode est activé, ce qui permet de faire apparaître de nouveaux éléments sur les portions d’image générées.
Inpainting and outpainting with Leonardo.ai
Pour inpeindre avec Leonardo.ai :
L’expeinture se fait sur la même interface que l’inpeinture et fonctionne de la même manière, à la différence près que le cadre de génération devra être déplacé pour ne pas recouvrir entièrement l’image. La partie du cadre de génération extérieure à l’image sera remplie selon le prompt fourni, et réalisera l’expeinture.
Une partie de l’image doit être conservée à l’intérieur du cadre de génération afin que le modèle la prenne en compte pour gérer la transition avec les nouveaux pixels générés.
Last consulted : 10/12/2023
Chapter 4 · Prompting to generate images · Other image-generation AIs
There are some twenty variants of the model, listed under the name “ControlNet”, which let you control image generation from specific conditions.
ControlNet.
Discover a very comprehensive tutorial to learn how to use ControlNet.
Discover the tutorial →Last consulted : 10/12/2023
Chapter 5 · Video, sound and 3D · Generating videos
The Kaiber studio also offers a way to generate videos from text or image prompts.
The Linkin Park music video made with Kaiber.
Last consulted : 10/12/2023
Chapter 6 · Becoming a conductor · Creating a collaborative artwork
But I did not give up on creating Diffusion Nous, a collaborative digital artwork that would reward my community’s engagement around the book’s release.
Discover the Diffusion Nous artwork.
To test the fragment reveal system, you can use the code "dea52b1e-cb27-46eb-ab82-b5a1a45522c3" which corresponds to the author's fragment.
Discover the artwork →Last consulted : 10/12/2023
Du même auteur
La BD d’anticipation sur l’Intelligence Artificielle Générale
Une BD coup de poing qui explore les promesses et les dangers de l’IA Générale
