The author puts the latest AI reasoning models - ChatGPT 03 Mini, DeepSEEK R1, and Google Gemini's Flash Thinking - to the ultimate test across 10 different prompts. The prompts start easy and progressively become more complex, challenging the models' reasoning capabilities.

The key findings are:

• ChatGPT 03 Mini: Performed well on the initial prompts, providing quick and accurate responses. However, it struggled with more complex tasks like solving the "Humanity's Last Exam" question.

• DeepSEEK R1: Demonstrated strong reasoning abilities, breaking down problems into smaller steps and providing detailed explanations. It excelled at the creative problem-solving task but had issues with the coding challenge.

• Google Gemini's Flash Thinking: Delivered fast responses, often matching the speed of ChatGPT. It correctly identified the image as being created by Midjourney, showcasing its multimodal capabilities. However, it faltered on the search-based prompts, providing outdated information.

The author concludes that each model has its strengths and weaknesses, and the choice of the "best" model depends on the specific task at hand. The comprehensive test highlights the current capabilities and limitations of these advanced AI reasoning models, providing valuable insights for those looking to leverage these technologies.

For the first set of prompts, I tested the basic reasoning abilities of the three AI models - ChatGPT 03 Mini, DeepSEEK R1, and Google Gemini's Flash Thinking.

The prompts included:

1. Finding the number of Rs in "Strawberry": All three models were able to correctly identify that there are 3 Rs in the word "Strawberry". ChatGPT was the fastest, taking only 5 seconds, while DeepSEEK took 88 seconds to provide a detailed breakdown of its reasoning process. Gemini was also quick, answering in just a few seconds.

2. Comparing the values of 9.11 and 9.9: Again, all three models correctly identified that 9.9 is larger than 9.11. ChatGPT and Gemini were faster, while DeepSEEK took 37 seconds to arrive at the answer.

3. Determining whether the chicken or the egg came first: All three models agreed that the egg came first, as the first true chicken likely hatched from an egg laid by a bird that was almost, but not exactly, a chicken.

Overall, for these basic reasoning tasks, the models performed well, with ChatGPT and Gemini being the fastest, while DeepSEEK provided more detailed reasoning behind its answers.

You have a rope that is exactly 50 ft long and a building that is 75 ft tall. You need to measure the height of the building using only the rope and your own body (you are 5ft tall) - no other tools. Describe the steps to solve this problem.

ChatGPT (03 Mini): The answer provided by ChatGPT (03 Mini) does not make practical sense. The suggested approach of hanging the 50ft rope over the top of the building and using your 5ft body as a ruler to measure the remaining 25ft is flawed, as you cannot continuously stand on top of yourself to measure the full height.

DeepSeeK (R1): DeepSeeK provided a more logical solution by using the concept of similar triangles. The steps are:

1. Hold the rope vertically against the building, with one end on the ground.
2. Fold the rope until the length matches your height (5ft).
3. Count how many 5ft segments fit along the height of the building.
4. Multiply the number of segments by 5ft to determine the building's height of 75ft.

Google Gemini (Flash Thinking): Google Gemini's approach is similar to DeepSeeK's, also utilizing the similar triangles technique. The steps outlined are:

1. Hold the rope vertically against the building.
2. Fold the rope until the length matches your height (5ft).
3. Count how many 5ft segments fit along the height of the building.
4. Multiply the number of segments by 5ft to calculate the building's height.

Both DeepSeeK and Google Gemini provided a more coherent and practical solution to this creative problem-solving challenge compared to ChatGPT's approach.

The third prompt set focused on testing the models' abilities in logical deduction. Here's how the different AI models performed:

ChatGPT 03 Mini: The statement is paradoxical - neither true nor false in a consistent way. ChatGPT's response acknowledged the paradoxical nature of the statement, but did not provide a definitive conclusion on whether the statement above is true or false.

DeepSEEK R1: DeepSEEK determined that the statement above is true. It provided a thorough explanation, breaking down the logic and reasoning behind this conclusion.

Google Gemini: Google Gemini's response was that the statement above is false. It also went through a detailed thought process to arrive at this answer.

Based on the provided answer key, the correct answer is that the statement above is true. Therefore, DeepSEEK was the only model that got this prompt right, while the other two arrived at incorrect conclusions.

This prompt highlighted DeepSEEK's strength in logical reasoning and its ability to provide a well-explained answer, even for complex, paradoxical statements. The other models struggled to fully grasp the nuances of this logical deduction problem.

Prompt engineering is a powerful tool for unlocking the full potential of AI models. By crafting well-designed prompts, you can leverage these reasoning models to boost your productivity and strategy across a wide range of tasks.

Some key benefits of using prompts with reasoning models include:

1. Improved Accuracy: Reasoning models excel at breaking down complex problems into smaller, more manageable steps. This can lead to more accurate and reliable outputs compared to traditional language models.

2. Enhanced Creativity: Prompts that encourage creative problem-solving can inspire these models to generate innovative solutions and ideas.

3. Streamlined Workflows: Carefully crafted prompts can automate repetitive tasks, saving you time and effort.

4. Personalized Assistance: Tailoring prompts to your specific needs and preferences can provide you with more relevant and useful outputs.

To get the most out of these reasoning models, it's important to keep your prompts relatively simple and focused. Avoid overly complex or lengthy prompts, as these models work best when breaking down tasks into smaller, manageable steps.

The free resource from HubSpot mentioned in the transcript can be a valuable tool in your prompt engineering arsenal. With over a thousand expertly crafted prompts across various categories, it can help you save time and unlock new possibilities for your productivity and strategy.

Remember, the key to success with these reasoning models is to experiment, iterate, and find the prompts that work best for your specific needs. By leveraging the power of prompts, you can unlock new levels of efficiency, creativity, and strategic thinking in your work.

The AI models were tested on their ability to create and modify a chess game with a specific rule change. Here's how they performed:

Chat GPT-3 Mini:

• Able to create a working chess game with the modified rule that the king can move like the queen.
• Correctly implemented the new rule, allowing the king to move multiple spaces like the queen.
• However, it had some issues with the endgame logic, not properly detecting checkmate situations.

DeepSeeK R1:

• Struggled to create a working chess game, even with multiple prompts and attempts.
• The model was unable to properly handle the visual elements and icons required to build the game.
• Despite the reasoning capabilities, it failed to produce a functional chess game implementation.

Google Gemini Advanced:

• Like DeepSeeK, Gemini was also unable to create a working chess game without additional prompts and troubleshooting.
• The model was able to understand the task and provide step-by-step instructions, but could not translate that into a functional chess game.
• The visual and coding aspects proved challenging for Gemini's capabilities in this test.

Overall, Chat GPT-3 Mini demonstrated the best performance in this coding challenge, successfully creating a chess game with the modified king movement rule, though with some limitations in the endgame logic. The other models, despite their reasoning abilities, struggled to translate the task into a working implementation.

Chat GPT says there is no reliable way to tell which AI model produced the given image. It acknowledges the lack of a definitive answer.

Deep Seek was unable to extract any text from the image, even though it contained text, and therefore could not provide an analysis.

Gemini, on the other hand, analyzed the image in detail. It noted the photorealistic style, artistic flair, and concluded that the image was likely created using Midjourney, the AI image generation model. Gemini provided a thorough justification for its assessment, demonstrating its reasoning capabilities.

In this test, Gemini outperformed the other models in its ability to analyze the visual characteristics of the image and make an informed guess about the AI system used to generate it. The other models were unable to provide a meaningful response to this task, highlighting Gemini's superior image analysis capabilities compared to Chat GPT and Deep Seek in this particular scenario.

The author conducted a thorough test to compare the reasoning capabilities of three AI models: ChatGPT 03 Mini, DeepSEEK R1, and Google Gemini's Flash Thinking. The test covered a wide range of prompts, starting with simple questions and gradually increasing in complexity, including creative problem-solving, logical deduction, coding tasks, and even an unsolved mathematical problem.

The key findings from the comparison are:

1. Simple Reasoning Tasks: All three models performed well on the initial set of simple reasoning tasks, with ChatGPT 03 Mini and Google Gemini providing the fastest responses, while DeepSEEK R1 took more time but offered more detailed reasoning.

2. Creative Problem-Solving: For the creative problem-solving task, DeepSEEK R1 and Google Gemini provided more coherent and logical solutions using the similar triangles technique, while ChatGPT 03 Mini's solution was less convincing.

3. Logical Deduction: On the logical deduction prompt, DeepSEEK R1 provided the correct answer, while ChatGPT 03 Mini and Google Gemini gave different responses, with one stating the statement was neither true nor false.

4. Coding Tasks: For the coding task of creating a modified chess game, ChatGPT 03 Mini and DeepSEEK R1 were able to fix the provided code and create a functioning game, with some minor issues. Google Gemini, however, was unable to run the code without additional prompts.

5. Multimodal Reasoning: When presented with an image created by Midjourney and asked to identify the AI model, Google Gemini was the only one to correctly identify the image source, while the other two models were unable to provide a definitive answer.

6. Combining Reasoning and Search: When asked to provide the best AI model, the models had varying levels of success in combining their reasoning capabilities with search functionality to provide up-to-date and comprehensive answers.

Overall, the comparison highlights the strengths and weaknesses of each model, with no single model emerging as the clear winner across all the tested scenarios. The author suggests that the choice of the best AI model may depend on the specific task and requirements, and that a combination of models may be necessary to achieve the desired results.

The video explores the ability of various AI models, including ChatGPT, DeepSEEK, and Google Gemini, to tackle challenging mathematical problems. One of the problems presented is the Goldbach Conjecture, an unsolved problem in number theory that has remained elusive for hundreds of years.

The Goldbach Conjecture states that every even integer greater than 2 can be expressed as the sum of two prime numbers. For example, 4 = 2 + 2. Despite being one of the oldest unsolved problems in mathematics, the conjecture remains unproven.

The video demonstrates that even the advanced AI models tested, with their reasoning capabilities, are unable to solve this long-standing mathematical problem. When asked to solve the Goldbach Conjecture, the models acknowledge the difficulty of the problem and the lack of a general proof, rather than attempting to provide a solution.

This highlights the limitations of current AI systems in tackling problems that have eluded human mathematicians for centuries. The video suggests that while these models excel at various tasks, they are not yet capable of solving complex, unsolved mathematical problems that require groundbreaking insights and proofs.

The inability of the AI models to solve the Goldbach Conjecture serves as a reminder of the ongoing challenges in artificial intelligence and the continued need for human ingenuity and mathematical breakthroughs to advance our understanding of the fundamental nature of numbers and their relationships.

The testing of the different AI reasoning models across a variety of prompts has provided some interesting insights:

• ChatGPT 3.5 mini performed well on the initial simpler prompts, quickly providing accurate answers. It struggled more on the complex math and coding challenges, but was able to adapt and fix the chess game code.

• DeepSEEK R1 took longer to process the prompts, but demonstrated strong reasoning abilities, particularly on the geometric problem-solving and logical deduction questions. However, it had difficulty with the coding task.

• Google Gemini's Flash Thinking model was the fastest overall, but had some limitations. It was unable to handle the coding prompt or the image analysis question, but provided a comprehensive answer on the "best AI model" query.

• None of the models were able to solve the longstanding Goldbach Conjecture math problem, highlighting the limitations of current AI reasoning capabilities when it comes to unsolved complex mathematical problems.

Overall, the results suggest that each model has its own strengths and weaknesses. ChatGPT excels at general reasoning and language tasks, DeepSEEK is stronger at complex logical and mathematical problems, while Gemini prioritizes speed. The choice of model may depend on the specific needs of the task at hand. Continued advancements in AI reasoning will be necessary to tackle the most challenging problems.