25 Prompt Battles and the Conundrums of Logos

Prompting and Coding

Anyone who has used ChatGPT might have an idea of what prompting is. Disguised as chatbots, the text prompts of Large Language Models (LLM) seem like common sense interfaces through which one can have a conversation with a given LLM robot. After years of chatting on mobile devices, swapping the human on the other end of the channel for an LLM can feel natural in terms of communication paradigms and affordances.

As some might know, swapping a human for a robot (or computing machine) as a communication partner is what Alan Turing suggested in his seminal 1950 article “Computing Machinery and Intelligence.” Turing first conceived a game, dubbed the “imitation game,” where an interrogator remotely asked questions to two subjects (A and B) to try to discern their respective genders. The interrogator and subjects were situated in separate rooms and were only able to communicate textually via teleprinter or teletypewriter (TTY). He then reformulated the gender-guessing game by swapping person A with a computing machine to discern if the machine could “pass” as a person when answering questions from the interrogator, specifying that “the best strategy is to try to provide answers that would naturally be given by a man” (443). A strange game indeed, yet, besides the question of “passing” as a human and one of a specific gender, the configuration he describes, whereby one types questions and receives written answers via TTY, is more or less what conversing with a chatbot entails, albeit with HTML textboxes replacing the TTY interface. The question that is worth asking here is: Is this prompting?

According to the OED, prompting signifies the action of persuading, suggesting, urging, dictating, or reminding someone to do or say something. The questions Turing’s interrogator poses do indeed act like this: “Please write me a sonnet on the subject of the Forth Bridge,” “Add 34957 to 70764,” “Do you play chess?” etc. (442). While this common meaning of prompting can make sense from the interrogator’s perspective, it takes on a very specific meaning when the receiving entity is a computer.

Before the term prompting entered the common generative AI nomenclature, circa 2020, a prompt in computer nomenclature typically denoted (and still does) the symbols displayed on a computer terminal’s command line. When a terminal is opened it usually displays a string of characters (e.g., user@host:~$user at sign host colon tilda dollar sign for a bash shell) that tells the user its user name, the host computer name the terminal is connected to, and the type of “shell” program that will be used to interpret the commands the user types and issues to the connected computer. A command line is composed of the prompt plus a certain command (e.g., user@host:~$ shutdown). As its name suggests, prompting on the command line commands the computer to execute written instructions. Clearly, prompting here does not signify persuading, or suggesting, or urging, or reminding, but strictly means dictating the computer to do something. Prompting here is prescriptive as it commands the computer to execute precisely formulated instructions and report on their results.

For generative AI, prompting means something slightly different. Popularized with the advent of GPT-3.0 (Brown et al. 2020) prompting denotes the act of giving so-called natural language instructions to a generative pretrained (GPT) model that in turn formulates a response without further training or updating of its parameters. Natural language here designates that the formulation of the instructions follows more “natural” expressions in terms of common use of language than the command line instructions described above, whose formulations and semantics are specific and stringent. Prompting “Please write me a sonnet on the subject of the Forth Bridge” to a pretrained language model makes sense, but it does not when written on the command line.

It is important to note that prompting is not training, strictly speaking. Only when a model is trained and configured with adequate parameters can it be prompted with natural language instructions. It might be useful to write mathematical symbols to make some of these notions more concrete and graspable. A model can be written as a conditional probability P(y|x;θ) where x is the input of the model, y is the predicted output, and θ are the parameters of the model. Before the advent of GPT, a model needed to be fully supervised while it was trained, meaning that the correct input-output pair examples (x,y) had to be manually labeled for the model to learn which parameters θ produce the optimal conditional probability. With the advent of GPT, models are pretrained unsupervised on large data sets and can predict the probability of text/token occurrence following observed text (e.g., “My last ___ is Gauthier”). One can manually “fine-tune” such pretrained models with a subsequent supervised learning process with considerably smaller data sets that are geared toward a specific task (e.g., translation, question and answers, etc.). Fine-tuning a model simply means that the parameters θ have initial base values following an unsupervised learning cycle and are incrementally updated by a supervised learning cycle. The advantage of fine-tuning is that it performs well under certain benchmarks. The drawback of this approach is that labeled input-output pair examples are required for every task the model is fine-tuned to solve (Brown et al. 2020, 3). As with the fully supervised approach, this introduces limitations as it does not scale to the level that unsupervised training does. Besides, fine-tuning does not squarely correspond to the definition of prompting formulated above. Why? Because parameters θ are updated during a semi-supervised training cycle, something prompting does not require.

Following our notation, prompting implies that parameters θ stay fixed. It is the input x that undergoes a rewriting to make it correspond more closely to the inputs the model was pretrained with unsupervised. In other words, a loosely formulated input x is rewritten by a function ${𝑓}_{\mathrm{prompt}}(x)=x\text{′}$ that transforms x into an input-form x′ that a given pretrained model can generate word occurrences with, and in so doing, predict the output y. Based on the work of McCann et al. (2018), Radford et al. (2019) showed how, rather than needing task-specific fine-tuning with input-output pair examples, task descriptions could simply be specified with words in the input. For example, input x “What do Francophones mean when they say: ‘Je t’aime . . . moi non plus’?” could be reformulated as a recognizable task x′ for the model “French: Je t’aime . . . moi non plus English: _______.” With this rewritten input, the pretrained model can fill in the blank with words it knows how to predict the occurrence of and is thus able to make decisive choices. This can simply be written P(x′;θ) where parameters θ are fixed, x′ is the rewritten form of input x, and output y is a generated word sequence with the highest probability. This approach was formalized by Brown et al. (2020) who showed how unsupervised pretrained models can be used for a plethora of tasks, so long as the tasks are properly formulated as inputs, or prompts.

This type or input rewriting, here denotated by ${𝑓}_{\mathrm{prompt}}(x)=x\text{′}$, can be done automatically “under the hood” by AI systems without the interrogator knowing it, or can be made explicit in the interrogator’s own prompt formulation. This is what prompt battles are all about: mastering prompt formulations to get the most out of a given text-to-text, text-to-image, text-to-video, or text-to-song generative AI system. Given a certain theme, an image, or an actual text prompt, contestants formulate advanced prompts that are submitted to generative AI systems to generate media (text, image, video, sound) that is then judged by the audience. The textual prompts that contestants write are usually projected on a screen, so that the audience can see them being formulated live, effectively turning what is typically done privately, in a one-to-one fashion with a generative AI system, into a public display.

On the original website Prompt Battle, one can read: “Do you have what it takes to become a Prompt Engineer?” (n.d.). Since tuning a model shifted from being an expert data set labeling/training problem to a prompt formulation/rewriting one, “prompt engineering” now stands as a type of “engineering” duty whose aim is to formulate optimal prompts for given generative models. I put “engineering” in quotations here since one does not need a university degree in computer engineering to formulate proper prompts. Indeed, learning how to prompt involves a doing—a learning-by-doing—and there is currently a plethora of primers online that are meant to train individuals in various prompt formulations and heuristics. This opens the door of technical positions to those from various backgrounds, including Humanities-educated personnel, who were trained to write eloquent descriptions and master complex sentence formulations. As one of the cofounders of OpenAI, Andrej Karpathy, ironically declared in a tweet: “The hottest new programming language is English” (2023).

This statement somehow joins two worlds that were not always made compatible: natural language and computer programming language. This is perhaps why it can sound ironic. In general, computer programming or coding, live or not, is not typically done in natural language. Turing already discussed this incompatibility in his 1950 article, using terms such as “imitation” or “mimicry” to draw a connection between these two worlds. As explained previously, computer terminals do not understand English prose, even if they have a prompt. Compilers and interpreters, used to turn source code into executable formats, do not either. In the introduction of their unusual User’s Manual, Blackwell et al. (2022) give an open-ended definition of live coding (i.e., a form of computer programming), composed of a multitude of statements: “What is live coding? . . . Live coding is about people interacting with the world, and each other, in real time, via code. Live coding is about making software live. . . . Live coding involves a critical orientation toward the otherwise conventionalized work of programming and software engineering. . . . Live coding makes software strange, allowing us to see beyond routine practices and interpretations of code. . . . Live coding involves showing the screen or making visible the coding process as part of a live performance” (2–3). The epitome of live coding events are algoraves: concert-like events where performers write code to program live music or sound. Just like prompt battles, the code performers write live is typically projected on large screens so that the audience can see the code notations/formulations that affect the sound. The similarities between prompt battles and algoraves are quite striking (see Figure 25.1). Both practices turn the usually private practice of writing text/code into a spectacle that directly addresses the audience’s senses through its visual and auditory effects. The striking difference, though, is that code notations of a given live coding language are more “cryptic” than the natural language ones of prompt battles. Even if the notations used to live code might vaguely resemble natural language semantically, the syntax typically does not. One needs to be knowledgeable about such syntactically marked notations to grasp the full meaning of the script that is written live. The question that is worth asking here is: Can live coding be seen as prompting?

The Emperor’s New Clothes

Based on the OED definition of “prompting,” I argued that prompting on the command line does not signify persuading, or suggesting, or urging, or reminding, but strictly means dictating the computer to do something. I will argue here that live coding does the same: It commands the computer to execute instructions that are written live. Prompting here is thus understood as being prescriptive, that is, prescribing the computer to execute something. It might be enticing to believe that prompting a generative AI works otherwise. After all, aren’t new forms of prompts about eloquently persuading or suggesting the AI system to say something or produce images in a certain way? I believe to take this perspective is to enshrine the interrogator’s perspective in our relation to technical beings, to keep them in Turing’s remote room that is only accessible through TTY. Prompt engineering a text is to formulate it into coded instructions that will ultimately dictate the computer to do something, regardless of the encoding scheme or generative system used “under the hood.” I am emphasizing the ultimately here since there is necessarily a (long) chain of translations that transform written prompts into actions performed by executable formats tailored for a given CPU and/or GPU. These formats are “machine readable” and not “human readable,” as they have no visual, let alone linguistic, representations.

As Wendy Chun warns us, “to make the argument that code is automatically executable, the process of execution itself not only must be erased, but source code must also be conflated with its executable version” (2011, 23). This is one of the problems of the interrogator’s perspective merely looking at what comes in and out of the TTY prompt, inducing the feeling that what the interrogator says or writes happens automatically. In conflating written prompts with machinic execution, one conflates logos (words, speaking) with action. This conflation resonates with the ancient poem of Parmenides where speaking (logos) and Being are entrusted (if not conflated) with one another. As Martin Heidegger (1992) observed, the Parmenidean logos is pure alētheia whereby the Being of beings is unconcealed and thus presenced; logos is the ultimate presencing of Being, its very own manifestation. While Heidegger laments our forgetting of the Parmenidean logos and identifies writing technologies as the epitomes of such forgetting (§5.2), I believe this type of conflated logos to be the specter that nurtures and sustains the interrogator’s perspective and warrants the idea that English is nowadays “the hottest new programming language.” Understanding computation through the prism of this ancient logos is what makes software strange, I believe, as it erases the intricate question of the situatedness of technical beings, effectively making them subaltern beings.

While it can be argued that prompting and live coding are not strict mathematical sciences but “genuine crafts” (Blackwell et al. 2022, 217), these learning-by-doing practices are nonetheless interrogator-type practices that solely engage with what comes in and out of the TTY. To truly engage with the technical beings that ultimately execute prompts, one would need to bypass the TTY and enter the remote room where these beings perform computation. Gilbert Simondon formulates a critique of hylomorphism in very similar terms, questioning if it is enough to look at what comes in and out of the workshop of craftsmen molding bricks to grasp what the actual process of brick-forming really entails. He concludes that “it is not enough to enter the workshop and work with the artisan: one should enter the mould itself to follow the operation of form taking at different levels of magnitude of physical reality” (2005, 46, my translation).

Computational operations do indeed take effect at various levels of physical reality. To simply stay put at the prompt is to take an age-old position toward events, a position that Western thought on logos has constructed over centuries, which is there for us to take unquestionably. While text is back and thriving within our current generative AI paradigm, it is time to devise methods and concepts that can help us enter the mould where computation happens and unmask Turing’s person A, so we can start developing a relation with technical beings that goes beyond the subaltern position we have always given them through our linguistic constructs.

References