This is exactly the sort of thing that teachers should be watching right now. Instead of recycling the same lesson plans they have been using for 15 years, it is time to look at how AI can make their job better, learning more engaging, and really developing their students into people prepared for the world they will join in the coming years. AI is more than a toy and people need to learn how to properly use it. The students are going to know more about this than their teachers and it is going to really put the teachers at a disadvantage.
Thank you Ethan and Lilach for a great video. I was thrilled to see that everything presented goes along well with everything I have been sharing on my RUclips Channel dealing with AI integration in education. AI Literacy is such an imperative that all of academia needs to work to better address it. This video will serve as a great resource to help with this issue.
A helpful video, thank you for doing this. One important correction for you to consider. The statement “They are essentially an auto-completion mechanism on steroids” is problematic. First, generative models feature an analytical component that is not just statistical probability, but instead is about understanding input, for example: summarizing a text effectively is not a probabilistic exercise. Second, there is a knowledge component thanks to training, for example when translating generative AI is able to access knowledge and context to translate more effectively than any previous software. The autocomplete example is more apropos to older, pre-Transformer models: the transformer enables a much richer interaction with the input that is multi-layered and deep. Finally, you correctly point out the risks of bias, which of course is not a function of autocompletion (e.g. the problem is not echo chamber bias): instead the bias in these models is a result of training which is distinct from input.
From Mike Burnham: "When an LLM predicts the next word it’s not looking back into a database of text and picking the word that appeared most frequently in similar contexts. It’s modeling the meaning of a sentence, then looking at its internal model of reality captured in word embeddings and picking the one that makes most semantic sense according to its model. It picks ‘dog’ as the next word because it judges that ‘dog’ makes the most sense according to its semantic understanding of the text. But don’t take my word for it, here’s Geoffrey Hinton and Ilya Sutskever, arguably the two most influential researchers in AI, making the exact same argument.
All fair points, but "auto-completion" is not the best way to describe the behavior of modern LLMs. The semantic and contextual limitations of autocomplete is qualitatively distinct from what modern LLM is capable of; e.g. the orders of magnitude of increased complexity in these models isn't simply a difference in notation, it has created something new. The notion of "auto-completion" suggests simply finishing an existing input, but this would suggest that these models don't create anything new and only effectively perform a single function. Thus, while it is fair to say that these models are probabilistic rather than deterministic, it doesn't follow that everything they are capable of is merely predicting the next word in a sequence. The use of non-linear activation functions and the development of the transformer have created significant changes that make a characterization like "auto-complete" dated. @@6681096
With Chat GPT I have been able to create summarized notes on topics like IFRS to help my students develop level 1 understanding on this vast topic . I hope to do the same for US GAAP and Audit standards . N see how we go from there .
Please consider limiting the background music and the constantly flashing images. It's hard to filter through all that audio-visual input just to hear. I tried muting it and reading captions, but the screen has too much information.
Very helpful; however, as this AI stuff is new to me, I found that the Professors spoke too fast for me to absorb what they were saying. Had to reduce the speed to 0.75x. Slow down a bit please. Information is great!!!!
Great intro but be careful with oversimplifying and generalizing. “Prior to the last couple of years what AI meant was usually about machine learning algorithms and prediction.” -> What about symbolic AI? “All AI models are based on prediction.” -> It could be argued that this is only one of the paradigms.
Ask smart questions based on discussions in class, that way you won't need to "detect" for AI Maybe it's not about catching them "cheat", but rather they're using AI cuz tests ask dumb questions
Language skills certainly, but this has always been true with search also. Linguistics is the study of language[s], you don’t need to study language in order to use a language.
@@qjames Study in that context is a noun referring to field of research, so it can’t be used as an infinite verb (“to Study”). I understood what you meant, but AI gives you a finite amount of token.
This is exactly the sort of thing that teachers should be watching right now. Instead of recycling the same lesson plans they have been using for 15 years, it is time to look at how AI can make their job better, learning more engaging, and really developing their students into people prepared for the world they will join in the coming years. AI is more than a toy and people need to learn how to properly use it. The students are going to know more about this than their teachers and it is going to really put the teachers at a disadvantage.
😊
Way to go! This is EXACTLY what needs to be said and SHARED with faculty across the globe!
Thank you Ethan and Lilach for a great video. I was thrilled to see that everything presented goes along well with everything I have been sharing on my RUclips Channel dealing with AI integration in education. AI Literacy is such an imperative that all of academia needs to work to better address it. This video will serve as a great resource to help with this issue.
One of the more concise and profound takes on the topic that I've come across. Thanks for putting this together.
Such a classy team!
Wonderful, I look forward to the rest of the series!
A helpful video, thank you for doing this. One important correction for you to consider. The statement “They are essentially an auto-completion mechanism on steroids” is problematic. First, generative models feature an analytical component that is not just statistical probability, but instead is about understanding input, for example: summarizing a text effectively is not a probabilistic exercise. Second, there is a knowledge component thanks to training, for example when translating generative AI is able to access knowledge and context to translate more effectively than any previous software. The autocomplete example is more apropos to older, pre-Transformer models: the transformer enables a much richer interaction with the input that is multi-layered and deep. Finally, you correctly point out the risks of bias, which of course is not a function of autocompletion (e.g. the problem is not echo chamber bias): instead the bias in these models is a result of training which is distinct from input.
From Mike Burnham: "When an LLM predicts the next word it’s not looking back into a database of text and picking the word that appeared most frequently in similar contexts. It’s modeling the meaning of a sentence, then looking at its internal model of reality captured in word embeddings and picking the one that makes most semantic sense according to its model. It picks ‘dog’ as the next word because it judges that ‘dog’ makes the most sense according to its semantic understanding of the text. But don’t take my word for it, here’s Geoffrey Hinton and Ilya Sutskever, arguably the two most influential researchers in AI, making the exact same argument.
All fair points, but "auto-completion" is not the best way to describe the behavior of modern LLMs. The semantic and contextual limitations of autocomplete is qualitatively distinct from what modern LLM is capable of; e.g. the orders of magnitude of increased complexity in these models isn't simply a difference in notation, it has created something new. The notion of "auto-completion" suggests simply finishing an existing input, but this would suggest that these models don't create anything new and only effectively perform a single function. Thus, while it is fair to say that these models are probabilistic rather than deterministic, it doesn't follow that everything they are capable of is merely predicting the next word in a sequence. The use of non-linear activation functions and the development of the transformer have created significant changes that make a characterization like "auto-complete" dated. @@6681096
Really helpful. Hopefully will be updated soon.
A gud presentation indeed keep updating us on the advent of AI
This video series is awesome. Could you put these Practical AI videos into a Playlist so they are easier to share?
Thank you! Excellent resource.
With Chat GPT I have been able to create summarized notes on topics like IFRS to help my students develop level 1 understanding on this vast topic . I hope to do the same for US GAAP and Audit standards . N see how we go from there .
Loved the links to the references made in the video, thanks for that!
Thank you!
Very concise take on a complex topic. When will the next video be posted?
Every weekday this week at 11 AM ET 👍
Very informative, thanks
Please consider limiting the background music and the constantly flashing images. It's hard to filter through all that audio-visual input just to hear. I tried muting it and reading captions, but the screen has too much information.
agree with you. he moves through this rapidly. alot of digest.
I am not convinced its all undetectable... yet.
Very helpful; however, as this AI stuff is new to me, I found that the Professors spoke too fast for me to absorb what they were saying. Had to reduce the speed to 0.75x. Slow down a bit please. Information is great!!!!
When is the whole series available? Many thanks
There are five videos in total, with each being released at 11 AM ET this week.
Great intro but be careful with oversimplifying and generalizing.
“Prior to the last couple of years what AI meant was usually about machine learning algorithms and prediction.” -> What about symbolic AI?
“All AI models are based on prediction.” -> It could be argued that this is only one of the paradigms.
take out the music please - other than that, I love the content.
agreed. music isn't helping.
Ask smart questions based on discussions in class, that way you won't need to "detect" for AI
Maybe it's not about catching them "cheat", but rather they're using AI cuz tests ask dumb questions
You need to have a mastery of linguistics to even use it correctly. I don’t think it can help as much as people assume
Language skills certainly, but this has always been true with search also.
Linguistics is the study of language[s], you don’t need to study language in order to use a language.
@@qjames Study in that context is a noun referring to field of research, so it can’t be used as an infinite verb (“to Study”). I understood what you meant, but AI gives you a finite amount of token.
... better start with Lawyers and MPs, doctors, judges and police ..
.. start the normal way .
Thank you!