Starting Over (3)
Speaking through Patterns
This letter continues along the path to re-building courses for the AI era in education.1 We had sketched out a general road-map towards an AI-allied course (▷), and developed specific visions for the future in our last newsletter (▷). I started to write a review of concepts of learning, and educational objectives to propose concrete ways to reimagine our target courses. Then I threw it all out – it was becoming a heap of ad hoc solutions, and it became increasingly unclear how anything of generality could be said. The solution I am introducing in this newsletter feels very different. It began in a similarly exploratory fashion, which had to be discarded, rebuilt, discarded again – but then it began to take form in a somewhat unexpected way. And I think it is at last at a stage where one can talk about it.
Let me introduce you to the idea of a pattern language.
Context
The Sentient Syllabus Project was founded on the realization that generative AI will change everything about education: what we learn, how we learn, what we do, and how we assess – but this also means change will happen at all scales. Questions that arise across different scales are notoriously difficult to address in practical ways, because they require us to think through design problems in multiple hierarchies and systems.
Just think of the discussions around “plagiarism” with ChatGPT. Should we worry that students are going to have a computer do their work? Should we accept that we cannot control who is doing the writing? Should we devise assignments that an AI cannot do well? Should we instead aim to have the learners who trust us with their education become competent users, realizing that this will prepare them better for their future?
Organizing our thinking in a complex domain is hard, even harder when things change in unpredictable ways, all the time.
Systems thinkers since the 1970s have embraced an approach that uses patterns to design solutions for such unsteady scenarios. Complex problems. Dynamic change. Value pluralism. Conflicting objectives. If one needs to work in such domains, no single set of rules, no reliance on authority will get one very far. Good approaches do not command a thing to be in a certain way, but provide ways for things to self-organize according to their own nature. Just like with language: we do not communicate a complex meaning by somehow attaching it to a unique label. We decompose it into pieces of meaning, associate those with words, and let the words self-organize according to their inherent affinities – the rules of grammar. Meaning emerges.
The pattern concept has shaped many domains of design and engineering, most visibly architecture, through Christoper Alexander’s groundbreaking A Pattern Language (1977) – a magnificent book whose lasting influence on our culture can hardly be overstated.
I have worked to adapt this approach to the problem we are trying to solve: rebuild academia while making sure everything stays intact. Focus on invariants. Build solutions that automatically scale with the problems. And be the change, to make sure change happens on our own terms. What has emerged from this thinking is PLAAI:2 a Pattern Language for AI-Augmented Instruction as a responsive way to structure knowledge, and as a practical approach to designing education.
Designing with Patterns
In the words of Christopher Alexander:
[A] pattern describes a problem which occurs over and over again in our environment, and then describes the core of the solution to that problem, in such a way that you can use this solution a million times over, without ever doing it the same way twice (1977, x).
Like words in spoken language, patterns for architectural design populate Alexander’s groundbreaking A Pattern Language (1977). His patterns address issues at scales from urban planning to interior design, from country towns (33 ff.) to front door bench (1121 ff.). But a pattern alone does not make a language, and the real power of the approach does not come from the patterns themselves, but from the fact that they can be extended and combined – just like the expressive power of a natural language is not due to its vocabulary, but due to its syntax that makes countless combinations meaningful. This can’t be emphasized enough because it has often been overlooked when Alexander’s ideas were adopted elsewhere: a pattern language is not just a collection of patterns. A pattern language is a system of meaning-creation based on the combination of patterns.
A goal of the “Pattern Language for AI-Augmented Instruction” (PLAAI) is to become a practical approach that is independent of specific tools, resources, or pedagogies to personalize instruction, increase engagement, and facilitate richer learning experiences. To achieve this, it should naturally allow to create and combine patterns and to build rich constructs, like modules, courses, or entire programs, and it should be straightforward to consider alternatives at every step.
This is achieved by giving PLAAI patterns the structure of an ontology.3
Not to be confused with its philosophical counterpart – which explores the nature of being and existence – the field of information science and knowledge engineering considers an ontology as a structured way of organizing and categorizing concepts, a pragmatic tool for structuring information. The resulting map of terms fosters a shared and common understanding, which is built from explicitly stated relationships between concepts.
The terms in the ontology are our patterns, and the relationships impart a structure on them that is analogous to grammar in human language. Grammar defines how different types of words – nouns, verbs, adjectives – can attach to each other, or can substitute for each other in a way that creates meaning. With PLAAI we achieve something similar by identifying and defining the concepts, and explicitly listing the relationships between them. Thus, the PLAAI ontology is built around educational concepts like lecture, pedagogy, and assessment, these are arranged as hierarchies, but we also define relationships through which patterns can augment each other – like adjectives can modify nouns, or adverbs add meaning to verbs – and through which patterns can substitute for each other – like prepositions can be replaced with different prepositions to construct new relationships. A concept like a lecture for example can be exchanged with a panel discussion or a Socratic dialogue for the purpose of transmitting information.
The resulting potential of this approach is inspiring.
A Pattern Example
So, how should we think of a pattern? It is probably best to start with a worked example: the Lecture pattern (LECTR). Augmenting lectures with AI systems may be a non-obvious choice, after all: what part of us, ourselves, speaking about a topic in our domain of expertise, could the AI usefully replace? But developing exactly this pattern turned out to be quite remarkable in illustrating the transformative potential of collaborating with AI systems. Here is what turned up.
LECTR: Lecture
Definition: Transmission of information about a defined topic through oral interpretation.
Task:
The purpose of the Lecture pattern is to transmit information through a human, oral interpretation. Despite the long tradition and ubiquitous use of lectures, there is significant scope for improvement, including: having learners participate actively, individualizing the pace and the instruction level, and addressing missing prerequisite knowledge.
Details:
It would not be wrong to consider the lecture to be the defining activity of the academy: an authority, proclaiming their views to an audience of learners. There are apparently significant benefits to listening to an other, which engages both verbal and non-verbal channels of communication. Merits that are specific to the lecture format include:
Speech: many learners absorb information more easily from listening to explanations than from reading them.
Pacing: the typical pace of speech is well suited for comprehension.
Scale: the typical length of a lecture matches our attention spans and the amount of information we can pick up at one time.
Responsiveness: during live delivery, a lecture can adapt to learners’ questions or comments.
Focus: lectures focus learners together as a group in a single event, which has an important social function. Through that, they constitute the foundation of material that every learner can be assumed to know.
Practice: much is known about how to lecture well, much has been written about best practice, and many tools exist to support lecture delivery.
These benefits are significant, but each point deserves to be reconsidered, and may be improved.
Some disciplines still have a culture of reading lectures out loud from prepared manuscripts. That is no longer considered best practice. Instructors who do this may be unaware how much their audience resents this mode of presentation; it may be better to simply supply the audience with the manuscript instead and use the speaking time to tell a story – for example: why the lecture contents became important and what significance it has for you personally. Sharing personal perspectives through your own voice, that is where the lecture format remains most convincing. Indeed, all explicit memories from past lectures that I have attended consist of such personal moments.
Consider what we could do. We all have been in the situation of discussing a significant piece of work, a discovery, or cultural artefact. But to examine it in its context requires an investment that we simply can’t properly fit into the limited time of the course. Take the Iliad for example. The climax of book 22, the death of Hector through Achilles’ spear, amid a complex web of mortal grievances, divine interventions, attempts to remain human and yet falling to fate, this all is an enduring foundation of human culture in a profoundly well crafted poetic format. But to discuss it? Would that not require pre-reading the 24 books of Homer, plus a good part of scholarship? Would these kinds of assignments inspire learners to engage more deeply, or turn them off the Classics forever? Or would we work from notes and summaries, stuck in the superficial presentation of storylines? We have another way now. It has become entirely reasonable to take an excerpt from such a text in its original form and present it in a lecture. Of course, the learners would lack context – but the AI can answer questions, and pursue them, and go deeper, to present exactly what a learner needs at the time they need it. This ability to switch between presentation and participation, between taking stock of what is known and exploring what is not, this is a significant departure from the way we used to work with content, and a significant opportunity to bring it to life and make it personal. Above all, we are now much better able to put the learner in control of their learning.
For the lecture format, audience engagement has always been an issue, since lectures are inherently passive modes of information transmission, not active learning. Treating them as opportunities for active involvement instead, and integrating dynamically generated content is a great opportunity. We are not even speaking of moving to full-on discussions as a replacement for lectures. Even a format in which content is partially developed in dialogue with selected audience members, while including the AI in the discourse, would help. If well done, this will allow the audience to engage with the situation vicariously, similar to watching a sports or gaming event, and engage with the material almost without realizing it.
Lectures occupy their allocated time slots fully, can neither be accelerated nor slowed down, and to individualize them in the way we just discussed takes thought. A general solution is to take check-in breaks, in which we switch between static and dynamic content and engage with the audience, to adjust the pace, to give learners an opportunity to catch up on explanations of concepts that might be unfamiliar to them, and to solve other issues with lacking preparation. At the same time, learners may bridge gaps due to a lapse in attention, which might otherwise prevent them from following the rest of the lecture – and it may provide interfaces for personalization. (As an aside, an excellent way to promote continued focus and avoid such gaps is to encourage effective note-taking during lectures.)
Similarly, lectures are given at a specific date, which may not be the right date for self-paced learning. This can be mitigated after the lecture, by AI assisted debriefing and follow-up, at the learner’s pace.
Personalizing the experience supports the crucial social benefits of a joint scholarly effort at an event. Imagine not “giving” a lecture but celebrating it, and develop ideas about framing the lecture that can support a sense of purpose and excitement for the audience.
AI Opportunities:
All activities below are implied to be collaborations with the AI. Many of the suggestions result in artefacts, ideally, these are organized in a knowledge repository.
Preparing the lecture:
Prepare a written script from the raw lecture slides, as a basis for further transformations.
Based on the script, revise lecture outlines to a length of one slide per minute or less.
Review the lecture contents. Ask for an unbiased critique of the lecture content that identifies potential areas of bias or underrepresentation of perspectives, and mentions alternative viewpoints. Ask for suggestions for interdisciplinary connections and real-world problems. Ask an AI with access to the internet about recent discoveries, controversies in the field, or related current events. Predict potential questions or confusion points that students might have, based on the lecture content, so they can be proactively addressed.
If the lecture is based on a manuscript, transform it into a set of keywords for free narration instead. These can be written on index cards to guide the lecture, or appear on the lecture slides.
Condense lengthy text in slides to an appropriate length: 40 words or less.
Generate a list of key concepts and related reading resources that students can review in advance.
Draft an announcement for the lecture, mention required or suggested audience preparation, manage expectations and create anticipation.
Create handouts.
Prepare a self-assessment quiz for students to assess their level of preparedness for the lecture.
Prepare speech-to-text: having a transcribed version of the lecture available will be useful. Apps and services are available, the technology is progressing rapidly. OpenAI's Whisper transcription language model can be run locally in a Jupyter notebook.
During the lecture:
Use check-in breaks every ten minutes or so. This provides opportunities to involve the AI when you realign the lecture with the audience. During the break, ask the AI to summarize where the lecture is at, based on the script, or on the slides that were covered. Or ask it for questions it would like to have clarified: unlike a human audience, the AI will always have something to say. Or ask the AI to summarize the progress of the lecture so far. Then ask the human audience whether this is a correct understanding. This introduces a stimulating, competitive aspect to understanding the contents.
If a real-time transcript is available, the AI can craft self-evaluation questions for students at check-in breaks.
If a chat is available, the AI can summarize the general understanding among the members of the audience, and identify shared problems. We can anticipate that students are more motivated to contribute feedback to the chat if their contributions result in something constructive, or get them noticed. (Chat contributions may be a component of participation marks.) Encourage students specifically to contribute questions.
Periodically ask the AI to anticipate upcoming points in the lecture based on the content covered so far. "What do you think will come next." This will focus the audience's attention on the structure of contents: was the AI right?
Post-lecture:
Have the audience collaborate with the AI to generate a shared set of possible follow-up topics - points of clarification, or unanswered questions - for the next lecture.
Create an FAQ document based on questions that were asked in the chat, or on common areas of confusion in similar topics.
Ask for a summary of the lecture in various formats (e.g., mind maps, bullet points, long-form text, flash cards, study notes) to support different learning styles.
Generate practice problems or application scenarios based on the lecture's key points.
Analyze the chat transcript (if available) to summarize student comments and sentiment, identify frequently asked questions, and note areas of confusion or high engagement.
To support continuous and self-paced learning, set-up a debriefing process that students can engage with at the time that is most suitable for them. Include the lecture script and transcript in a knowledge repository to support this.
Generate individualized homework assignments that are tailored to different learning levels, levels of preparation, and learning styles.
Generate a glossary of important terms and concepts from the lecture to be made available in a knowledge repository.
Implementation:
Treat the lecture as an exciting, inspiring event, and share the excitement with your audience. The use of AI in the preparation and post-processing of lectures follows established patterns of collaboration. Transform the lecture into a script, as a resource to interface with the AI and enable its contributions to the content. Share results in a knowledge repository. Interleave your delivery with check-in breaks, to add a discursive, participatory dimension. And: consider alternatives with a stronger active-learning component.
These are the descriptive components of the lecture pattern. Getting there required solving a number of some organizational and technical issues.
Data and Deployment
Some thought had to be given to the data management of an extensive, complex data product like PLAAI. Patterns should be easily accessible and there should be ways to browse and search. Obviously, the patterns need to be human-readable, and they should be nicely formatted. But they also need to be computer readable and easy to reformat into different views. For both types of views there are several options to weigh: Google docs, Markdown files, or web pages come to mind for the human readable views. The machine representation could be a spreadsheet, like a Google sheet, a custom data grammar, XML markup, or the ubiquitous JSON format, or even an actual database like PostgreSQL, and each option has its own pros and cons. It is crucial that there is one, and only one point of editing the patterns, and one needs to decide which type of asset that should be.
As for the data technology, although the data was initially written into a spreadsheet, we need a richer data format for future extensibility and JSON provides a good compromise between still being human-readable, and playing well with a large ecosystem of data management tools. Thus patterns are defined in individual JSON files and stored in a directory of a GitHub project. They are converted to Markdown files for display as Web pages on the GitHub site. Edits always flow from JSON to Markdown. As well, the reference tree, the dictionary, and other views are produced from the files.
Collaboration
To cite Christopher Alexander once more:
[In certain patterns] we believe that we have made some progress [...] but that with careful work it will certainly be possible to improve on the solution. In these cases, we believe it would be wise for you to treat the pattern with a certain amount of disrespect – and that you seek out variants of the solution which we have given, since there are almost certainly possible ranges of solutions which are not covered by what we have written. [...]
We hope, of course, that many of the people who read, and use this language, will try to improve these patterns [...] and we hope that gradually these more true patterns, which are slowly discovered, as time goes on, will enter a common language, which all of us can share.
You see then that the patterns are very much alive and evolving. In fact, if you like, each pattern may be looked upon as a hypothesis like one of the hypotheses of science.
(1977, xiv f.)
If this was true of a work comprising over 1,200 pages, and written over years, how much more true must this need to explore, to experiment, and to co-develop be, for our education patterns that have been only just conceived? I wholeheartedly agree with Alexander's sentiments – and have designed PLAAI with collaboration in mind.
Collaborators world-wide should be able to participate in the project – and this means decisions need to be made about editorial control, how to acknowledge contributions, how to ensure data integrity, and robust storage, in short, a set of supporting technologies and procedures to ensure this all proceeds in an orderly fashion. The way to go here is to build on the experience of the very large open-source software development community, and follow their approaches. This has been well established on GitHub over years. Github offers storage of software project assets, as well as hosting of associated Web pages that can display documentation – like the PLAAI patterns.
Thus read access to PLAAI assets is simply Web access, patterns will be hyperlinked with each other, for anyone to explore. Resources – prompts, for example – can be hosted there as well.
Edit access and other contributory access will follow the tried-and tested mechanics that GitHub communities have implemented. This requires no particular technical skills, and detailed instructions will follow.
A link to the developing PLAAI white paper is here (▷).
A link to the entry point on the GitHub Website is here (▷).
A link to the GitHub project is here (▷).
I am rapidly adding content so it may be a bit too early to go exploring. The next update will be much sooner.
TL;DR
A pattern language is proposed for the composition of AI-Augmented instructional patterns. Patterns publish best-practice information for AI in educational topics, and support composing the resulting patterns. Data structures have been built, nearly 300 terms have been organized into a hierarchical structure, and assets can be shared via GitHub.💡
References
ALEXANDER, Christopher; ISHIKAWA, Sara and SILVERSTEIN, Murray (1977). A Pattern Language : Towns, Buildings, Construction. New York: Oxford University Press.
GAMMA, Erich; HELM, Richard; JOHNSSON, Ralph and VLISSIDES, John (1995). Design patterns : elements of reusable object-oriented software. CD edition (1998). Reading: Addison-Wesley.
Feedback and requests for interviews, speaking, collaborations or consultations are welcome at sentient.syllabus@gmail.com . Comments are appreciated here.
Sentient Syllabus is a public good collaborative. To receive new posts you can enter your email for a free subscription. If you find the material useful, please share the post on social media, or cite it in your own writing. If you want to do more, paid subscriptions are available. They have no additional privileges, but they help cover the costs.
Cite: Steipe, Boris (2023) “Starting Over (3): Speaking through Patterns”. Sentient Syllabus 2023-05-19 https://sentientsyllabus.substack.com/p/starting-over-3 .
I wish to acknowledge some contributions by ChatGPT (both the GPT-3.5 version of 2023-02-13 and the GPT-4 version 2023-03-14) in response to my prompts, for grammar, expression, and summarization, but also for collaboratively “thinking” through technical issues. I take full responsibility for facticity.
Pronounce PLAAI as “play”, in a homage to Maria Montessori, who saw play as the essence of education, and to Hans-Georg Gadamer who situated play at the centre of his aesthetics, and ultimately epistemology.
Alexander organizes his patterns as a progression from large scale to small scale – from urban planning to interior design (1977). Gamma et al. (1995) categorize their patterns according to roles instead: {creational, structural, behavioural}. PLAAI goes significantly beyond this with its full ontology.