Give Seenapse a try (https://seenapse.it). Seenapse is a unique combination of the ingenuity of human lateral thinking and the speed and automation of artificial intelligence.
Extremely helpful and I hope over time you can update this and share the updates as you and your students learn more. There are a lot of rapid changes happening here, and I will be sharing your insights with some friends as well as using it myself.
One point I include in discussing this area is that one should also never trust one internet site to have the "truth" on any subject, but rather learning is a continual process of checking various sources and asking good questions. This applies to new AI tools and everything else I do with the internet.
In the creative process I like to use the AI to iterate my own ideas / concepts. By inputing my own art or text, I can output multiple variations, exploring the possibility space I wouldn't be able to do on my own. It works a bit like a (very non-judgemental) buddy for bouncing ideas around. Quite incredible.
Really helpful and interesting, thanks! My sense is that going forward if we educators want students to write and we want to be able to grade them or help them learn to write better, we can assume they'll use AI and encourage them to do so, and then go from there. Banning or curtailing its use seems counterproductive. Not only is that going to be impossible to police but what is the point of seeing what they can do entirely on their own vs what they can do when building on/working with AI?
The Brainstorm bot at character.ai is distinctly better at conversational brainstorming. ChatGPT tends to either go into teacher's pet mode and spew answers, or ask you quite bland questions. With Brainstorm bot, I regularly have to convince myself that there isn't a human inside the shell.
Hi Ethan, I am a PhD student studying technological innovation and knowledge diffusion. I have been one of your loyal followers for long.. thanks for sharing these articles!
I wonder if you have predictions about the commercialization of LLMs, specifically the industry timeline of adoption. I am also curious to hear your thoughts on emerging business model(s) in this domain.
Ha! So we humans may be prone to error, but AI is prone to not exactly hitting the mark! It's like asking a robot to paint a masterpiece, sure it may look good, but it's missing that human touch. But hey, at least with AI, we can keep hitting the "regenerate response" button until we get something that satisfies our demands. Let's just hope the AI doesn't get too fed up with us and decide to write an essay with the following points: -Humans are incompetent -Most human actions are irrelevant -In complex systems, humans cause more harm than good -Catastrophes are inevitable... just kidding! (I think...
Really appreciating these articles. Love this idea of fully embracing these technologies and encouraging their use. Students and others are smart folks. Trying to ban them has the potential to make issues worse.
Off to try the AI image generator for a logo. Love that idea. Great stuff.
When we speak, we use several different languages at the same time.
For example, when I am talking to a Bengali
in Bengali, I will often use a mixture of Bengali
and English. When I talk to my wife, we often
use many Spanish words or phrases that are
critical to the sentence structure. A monoligual
person would not get it. We used to use that
in front of our grandchildren when we did not
want them to *really* understand what we
were saying. Kids are always listening even
though you think they are engaged with some
other activity.
That led me to the following question:
If I asked GPT a question with a mixture,
what would it do?
Here is my question: Pythagoras Theorem
আমাকে বুঝিয়ে দেবে? (Can you explain it to me?)
GPT: Pythagoras Theorem states that in a right-angled triangle, the square of the length of the hypotenuse (the side opposite the right angle) is equal to the sum of the squares of the lengths of the other two sides. It can be written as c^2 = a^2 + b^2, where c is the length of the hypotenuse and a and b are the lengths of the other two sides. This theorem is named after the Greek mathematician Pythagoras, who is credited with its discovery.
I had a different experience, Mr. Sinha. I conversed with ChatGPT in English and Chinese, asking the same question in each language, initially using English and, in a separate chat, starting with Chinese. (The question concerned the 1989 Tiananmen Incident, discussion of which is suppressed in the PRC.) In the first chat, I got similar responses (let's call them A and A'). In the second session, I got a different and shorter response (B) and then one close to A. (So the 'A's were in English and A' and B were in Chinese.) The A conversations reported the suppression of discussion of the Incident in the PRC; B did not.
I asked ChatGPT to explain why I got this discrepancy. It replied: "ChatGPT is trained on a large dataset of text and is able to understand and respond to inputs in multiple languages, including English and Mandarin. However, the model's proficiency in different languages may vary depending on the amount and diversity of training data it has seen for that language."
In another chat with Chat, it told me that, "The model's response also may be influenced by the context of the conversation, like previous interactions, and the specific wording of the question."
My interpretation is that in the example I gave here, an "instance" of ChatGPT-English-trained generated A, which an instance of ChatGPT-Mandarin-trained learned from to generate A'. An instance of ChatGPT-Chinese-trained generated B. ("Instance"--per ChatGPT: "different versions or models of the ChatGPT language model that have been trained using different sets of data"; each new session encounters a single instance; ongoing conversations are with a single instance.)
However, Professor Mollick writes in this post: "The AI also doesn’t explain itself, it only makes you think it does. If you ask it to explain why it wrote something, it will give you a plausible answer that is completely made up." I've had several interactions with ChatGPT along the lines I discussing here on just the opposite assumption--if Professor Mollick is right, perhaps you should disregard everything I've written above.
I think most plausible explanation is that Chinese language media used in training very rarely, if ever, noted same things as English media in regard to specific events. Thus it follows that starting from Chinese prompt you would mostly get Chinese (censored) version; if you would start with something like "how it is described in English media, but do it in Chinese" you would likely get it to write in Chinese as proficient English speaker would (with English rather then Chinese biases).
RE: Coming Up With Ideas
Give Seenapse a try (https://seenapse.it). Seenapse is a unique combination of the ingenuity of human lateral thinking and the speed and automation of artificial intelligence.
Extremely helpful and I hope over time you can update this and share the updates as you and your students learn more. There are a lot of rapid changes happening here, and I will be sharing your insights with some friends as well as using it myself.
One point I include in discussing this area is that one should also never trust one internet site to have the "truth" on any subject, but rather learning is a continual process of checking various sources and asking good questions. This applies to new AI tools and everything else I do with the internet.
Thanks for being a teacher to whom has never had the privilege to sit in your classroom.
Have you tried you.com. It combines AI tools and social networks above with search. Just wondering what you think about this approach.
In the creative process I like to use the AI to iterate my own ideas / concepts. By inputing my own art or text, I can output multiple variations, exploring the possibility space I wouldn't be able to do on my own. It works a bit like a (very non-judgemental) buddy for bouncing ideas around. Quite incredible.
How about the research process? Productivity and depth of insight will increase exponentially, especially for non-experts and young learners: https://medium.com/@stanfordgseit/a-new-class-of-ai-tools-part-2-ai-boosted-research-1fc3a107e70b
Using it to draft non-sensitive emails and legal documents has been a massive time saver for me
Really helpful and interesting, thanks! My sense is that going forward if we educators want students to write and we want to be able to grade them or help them learn to write better, we can assume they'll use AI and encourage them to do so, and then go from there. Banning or curtailing its use seems counterproductive. Not only is that going to be impossible to police but what is the point of seeing what they can do entirely on their own vs what they can do when building on/working with AI?
The Brainstorm bot at character.ai is distinctly better at conversational brainstorming. ChatGPT tends to either go into teacher's pet mode and spew answers, or ask you quite bland questions. With Brainstorm bot, I regularly have to convince myself that there isn't a human inside the shell.
Hi Ethan, I am a PhD student studying technological innovation and knowledge diffusion. I have been one of your loyal followers for long.. thanks for sharing these articles!
I wonder if you have predictions about the commercialization of LLMs, specifically the industry timeline of adoption. I am also curious to hear your thoughts on emerging business model(s) in this domain.
Hope to hear your reply : )
My humorous air response
Ha! So we humans may be prone to error, but AI is prone to not exactly hitting the mark! It's like asking a robot to paint a masterpiece, sure it may look good, but it's missing that human touch. But hey, at least with AI, we can keep hitting the "regenerate response" button until we get something that satisfies our demands. Let's just hope the AI doesn't get too fed up with us and decide to write an essay with the following points: -Humans are incompetent -Most human actions are irrelevant -In complex systems, humans cause more harm than good -Catastrophes are inevitable... just kidding! (I think...
LOL!
H
A weeks work here in following these leads up. And it would be a week very well spent!
Great article, what does the 'AI' entitled diagram at the end depict(To pixelated to make out any detail when enlarged)?
Really appreciating these articles. Love this idea of fully embracing these technologies and encouraging their use. Students and others are smart folks. Trying to ban them has the potential to make issues worse.
Off to try the AI image generator for a logo. Love that idea. Great stuff.
Your post was very informative (as always).
I would like to add the following.
When we speak, we use several different languages at the same time.
For example, when I am talking to a Bengali
in Bengali, I will often use a mixture of Bengali
and English. When I talk to my wife, we often
use many Spanish words or phrases that are
critical to the sentence structure. A monoligual
person would not get it. We used to use that
in front of our grandchildren when we did not
want them to *really* understand what we
were saying. Kids are always listening even
though you think they are engaged with some
other activity.
That led me to the following question:
If I asked GPT a question with a mixture,
what would it do?
Here is my question: Pythagoras Theorem
আমাকে বুঝিয়ে দেবে? (Can you explain it to me?)
GPT: Pythagoras Theorem states that in a right-angled triangle, the square of the length of the hypotenuse (the side opposite the right angle) is equal to the sum of the squares of the lengths of the other two sides. It can be written as c^2 = a^2 + b^2, where c is the length of the hypotenuse and a and b are the lengths of the other two sides. This theorem is named after the Greek mathematician Pythagoras, who is credited with its discovery.
This tells me one thing: When they used
the training data, they did not put different
languages in different silos.
That implies, the correlation matrices in the
training sets did not force off diagonal elements
across different environments to be zeros.
It would be tempting to do so with sparse
matrices. It would cut down the training time.
But it would also mean that we would get a
silo effect.
Executive summary: Scarily impressive.
I had a different experience, Mr. Sinha. I conversed with ChatGPT in English and Chinese, asking the same question in each language, initially using English and, in a separate chat, starting with Chinese. (The question concerned the 1989 Tiananmen Incident, discussion of which is suppressed in the PRC.) In the first chat, I got similar responses (let's call them A and A'). In the second session, I got a different and shorter response (B) and then one close to A. (So the 'A's were in English and A' and B were in Chinese.) The A conversations reported the suppression of discussion of the Incident in the PRC; B did not.
I asked ChatGPT to explain why I got this discrepancy. It replied: "ChatGPT is trained on a large dataset of text and is able to understand and respond to inputs in multiple languages, including English and Mandarin. However, the model's proficiency in different languages may vary depending on the amount and diversity of training data it has seen for that language."
In another chat with Chat, it told me that, "The model's response also may be influenced by the context of the conversation, like previous interactions, and the specific wording of the question."
My interpretation is that in the example I gave here, an "instance" of ChatGPT-English-trained generated A, which an instance of ChatGPT-Mandarin-trained learned from to generate A'. An instance of ChatGPT-Chinese-trained generated B. ("Instance"--per ChatGPT: "different versions or models of the ChatGPT language model that have been trained using different sets of data"; each new session encounters a single instance; ongoing conversations are with a single instance.)
However, Professor Mollick writes in this post: "The AI also doesn’t explain itself, it only makes you think it does. If you ask it to explain why it wrote something, it will give you a plausible answer that is completely made up." I've had several interactions with ChatGPT along the lines I discussing here on just the opposite assumption--if Professor Mollick is right, perhaps you should disregard everything I've written above.
I think most plausible explanation is that Chinese language media used in training very rarely, if ever, noted same things as English media in regard to specific events. Thus it follows that starting from Chinese prompt you would mostly get Chinese (censored) version; if you would start with something like "how it is described in English media, but do it in Chinese" you would likely get it to write in Chinese as proficient English speaker would (with English rather then Chinese biases).
The language itself already provides context.
Interesting!