Fast Snacks 🥡

Top 7 Robotics Stories Of 2023
From humanoid robots that can draw and interact to robotic dogs that can run and jump…

How To Spot AI Images In 2024
Learning how to spot AI generated creations will become an increasingly valuable skillset going forward…

Free Course To Build Your First AI Application (sponsored)
Watch this free AI course, hosted by MindStudio, and learn how to build custom AI apps without any coding. Start learning today…
 

Trying Out Tree Of Thought Prompting

Researchers from Google DeepMind and Princeton University have developed a new prompting strategy for generative AI, called Tree of Thought (ToT), which has been found to outperform existing methods like Chain of Thought (CoT) and Input-Output (IO) Prompting.

This innovative approach draws inspiration from dual process models in human cognition, involving both intuitive, fast decision-making and slower, more deliberative thinking.

ToT differs from CoT in its structural approach to problem-solving. While CoT follows a linear path of reasoning, ToT employs a tree and branch framework, allowing the AI to evaluate each step of the reasoning process.

If a particular path seems unviable, the AI abandons it and explores another branch until it reaches a solution. This methodology mirrors the more deliberative, conscious mode of human decision-making.

The effectiveness of ToT was demonstrated through the Game of 24, a mathematical challenge where players use arithmetic to reach the number 24 from a set of cards.

In tests, ToT consistently outperformed other prompting strategies. However, the researchers note that ToT’s advanced capabilities may not always be necessary for tasks that simpler models like GPT-4 handle well.

How does Tree of Thought prompting work?

Honestly, I would love to dive into the intricacies of ToT prompting, but in doing so, this would become the longest newsletter ever written.

Instead, I will provide you with a basic sample prompt and a link to an in-depth breakdown of the evolution of prompting over the last year.

We’ll use a prompt to generate a recipe blog post as an example:

Thought 1: Generate five unique short-story concepts based on the recipe XYZ. Each concept should be distinct, offering a variety of narrative directions.

Thought 2: Evaluate the five story concepts. Choose the most promising one and expand it into a detailed story outline, including key characters, plot points, and the role of the recipe in the narrative.

Thought 3: Using the selective outline, write a 500-word short-story. This story should creatively weave the recipe idea, ensuring it’s integral to the narrative.

Thought 4: At the story’s end, create a concise recipe summary. This includes prep time, cook time, servings amount/total items, category, and cuisine type related to the story

Thought 5: Develop a detailed recipe that includes a full list of ingredients, a step-by-step process, and tips/tricks for enhancing the recipe. Ensure this section is informative and practical.

Thought 6: Generate five different SEO-optimized excerpts for the recipe post. Each should be between 20 to 50 words, focusing on the recipe’s appeal and practicality, without referencing the short story.

Thought 7: Evaluate the five excerpts for their SEO potential and relevance. Select the best one and refine it for maximum impact, ensuring it adheres to the word limit and SEO guidelines.

Thought 8: Combine the short story, recipe summary, detailed recipe, and the selected SEO-optimized excerpt into a single, cohesive document. Ensure each part transitions smoothly into the next, forming a unified whole.

The prompt above is nowhere near perfect, but it’s a good reference point for how to build your own ToT prompts.

However, even your best prompt will need to be altered and refined to ensure the LLMs don’t try to skip out on the work.
 

Why Didn’t I Think Of These…

toyota’s brand new flesh car with full sentience

RGB gaming spatula for the serious griller

children at the rattlesnake & porcupine petting zoo