I’ve lost track of the original source (post? toot? tweet?), but I recently ran across a note discussing how we, as educators, would be a lot better served by building and training our own Artificial Intelligence systems rather than relying on the generic ones produced by OpenAI (or anyone else).
I also attended a meeting last week where someone shared a piece of software they built to use AI systems to help integrate EDIA into learning outcomes and course outlines. Aside from my concerns about the fact that incorporating EDIA into teaching and learning should be informed and intentional, I was also unconvinced about the system.
The audience was non-technical, so there wasn’t a lot of detail, but from what I can see, the system was making structured queries to the GPT-3.5 API to construct course materials. If the API returned an unacceptable answer, as determined by a “constitution”, the query was run again. Constitution in programming is not a term I’ve seen before, so I certainly have some reading to do, but I assume it’s similar to ChatGPTs second model which prevents it from generating bigoted statement. You always stand the chance of the model never getting an answer you can use, but generally this approach works, even if it is potentially wasteful.
Building a structured system around GPT makes sense. Again, I don’t know exactly what’s in the queries, but the idea that the system should have a purpose that you can give your specific information, feels like a much more useful tool than a chat bot that sounds confident, but gets everything wrong.
I would still prefer an approach that centred actual knowledge, but improved structure would certainly be a step in the right direction.
Beyond improved structure, we should also consider improved training for AI systems used in teaching and learning, especially in higher education. Having an AI that understood what a lab report looked like, or better yet, what a lab report looked like for the course you were teaching would make it a much different tool, than just relying on a generic system which might or might not start your students off with in a meaningful position. If an AI system could concretely mention, “you’ve forgotten your results section” or “there seems to be an error in your calculations”, it would be much more effective as the tutor that some people are seeking.
Another place where I’d like to see us look much more deeply at how systems are trained and on what, is in transcription.
The transcription in Zoom is pretty good. If you’re having a chat it more or less gets the words right. It’s also alright for a lot of standard meetings, but when you start using it for specific topics, it tends not to be able to transcribe more specific terminology. Once we start trying to teach, especially upper level concepts, it often devolves in to utter uselessness.
Now I’m not saying Zoom’s not doing more work to improve the tool, I’m just saying that if you look across a higher education institute, a *lot* of different words that get used. Those words mean things and better yet, depending on your discipline, those words mean *different* things.
I think the answer to that is again to have tools trained to support different discipline, or even different instructors. That way you have the benefits of a live transcription, but also one which is correct, without someone having to spend hours correcting the transcript file.
This is definitely just a starting point. I don’t really know where to go with these systems, nor how much of my time and energy should be directed at them. I do think Tim Bray has the right of it that we can’t expect any generic system to reflect the light we need for teaching and learning.
The data sets that current LLMs are trained on are basically any old shit off the Internet, which means they’re full of intersectionally-abusive language and thinking. Quote: “Feeding AI systems on the world’s beauty, ugliness, and cruelty, but expecting it to reflect only the beauty is a fantasy.”
Tim Bray
https://www.tbray.org/ongoing/When/202x/2023/03/14/Binging
And we do tend to try to avoid giving our students “any old shit” when teaching.
Generally, as educational institutions, we lack the resources to really build the tools we need and I’m honestly terrified of the next generation of EdTech salespeople. I guess for the time being advocacy for critical thinking around these tools is a starting point and more generally pushing for greater control of our software in every context.
Since I started writing this, Arthur Spirling, published a similar call[https://www.nature.com/articles/d41586-023-01295-4] in Nature (so I feel like I’m at least roughly headed in the right direction with this). His call specifically focused on how scientists should engage with open-source AI systems. I *still* think there’s a significant gap in resources and skills to really do this effectively, but it would be great to see some actual open, or public initiatives to help build these tools for the public good.