Thank you so much for this outstanding video; its complexity and advanced nature truly stand out. I greatly appreciate you sharing your research and insights on how to circumvent errors and biases. I confess, I’ll need to revisit this video a few more times before I can fully incorporate the techniques you’ve so brilliantly demonstrated. As always, your work and unique presentation style continue to astound me.
David, you are the best! Thank you so so much for your support and encouragement. I love reading your comments, they are so insightful and refreshing! I am so humbled and grateful to know that you enjoy my content and if you ever need any help feel free to connect with me! ❤
Hi Goda. Recently discovered your channel and finding it very inspiring. I haven't been through all of your content yet, which when I do, might answer this request. Would you be able to suggest where would be the best to start point please Best wishes Lee
I love the story of the university lecturer that failed his entire class because he asked "chatgpt" if it wrote each student's paper he pasted in 😅🤦🏻♂️
I doubt that anecdote is true but it embodies a question I’ve had for several months-whether ChatGPT keeps a record of each user’s requests, and if it uses this record in answering other users’ requests. Again, I doubt that the robot really told the professor that it wrote papers for the students he named for it to check. To use President Trump’s neologism, this is likely “fake news.”
Hey Goda, I just watched your video and I must say that it was really informative and well-made. I was wondering if I could help you edit your videos and also make highly engaging shorts for you?
They claim AI will cause the end of humanity. lol I think AI is mostly hype considering the current hallucination factor. Thanks for talking about this timely topic.
I’m not sure your demo of DiVeRSe is very accurate. By adding the 5 exemplars all that was shown is that adding exemplars (aka few shot prompting) can improve model performance. However the brilliance of DiVeRSe comes from trying multiple different ways of wording the prompts to encourage different reasoning paths and then using a voting verifier to determine which reasoning paths are most likely to lead to the correct answer. For example, if one of the reasoning paths the model took was Alexis spent a total of $30 + $46 + $38 + $11 + $18 = $165 then this part of the reasoning would be identified by the voting verified as incorrect because the math in this *step* is wrong. (Even though it may still somehow arrive at a correct conclusion). The idea is that if all the reasoning steps are correct then it is more likely that the final answer is correct. You can then take a majority vote (self consistency) of the answers that have correct reasoning paths and you are much more likely to arrive at a correct answer overall.
You are right! I tried to simply the example to show one step of this Method. Same with AMA, but your explanation is brilliant. Thank you for sharing this with me!
Very interesting. Prompting is key, of course! In order to get accuracy and self avaluation, I've been trying to implement the "Tree of Thoughts", but I got stuck cause the limit time response for the chatgpt api in google sheets is just 30 seconds.
I think it would be best if language modules were taught the same way people learn in school. First, they should learn the school textbooks. After that, professional books, encyclopedias, scientific books, etc... They go through universities the same way as we do. If you already have this foundation, you can now release it to the Internet to continue learning there. But by then you will have a basic knowledge of the world. For example, if you see a news story in which the sky is red, you will know that it is not normal because it should be blue. But unfortunately, we humans don't write news, articles, or blogs about the fact that the sky is blue. Only if something is out of the ordinary.