Perhaps an unpopular opinion in a listserv for educators, but Chat GPT challenges many educational systems that attempt to train students to produce evidence as a proxy for learning. If a public beta toy can produce student grade content then students might be better served through education on how to use it critically, what goes under the hood (see https://beta.openai.com/playground for dials) and how to train their own personalised model so it speaks in their ‘voice’ and ‘style’ (see https://beta.openai.com/docs/guides/fine-tuning ) . Those kinds of activities are much more likely to be enjoyable and motivating for learners than acquiring skills that an AI is likely to do easily in the very near future. When you look at how OpenAI trained the RL model/layer of Chat GPT (see @DrJimFan’s twitter thread here for an explanation that made sense to me: https://twitter.com/DrJimFan/status/1600884299435167745 ) it, after all, does look a lot like what we do in standardised testing – ranking of responses as reward model to ‘encourage’ better response/higher grades in the student cohort.

 

I’d love to see more uptake of the less headline grabbing open source initiatives like https://huggingface.co/docs/transformers/model_doc/gptj & https://www.eleuther.ai/projects/gpt-neo/ for use in classrooms. With the right support, these can provide teachers and educators with school safe AI playgrounds. Having students create their own personal Chat GPT tutor would be both fun and educational. Having them ‘grade’ its outputs would instil the critical literacies needed to flourish in the age of ubiquitous AI. They could learn the basics of human-in-the-loop annotation, how to deliver experience to the machine, what’s good/bad practice, etc…

 

It’s a huge challenge, but I’m optimistic primarily because of the work done by the incredible AI4K12 team and the other contributors in these threads.

 

 

From: xxxxxx@lists.aaai.org <xxxxxx@lists.aaai.org> on behalf of Ed Chen <xxxxxx@nuevaschool.org>
Date: Friday, 9 December 2022 at 6:34 am
To: Dave Touretzky <xxxxxx@cs.cmu.edu>
Cc: xxxxxx@lists.aaai.org <xxxxxx@lists.aaai.org>
Subject: Re: [AI4K12] you must try ChatGPT today

Hi all,

 

Please note several of our students are using ChatGPT to have it write their essays and they are submitting them. 

 

Please feel free to alert your teachers to look for shifts in voice and style in their students' writings.

 

 

 

On Thu, Dec 8, 2022 at 12:25 AM Dave Touretzky <xxxxxx@cs.cmu.edu> wrote:

ChatGPT is a new chatbot released last week by OpenAI based on their
GPT-3.5 large language model.  (Stefania Druga referenced it in her
message about generative AI for kids.)  ChatGPT is astonishing.  It can
hold a conversation, write stories, write code in many different
computer languages, solve analogies, answer questions about hypothetical
situations, compose poetry, and more.  The world has been going nuts
over it, with over 1 million users signed up to play with it.

You have to try it and show it to your students.  Just follow this link:
https://chat.openai.com/chat

Here's a good New York Times article on the phenomenon:
https://www.nytimes.com/2022/12/05/technology/chatgpt-ai-twitter.html

This critical article in the Atlantic advises to "Treat it like a toy,
not a tool."
https://www.theatlantic.com/technology/archive/2022/12/chatgpt-openai-artificial-intelligence-writing-ethics/672386/

I think that's good advice.  While ChatGPT is very adept at holding a
conversation, a lot of what it says is utter nonsense.  Lots of people
have been pointing this out, but here are my own personal examples:

- It claimed that the isotopes of neon were radioactive and listed their
half lives.  In reality, the three major isotopes of neon are all stable
and have no half lives.  It just made up those numbers!

- When I asked it "Who is David Touretzky" it wrote a one-paragraph
biography of me that correctly identified me as a professor of computer
science at Carnegie Mellon, but the educational background and awards it
attributed to me were completely fictitious.

- It insisted that the cube root of -1 was i.  That's obviously wrong,
as i^3 = -i, not -1.  I tried to have an argument with it about this,
but I was unable to fully convince it that it was wrong.

- When I asked it to write Lisp code to generate all permutations of the
letters in the world JACKASS, it wrote code and showed me what the
output should look like, but when I actually ran the code it produced
garbage, not the desired result.

- It insisted that the LamDA language model was created by OpenAI, when
in reality it was created by Google.

Despite its limitations, you can have a lot of fun with ChatGPT.  I
asked it to compose a version of The Three Little Pigs where the
littlest pig was romantically involved with the wolf, and it did so.  It
wasn't a very good story, but the fact that it could do it at all is
astonishing.  We have not seen technology like this before.

The implications of this technology are hard to grasp at this point.
The dialog between humans and machines is about to get a lot more
interesting.

-- Dave

To unsubscribe from this list please go to https://www.simplelists.com

 

--

Ed Chen Director of Technology 650-350-4548 xxxxxx@nuevaschool.org   nuevaschool.org

Hillsborough Campus Grades PreK-8 6565 Skyline Boulevard Hillsborough, CA 94010 San Mateo Campus Grades 9-12 131 East 28th Avenue San Mateo, CA 94403

To unsubscribe from this list please go to https://www.simplelists.com

To unsubscribe from this list please go to https://www.simplelists.com