Our Take

Our thoughts on OpenAI’s new features for building education tools for low-resource settings.

News • November 2023

Overview

We’ve been having lots of fun chatting to people about AI in education over the last few weeks, shaping our thoughts, planning what’s next. Then boom – OpenAI throws it open again and release a whole bunch of new cool things (here).

See below for some of our thoughts.

GPT-4 Turbo with 128K Context Window

That’s a lot of tokens, which means a lot of words are allowed for prompts and replies. Now Claude could already get us here, through some magic (though our data scientists say it’s likely sliding window attention – like here https://github.com/mistralai/mistral-src).

The extended context obviously gives more comprehensive interactions – most obviously whole workbooks or teacher guides, or even whole curriculum frameworks. This basically means you can summarise more and quicker, rather than having to cut it into parts which was the previous solution pre-Claude. I see this as a big win for content integration, and summarising, or allowing policy makers to have Q&A with reports.

For our chatbot, one obvious win is it can hold a more in-depth discussion on a subject without losing the thread. This is crucial when providing nuanced explanations, when students or teachers are working through complex problems, or doing reflective dialogue, which requires referring back to earlier points in the discussion.

The big challenge here is that more tokens equal more money, so you have to still be cautious when you use this. I think this is one for the low-frequency work flows, i.e. co-piloting content with specialists.

This is probably why they’ve moved to make them cheaper

The reduced pricing for input and output tokens makes this more affordable for low-resource settings, but nowhere near low enough for child level high frequency use I’m afraid. Nonetheless, it will reduce the costs of the backend work noticeably.

For the developers, one major shift forward is the function calling updates (this took us quite a while to do for our formative assessment app)where to get the output to display on the app, you have to structure it as a JSON (an open file format) which it then shares with the app. If this is wrong, the app breaks and getting that consistently was a pain. Now OpenAI handles that for you!

The New Assistants API – reducing the need for large dev teams

This is exciting and makes the last year’s R&D investments from small firms seem quite foolish. For devs, many of the problems that they’ve built for –notably integrating databases -can now be handled by OpenAI. The accessibility of the code-interpreter, and the retrieval for your own documents now means you can just use their systems to save your education content, rather than build your own databases and optimise. In the short term, this will mean that people are more likely to have some local content in easily, which is a step forward.

I still think the way forward for educationists is to figure out what an optimal information storage and retrieval system is for using localised content in AI, rather than using an off-the-shelf model – something we are doing in the AI-for-Education.org work.

Equally, a major pain was ensuring that the models had ‘memory’ – essentially by repeating stuff back to it, until you hit the limit (now 120k don’t forget, like an elephant). Again, OpenAI now does that for you, reducing entry costs again (but meaning you’re tied into them or need to build if you leave).

Overall – this is exciting, and opens up the space even more. I wonder how many educational products for LMIC will make it into the plug-in store though. I’d love to see a Sierra Leone version of the Code.org plug in, for example.

We’ve gone Multimodal

Having a small child singing about A, B, C always reminds me that learning to read is not just looking at a book. We don’t work that way either. We use our vision, our hearing, speech, and other forms of communications.

Multimodal models enhance the range of possible educational tools, such as interpreting and/or creating textbooks that can combine images with text – like the fascinating work that Taleemabad are exploring. Now Text to Speech is widely available, you can use this to help visually impaired children – though it’s yet to be seen how quickly they can return the speech, or how stable your internet needs to be.

We’ve been playing around with these technologies in different components, but again, putting them together into a package is super helpful to reduce the costs of exploration.

Custom Models and Fine-Tuning

Given how big and good GPT-4 is, I’m not convinced yet that we’re ready to fine-tune it – not to get maximum value from the investment. I think that one of the big wins from AI will be just making the existing content better structured and more machine readable – but we need to do this step before we think about fine-tuning – and if we’re going that far, then the most exciting opportunity is to look at a custom model.

Imagine if GPT-4 knew that learning styles were, well, not recommended 😊 Imagine if it knew what we know about how to help teachers in developing countries. This is possible now it seems – be great for the community to come together to work with OpenAI to do this!

Article by Dr Paul Atherton and Dr Oliver Garrod, Fab Inc.

Author

Dr Paul Atherton

i

CEO & Founder

Fab Inc. & AI-for-Education.org

Explore the Ideas for How AI Can Help With Education

AI has the potential to transform approaches in education, both inside and outside the classroom. Here, we explore some suggestions of where and how AI technology may be used throughout the education system.

We have established 19 areas of the education system where we envision the use of AI and for each of these areas we have described potential appliations. By outlining these applications we hope to frame dialogue within the community and to establish priorities for investment and development.

Sign Up

Join our mailing list to keep up to date with news and events.

Community          Knowledge          FAQ
          Privacy Policy

AI-for-Education.org was set up by Fab Inc. in partnership with Team4Tech. We are grateful to the Bill & Melinda Gates Foundation and the Jacobs Foundation for their support.

Powered by FabData.IO