The invention of machines offering functions assumed to be expressions of human intelligence has a long history. For instance, the ability to perform arithmetic operations was assumed to be an expression of human intelligence. The invention of the world's first automatic calculating machine [5] by Blaise Pascal, known as the Pascaline, had a significant impact on society during the 17th century. While the exact reception varied among different segments of society, the overall response was positive and marked an important milestone in the development of mechanical calculators. The term Artificial Intelligence, of course, did not exist at that time but could be described retrospectively to capture the nature of that innovation at the time. The Pascaline did not immediately revolutionize society or replace human calculators, as its cost and limited functionality restricted its widespread adoption. Technology evolved and gradually the performance of machines exceeded that of humans. Machines started to beat humans in Chess [3], ten years later Go [11] and many other human abilities. As the performance of computers is increasing exponentially and neural nets are being trained on previously unimaginable large data sets, Artificial Intelligence is evolving at an alarming speed. Ray Kurzweil called the point at which Artificial Intelligence would exceed Human Intelligence the Singularity [6].
The Singularity may not be here just yet but developments such ChatGPT make it quite clear that the impact of technology on society is much more profound and that technology is moving at previously unimaginable speeds. Unlike with Blaise Pascal's calculating machine this technology is affordable and not just relevant to the intellectual elite. In other words, it is extremely likely for AI to transform the very fabric of society. Some time ago, in his book "Program or be Programmed" Rushkopf [9], defined a critical ultimatum underlining the urgency of Digital fluency [2]. Essentially, if one is not willing to take control of technology, i.e., "to program", then the machines will take over and "program" humans. While his position may be extreme it illustrates possible consequences of ignoring technological development. How can society coexist with seemingly superior technology where machine intelligence exceeds human intelligence? Going back to the examples of games we can observe two fundamentally opposite strategies to deal with the looming Singularity: capitulation and convergence. Lee Sedol, after losing against Google's AlphaGo decided to capitulate. He saw no more value in further pursuing this career as a Go player. After all, in the context of
Computational Thinking (CT) is an iterative process synergizing human abilities with computer affordances. The field of Computer Science education research [10] is still debating what exactly CT really is [7] and how it is different from Computer Science and programming. However, there are some common characteristics describing the CT process. Figure 1 shows the original three "A" CT process [8]consisting of the abstraction, automation and analysis stages. This figure is being shared widely, including in Wikipedia.
1. Abstraction: Problem Formulation: Problem formulation attempts to conceptualize a problem verbally, e.g., by trying to formulate a question such as "How does a mudslide work?," or through visual thinking, e.g., by drawing a diagram identifying objects and relationships.
The solution needs to be expressed in a non-ambiguous way so that the computer can carry it out. Computer programming enables this expression. The rule in Figure 1 expresses a simple model of gravity: if there is nothing below a mud particle it will drop down. The Yin and Yang in the center of the Computational Thinking process (Figure 1) describe the two complementary forces which make up the synergy between human abilities and computer affordances. There are CT models such CT unplugged [1] exploring Computational Thinking without the use of computers. However, for our purposes to explore the role of AI we are exclusively interested in use cases where we do have computers.
The Computational Thinking process can now be reconceptualized as proompting (according to the Urban Dictionary this term has been introduced informally into the English language). By proompting we refer to the process of writing or modifying a prompt to an AI system such as ChatGPT or Midjourney. Lets now revisit the 3 stages of CT from the Proompting angle with an explicit eye on the role of humans and computers:
1. Abstraction: Problem Formulation: again, the human drives the CT process by starting with a question. We use the same question from before "How does a mudslide work?" 2. Automation: Solution Expression: Unlike with our previous example the user is not planning to program a simulation of a mudslide but is planning to use one or more AI systems to create some kind of output. Users may have different formats of output in mind. Do they want text, images, videos, audio, code or perhaps even some combination? The user requirements combined with pragmatic concerns such as previous experience will determine which AI system users will use and how they will formulate a promising prompt. The initiative here is still human. Users need to select, and combine AI systems. They need to express a prompt. 3. Analysis: Execution & Evaluation. An AI system may create one or more responses which users will have to evaluate. The analysis stage may be quite simple such as picking an interesting looking picture produced by an AI system such as DALL-E, or it may be quite intricate such as reading and perhaps even running code produced. The connection between input (prompt) and output may be difficult to understand and in some cases even unpredictable. For this stage to be productive humans will need to have some understanding of the various affordances of the different AI systems.
A single iteration is unlikely to produce the desired result. Users may have to adjust their problem formulation (abstraction), change the expression of their prompt (automation), or further process the output of the AI (analysis). That is, they will engage in proompting. In the next sections we present some proompting examples.
Here are some basic examples of proompting.
The user is an elementary school teacher trying to teach the concept of mudslides to 4th graders in a STEM class. She starts the CT process by formulating the original question from Figure 1: "How does a mudslide work?".
Automation: Moving on to automation she picks ChatGPT 3.5 and uses the main question 1:1 as prompt "How does a mudslide work?"
The output is elaborate text introducing the general notion of a mudslide A mudslide, also known as a debris flow or a landslide, occurs when a mass of soil, rock, and water rapidly moves down a slope. It is typically triggered by heavy rainfall, snowmelt, earthquakes, or volcanic activity.
Here's a general explanation of how a mudslide works:
A bullets list with 6 points including "Triggering event", "Saturation and liquefaction", "Failure of slope material", "Movement", "Expansion and destruction", and "Deposition" follows. After a brief analysis, the teacher decides that the output is too complex for 4th graders.
Automation 2: She keeps the abstraction but expands the prompt: "please simplify for an 8 year old and be brief"
Sure! A mudslide is like a big, heavy pile of wet dirt and rocks that suddenly moves down a hill very quickly. It happens when there is a lot of rain or water, and the ground gets too weak and slippery. The dirt and rocks can destroy things in their path like trees, houses, and roads. After it stops moving, it leaves behind a big mess. It's important to be careful and stay away from areas where mudslides can happen.
Automation 3: This is a good start but to get a more hands on experience for her students the teacher would like to see a simulation. She proompts: "write a particle simulation of a mudslide using JavaScript" She got a short JavaScript program from ChatGPT including some explanation of what it does "The particles are randomly generated with different velocities and sizes, and they fall from the top of the canvas to the bottom, simulating the movement of a mudslide." She knows how to copy the program into the browser. The code actually runs but there is no evidence of a mudslide. The teacher does know a little bit of JavaScript but is not able nor does she have the time to fix the code.
Somewhat frustrated, the teacher wishes to at least draw a picture of a mudslide. Using DALL-E via the BING interface she enters the "mudslide" prompt only to be greeted by a "This prompt has been blocked. Our system flagged this prompt because it may conflict with our content policy" BING message.
Alan Kay's 50+ year old vision of educational computing devices for children was called DynaBook [4] depicting Jimmy and Beth using hand held devices including networking and programming: "Two nine-yearolds were lying on the grass of a park near their home, their DynaBooks hooked together." With more than 20 iterations within the Computational Thinking process (Figure 1), Proompting for the image to the right was quite involved. The prompt used for Figure 2 right was "two 9 year old kids collaborating with each other using iPads, sitting outside in the grass, cartoon style." Without the "cartoon style" part DALL-E (via MS Bing) flat out refused to create an image with the warning "Unsafe image content detected." The same was not true when using DALL-E directly. It is not clear why this image was considered unsafe.
The challenge suggested by this example is a high degree of unpredictability inherent to proompting. That is, the precise mapping from step 2 in Figure 1 (automation) to step 3 (analysis) can be difficult to comprehend because of the unforeseeable nature of many of the AI tools. Especially for the unexperienced users the result can be quite a surprise.
The third experiment is trying to use DALL-E to replicate the work of OP-Art artist Bridget Riley, who created the work "Shiver" in 1964 (Figure 3). In this experiment proompting consist of abstracting the geometrical principles of the original picture (Figure 1, step 1), automating the image by formulating prompts (step 2), analyzing the image generated (step 3) and repeating the CT process (steps 1 -3) as many times as necessary.
The work consists of a simple matrix of 31 x 31 grid squares. In each white field is a black triangle. From row to row, the triangles are arranged and twisted slightly differently. The first and last rows are identical horizontally and vertically. The AI generates a matrix containing triangles. However, it does not follow the more specific description. The program thus generates a better matrix by additionally specifying that the first triangles point downwards and the rotation can be to the right or to the left. The first result is somewhat better, but far from satisfactory. The other results move away from the target again and are partly very questionable. In the last image, for example, numbers and letters appear. Maybe we have to tell the AI to stay as close as possible to our description.
Overall our experiences show that AI holds great promise but, in its current state, can also be the source of great frustration. At a time where AI as an end-user technology is exploding it will become nearly impossible for users to track the rampant evolution of AI tools. It will take AI tools to be able to use AI tools. Already extensions to AI tools enable the proliferation of next-generation meta-AI tools bundling up several AI tools into new services. The mere speed of this process is raising big concerns on multiple levels. Education will be profoundly transformed in good and bad ways. In this kind of world with so many quickly moving parts the idea that "Proompting is Computational Thinking" may provide some calm by offering some resilient principles. It will be necessary to establish common principles to all these systems. Here are some early principles that we have noticed:
1. Proompting is Computational Thinking is a universal framework offering some resilient principles that are valid across different AI tools. 2. Proompting is a social process: The opaqueness of most AI tools makes it difficult to predict the output resulting from a prompt. The community of users showcasing examples and sharing suggestions is essential to work productively with AI tools. Some AI systems such as Midjourney include this kind of social embedding. 3. The unpredictable nature of prompting may require reverse Computational Thinking. Some AI tools offer early support to reverse the Computational Thinking process. For instance, users can provide existing images to make Midjourney describe that image as a prompt. In other words, these tools make the proompting processes more transparent by providing a path back from Analysis (step 3) to Automation (step 2). 4. Competence models need to shift from answering questions to posing them. Traditionally the competency of individuals is judged by their abilities to answer questions. For instance, can you label the various parts of the human body in pictures of anatomy? Emerging competency models will be about the skills to pose questions and to modify them. In other words, new models will be about proompting.