“My words fly up, my thoughts remain below...”
my thoughts remain below...”
Musings on the Unknowables of Large Language Models
Faraz Forghan Parast
ChatGPT, a language model developed by OpenAI, is almost everywhere these days, demonstrating an uncanny ability to generate human-like text. Its sophisticated algorithms and extensive training on diverse datasets give ChatGPT the potential to revolutionize various fields, including digital humanities, electronic cultures, and texts.
On the 16th of May, 2023, the event hosted by the Electronic Textual Cultures Lab (ETCL) at the University of Victoria featured a thought-provoking talk by Dr. Bernardo Bueno, titled "Navigating Digital Writing and AI: An Optimistic and Slightly Apocalyptic Examination". Dr. Bueno's insights into the ethical implications of using AI in literature, including issues of copyright and the potential disappearance of human writers, sparked a wave of dialogue and inspired the creation of this essay.
As a grateful guest at this event, I was deeply moved by the discussions and felt compelled to further explore their implications in digital humanities. This covers some of my musings on Dr. Bueno's talk, particularly his remark on how "we don't know exactly how it works yet," and aims to look deeper into the workings of ChatGPT and Large Language Models (LLMs), their unknowns, and the ethical considerations they bring forth.
It may also serve to open a small aperçu to the digital humanities community to take an active role in shaping the use and development of LLMs like ChatGPT. Such involvement could involve conducting further research into the ethical implications of AI, developing educational resources to increase understanding of LLMs, and fostering a dialogue about how these tools can be developed and used responsibly and effectively in the field of digital humanities. The journey into the era of AI-assisted writing is just beginning, and institutions like the ETCL have a bold banner to bear aloft in guiding its course.
Ineffability, Inscrutability or Complexity?
We hear that it is not clear how ChatGPT works - and this is, in the case of something so ubiquitous and explosive, rightfully a cause for uneasiness.
From a purely technical perspective, an LLM is essentially a colossal calculator that transforms text into integers (tokens, in NLP parlance), performs calculations, and then converts the resulting integers back into text. The core of these calculations is performed by an extensive artificial neural network, which consists of layers of neurons arranged in a lattice-like structure. The first layer, known as the input layer, is where the tokenized input is inserted.
Following the input layer are numerous hidden layers, which form the brain of the LLM. In the case of ChatGPT, there might be as many as 96 layers, interconnected in various ways. These hidden layer neurons, also referred to as parameters, each contain a weight.
They take some input, perform a calculation using the weight, and then send the result to all its upstream connections.
Despite the simplicity of this concept, the resulting complexity is so vast that it often leaves researchers unsure of how neural networks function - the "black box" effect that gives rise to unexpected outputs and emergent behaviors.
We do understand the technical aspects of the construction and operation of neural networks. They are composed of layers of interconnected nodes or "neurons," and they "learn" patterns in data by adjusting the weights of these connections based on the error of their predictions. In the case of ChatGPT, a transformer neural network is used, which is particularly good at handling sequential data like text.
To put the engineering genius and technical finesse involved into perspective, consider these facts: ChatGPT is built on a neural network with 175 billion weights that determine word probabilities. The weights in the neural network are adjusted through training using examples and statistical calculations, to minimize loss and reproduce examples appropriately, to generate text that sounds like humans. Word embeddings in a largescale matrix are used to represent words in a 12,288-dimensional space based on their statistical relationships.
The figures continue to become more impressive when we consider that ChatGPT was trained on 570 Gigabytes of text, which is the equivalent of 5.7 billion pages of text. If a human could read 1000 pages a day, it would take them 4000 years to go through this veritable mine of knowledge. The training cost an estimated $12 million in compute resources.
However, while we understand these components, how to build and connect them and their mechanics and mechanisms, the complexity of the networks involved makes it difficult to understand why they make specific decisions. For example, when ChatGPT generates a piece of text, it's challenging to pinpoint exactly why it chose a particular word or phrase. This is due to the high dimensionality of the network and the fact that the decision is influenced by the complex interplay of thousands or even millions of neuron weights.
Moreover, the training process of these models involves a large amount of randomness. The initial weights of the connections are usually set randomly, and the order in which training examples are presented to the model can also be random. This means that even with the same training data and network architecture, two models may end up behaving differently.
So, when it is said "we don't know exactly how it works," there are at least two aspects to consider as the root of this discomfort: unpredictability and a lack of interpretability - which just means that despite our knowledge of how neural networks learn from data and make decisions, the specific reasons behind a particular decision made by a complex model like ChatGPT can be difficult to decipher. This is an active area of research in AI, with ongoing efforts to develop methods for making these models more understandable and explainable.
Of course, it could be argued that performing unfathomably complex calculations and processing is indeed the raison d'etre of creating a neural network or an AI in the first place. We are, essentially, offloading the in vivo, wetware-based cognition in our own collective heads to the in silico processing resources of connected computers we have built.
We may observe, for instance, that we do not exactly know what goes on inside our own minds, or why a certain thought or memory or idea runs through us; this does not preclude our understanding of neurology or neuroanatomy, or what an fMRI machine can show us in terms of oxygenated blood moving to certain brain areas. We can only surmise that evolution chose this system of synaptic connection, neuronal activity, neuroplasticity and cognitive functioning and we are now here because of it. One can imagine that once the complexity of a neural network goes beyond a certain length, it becomes a structure that is able to solve the problem of linguistic production.
Interestingly, there is no other theoretical reason why AI needs to use neural networks other than the fact that they work. And therein lies the opener to the bottomless can of unknowability.
This inherent inscrutability of AI, often likened to a 'black box', presents both challenges and opportunities. We can observe the inputs and outputs, but the internal workings remain largely opaque, contributing to the uneasiness surrounding AI phenomena. This difficulty of comprehending is not a shortcoming or a flaw, but rather an inherent characteristic of these complex systems.
Where does the output of an LLM come from?
The transformation of inputs into outputs in LLMs is a complex process that involves multiple layers of computation. Each layer consists of numerous artificial neurons, which perform simple mathematical operations on their inputs and pass the results to the next layer. The final layer produces the model's output, which is a prediction based on the input.
ChatGPT responds and creates output based on a principle that is both simple and profound. It generates a continuation of text that appears as a reasonable successor to the given input, doing so one token at a time. A token can be as short as one character or as long as one word. For instance, in the sentence "ChatGPT is amazing," let us assume that there are three tokens: ["ChatGPT", " is", " amazing"]. The model calculates the probabilities of all possible next tokens and selects one based on these probabilities. This process is repeated until the desired length of text is generated.
The selection of the next token is not entirely deterministic. An element of randomness, controlled by a parameter known as 'temperature', is introduced to make the generated text more diverse and less predictable. A higher temperature value leads to more diverse outputs, while a lower temperature results in more focused and consistent outputs.
This process is inherently inscrutable due to its high dimensionality and non-linearity. Each neuron operates in a high-dimensional space, where it considers a multitude of factors simultaneously. The relationships between these factors are non-linear, meaning that small changes in the input can lead to large changes in the output, and vice versa. This makes it difficult to understand or predict how a particular input will be transformed into an output.
The transformation process is dynamic and context-dependent. The same input can lead to different outputs depending on the context in which it is used, and the model can adapt its behavior based on the inputs it receives. This adds another layer of complexity and unpredictability to the process.
Moreover, LLMs are stochastic in nature - during training, randomness is introduced in several ways, such as through the initialization of parameters, the shuffling of training data, and the use of dropout and other regularization techniques. During inference and output, randomization occurs through techniques like beam search and temperature scaling, which control the diversity of the model's outputs.
This stochasticity means that even with the same input, the model can produce different outputs on different runs. While this can be beneficial for tasks that require diversity and creativity, it also makes the model's behavior more unpredictable and harder to understand.
Furthermore, LLMs do not have a single, deterministic response to a given input. Instead, they have a distribution of possible responses, each with a certain probability, which makes building “cases” and initial conditions almost impossible and practically futile. Understanding this distribution, and how it is influenced by the model's parameters and the input data, is a major challenge in their study.
How does Training Influence the Model's Behavior?
LLMs like ChatGPT typically start training on a vast copy of the web. The goal is to predict the next word in a certain piece of text from the internet. This machine learning process involves adjusting the weights in the neural net to produce the desired output. The neural net might be shown the first half of a sentence from Wikipedia, and then have to guess the next word.
Initially, the outputs might be erroneous, but the weights of the neurons can be adjusted progressively and iteratively, to minimize the difference between the model's predictions and the actual data to get closer to the real answer. This process, known as gradient descent, can take weeks and cost millions of dollars in compute time, and is essentially a complex optimization problem in a high-dimensional space.
The training process shapes the behavior of LLMs in subtle and complex ways. During training, the model's parameters are adjusted to minimize the difference between the model's outputs and the target outputs. This involves a complex interplay of factors, including the choice of training data, the setting of hyperparameters, and the use of optimization algorithms. The exposure to a plethora of textual data from diverse sources allows the models to extract linguistic patterns and relationships between words, acquiring general language abilities and the capacity to generate contextually coherent text.
The effect of training on the model's behavior, the trajectory of this optimization process, and how it shapes the final model is not fully understood. It is known that different training datasets can lead to models with different behaviors, and that the choice of hyperparameters and optimization algorithms can affect the model's learning process. However, the exact relationship between these factors and the model's behavior is not clear.
It is clear though that the training data is the lifeblood of LLMs. They learn to mimic the patterns and biases in the data, but they do not semantically understand the underlying concepts or principles. This data dependence contributes to the inscrutability of LLMs.
For instance, if an LLM is trained on a dataset that contains biased or inaccurate information, it will learn to reproduce these biases and inaccuracies in its outputs. It cannot distinguish between reliable and unreliable information, or between facts and opinions. It cannot question the validity of the data, or seek out additional information to verify or refute the data.
This data dependence makes it difficult to predict or control the behavior of LLMs. It also raises important ethical and practical questions about the responsibility for the model's outputs, the potential for misuse or abuse, and the need for transparency and accountability. This lack of understanding makes it difficult to predict how changes in the training process will affect the model's behavior. It also makes it difficult to control the model's behavior or to ensure that it behaves in a desired way.
The models do undergo a fine-tuning process to tailor their capabilities for specific tasks. They are exposed to narrower, task-specific datasets where human reviewers follow guidelines to review and rate the generated outputs. This process incorporates techniques like reinforcement learning from human feedback (RLHF), which collects multiple rounds of feedback, ranking, and comparisons to improve the model’s performance iteratively.
Research in understanding and controlling training data influence is ongoing. For instance, techniques like influence functions are being used to understand how individual training examples affect the model's predictions. Other approaches involve carefully designing the training process, such as curriculum learning, where the model is trained on tasks of increasing difficulty, or adversarial training, where the model is trained to be robust to adversarial examples.
The Appearance of Emergent Behaviours
LLMs can exhibit unexpected properties that emerge from the interaction of their many layers and parameters. These emergent properties can lead to creative and useful outputs, but they can also lead to unpredictable and potentially harmful behaviour.
The emergence of unexpected properties is a fundamental aspect of complex systems, and it poses significant challenges for the use and understanding of LLMs. It is difficult to predict or control these emergent properties, and they can lead to outputs that were not explicitly programmed or anticipated.
Research in this area involves both understanding these emergent properties and developing methods to control them. For example, techniques like adversarial testing and formal verification are being used to uncover and understand unexpected behaviors. On the control side, approaches like reward modeling and safe interruptibility are being developed to ensure that the emergent behaviors of the model align with our values and intentions.
Despite these challenges, the emergence of unexpected properties or behaviours also offers opportunities for creativity and innovation. By exploring these emergent properties, we can discover new ways of using LLMs and develop new methods for harnessing their creative potential.
“ It’s Just Adding One Word at a Time ”
Smoke and Mirrors?
The human mind is a marvel of nature, a complex machine that processes, understands, and generates language in a way that is profoundly different from how AI models like ChatGPT operate. The human brain is a biological entity, shaped by millions of years of evolution, and it processes language in a deeply integrated way, involving a complex interplay of cognition, emotion, and social context. When humans use language to communicate, we draw on a rich textile of interwoven threads, of experiences, emotions, and contextual knowledge. We understand its nuances, the subtleties of metaphor, and the emotional weight of words. We can "read between the lines," grasp the meaning behind a sentence, understand the intent of the speaker, and respond in a way that is sensitive to the social and emotional context. We can feel the joy, sorrow, anger, or humor conveyed by language, and we can use language to express our own complex emotions and thoughts.
In contrast, and not to say that ChatGPT is not a remarkable achievement, it is crucial to understand that its apparent grasp of language is more akin to a mirage in the desert of comprehension. It is a testament to the power of machine learning and the richness of the training data that this illusion can be so compelling - but it is seeing something that is not really there.
The illusion is created by the two-step process we spoke of earlier: pre-training and fine-tuning, where the model learns to predict the next word in a sentence by analyzing vast amounts of text data from the internet. This helps it learn grammar, facts about the world, and even some reasoning abilities, but it also exposes it to various biases present in the data. Then, in the fine-tuning phase, as the model is further trained on a narrower dataset that has been generated with the help of human reviewers following specific guidelines, the model generalizes from this feedback to respond to a wide array of inputs and uses statistical patterns in the data it was trained on to generate responses that often seem remarkably human-like.
This distinction between human language processing and AI language processing is crucial for understanding the capabilities and limitations of AI models like ChatGPT. While they can appear to understand language, they are fundamentally different from human minds. They are tools, not people, and while they can be incredibly useful, they should not be mistaken for human-like entities. At the end of the day, it is not difficult to see where the seams come apart, and as miraculous as it appears, the façade of intelligence crumbles.
Understanding: A Rich Constellation of Mental Concepts
Human understanding is a complex process that involves the creation and manipulation of mental concepts. These concepts often contain rich information about the properties of objects, information that can be drawn on for a variety of uses that almost certainly were not originally encoded for. This ability to break out of functional fixedness and apply knowledge in unforeseen circumstances - creativity, if you will - is a hallmark of human cognition.
Our mental representations are often sketchy and incomplete, particularly for things that we have encountered only once and briefly. However, with repeated exposure, we can build up robust detailed representations. These mental representations, whether novel and unfirm or familiar and detailed, appear to be to a large extent purpose-neutral, or at least to contain information beyond that needed for the originally conceived purpose. This adaptability is a significant advantage in problem-solving flexibility.
Linguistic Production
Linguistic production in humans is a dynamic process that involves the use of sensorimotor functions in a covert way. Mental structures that originally evolved for perception or action appear to be co-opted and run "off-line," decoupled from the physical inputs and outputs that were their original purpose, to assist in thinking and knowing. This decoupling allows for the simulation of some aspect of the physical world, as a means of representing information or drawing inferences.
A significant aspect of human linguistic production is the use of mental imagery. This includes not only visual imagery but also auditory and kinesthetic imagery. Imagery is an example of mentally simulating external events. It involves analogue representations that functionally preserve spatial and other properties of the external world, rather than consisting of bundles of propositions.
Working memory is another example of simulating physical events through the off-line use of sensorimotor resources. It involves the temporary storage of verbal and visuospatial information, each of which is coded and maintained in something resembling its surface form. This sensorimotor nature of working memory is evident in the effects of phonological similarity, word length, and articulatory suppression on memory performance.
Language Processing in ChatGPT
ChatGPT, in stark contrast to the rich and dynamic processes of human understanding and linguistic production, operates on a fundamentally different principle. It lacks mental concepts, the ability to simulate physical events, and an understanding of language in the way humans do. It doesn't possess a concept of the world, objects within it, embodied meaning or abstract ideas; nor does it have a coherent sense of self, beliefs, desires, or intentions. The ability to reflect on past experiences or anticipate future ones, form mental images, or use working memory in the way humans do is also absent in ChatGPT.
Despite these fundamental differences, ChatGPT can generate text that appears meaningful and coherent - creating an illusion of understanding, but ultimately remaining a simulacrum of human cognition and linguistic production, devoid of genuine comprehension or creativity. Ironically, it is comparable to the experience of cognitive dissonance by professional translators: A translator who is proficient in two languages but not knowledgeable about the specific subject matter of a text can produce a passable translation by accurately converting the words and syntax from one language to another. However, without understanding the subject matter, the translator might miss nuances, make incorrect assumptions, or fail to convey the intended meaning accurately. The resulting translation might appear correct on the surface, but upon closer examination by someone knowledgeable in the subject matter, inaccuracies and misconceptions may become apparent.
Similarly, ChatGPT can generate text that appears meaningful and coherent, but it does not understand the content in the way a human would. It does not have an understanding of concepts, or a conceptual understanding of the topics it writes about.
While ChatGPT can mimic the surface features of human language production, the underlying processes are fundamentally different. LLMs like ChatGPT have been called Stochastic Parrots - The AI is like a sophisticated parrot that can mimic the sounds of human speech but doesn't understand the meaning of the words it's saying. The same is true for ChatGPT. Its outputs should be seen as the product of statistical patterns in data, not of genuine understanding or intention.
This is where the misuse of ChatGPT and other language models can pose a very real existential threat if such systems are indiscriminately used as "sources of truth" or decision-making mechanisms without informed human overwatch.
These AI systems can diagnose rare medical conditions in milliseconds and pass the bar exam better and faster than any human lawyer, but as they do not actually understand anything at even a fundamental or conceptual level and rely on big data and statistical patterns rather than models of the world, they can make mistakes and even "hallucinate" incorrect information, which can lead to serious consequences if they are integrated into our information ecosystem.
The lack of transparency and absence of conceptual understanding means that humans won't know what specifically is going wrong when misinformation is spread. It can lead to unexpected errors or biases in decision-making, such as in the case of automated hiring systems that might discriminate based on gender or race due to biases in the data used to train the systems. These systems can also have "Achilles heels" - vulnerabilities that can be exploited by hackers or other malicious actors, leading to security breaches or even cyberattacks.
The reliance on big data and deep learning models to approximate the world can lead to a phenomenon called "overfitting", in which the algorithms become too specialized to the specific data they were trained on, and therefore, fail to generalize well to new and unexpected situations. This overfitting issue can manifest in many ways, including in the form of chatbots that produce nonsensical responses or products that fail to satisfy needs, because they are based on narrow and limited data.
We don't Know What We don't Know
With their vast datasets, size and complexity, Large Language Models present a significant challenge in terms of transparency and understanding. They operate in a high-dimensional, nonlinear parameter space, making their behavior difficult to predict or interpret. Their training process is opaque, involving the adjustment of millions or even billions of parameters based on vast amounts of data. This process can lead to emergent behavior, where the model exhibits patterns or capabilities that were not explicitly programmed or anticipated.
Moreover, LLMs lack causal reasoning, meaning they do not understand the underlying cause-and-effect relationships in the data they process. They are also data-dependent, with their behavior and outputs heavily influenced by the data they were trained on. Finally, they exhibit stochasticity, producing different outputs each time they are run with the same input.
These characteristics contribute to the inscrutability of LLMs, making it challenging to understand why they produce the outputs they do. This opacity raises important ethical and practical concerns, as it can lead to unexpected or undesirable behavior, limit the ability to correct errors or biases, and hinder the development of effective governance mechanisms.
Language is, in essence, the medium through which we express our thoughts, emotions, and experiences. As such, it becomes deeply personal, shaped by our life experiences, cultural backgrounds, and the societies we live in. When we communicate with others, we intuitively rely on these shared experiences and mutual understanding for the language to carry the weight of our intended meanings. This dynamic becomes complicated when a machine that uses human-like language, such as an AI, is introduced into the equation.
Consider Rumi's "Lament of the reed;" Mowlana Jalal-eddin Mohammad Rumi, the renowned Iranian poet, opens his magnum opus, "Masnavi-e-Ma‘navi," an extended poem consisting of six volumes of couplets, with this poem:
Hark!
The reed flute cries,
Its sorrow profound,
With tales of estrangement that deeply resound.
Severed from the reed bed
Where once I did grow
My wailing makes tears from men and women flow.
Yearning to rend my breast,
My anguish to show,
The tale of longing's fire, fiercely aglow.
Whoever strays far from their roots and their kin,
Longs to reunite back at their origin
In every assembly, my warbling takes wing,
In tune with the joyous and mournful I sing.
Each finds a companion through what they believe,
The depths of my secrets they do not perceive.
My secret lies near in my sorrowful strain,
But eyes and ears cannot its essence attain.
Body, soul, from each other are never concealed,
Yet the soul's true complexion is never revealed.
The story is expressed in human language, rich with metaphor and emotion. These lines tell of longing, of pain, of the human condition — concepts that are intrinsically human. Yet, when read, each reader brings their own interpretation to the poem, shaped by their unique experiences and understanding.
This principle is not dissimilar to how we chat with AI. As humans, we automatically imbue the words with meaning. We interpret them through our own lens, not the AI's, because that is how we have always communicated. This is where the magic of language can lead to misunderstandings about what the AI is and does.
A prime example is Joseph Weizenbaum's ELIZA, the earliest known "chatterbot" program, particularly the famous case of his secretary. Despite knowing that ELIZA was simply a program, she reportedly asked to be left alone with it — indicating a level of anthropomorphization and emotional engagement. It wasn't that ELIZA understood or provided genuine emotional support; it was that its responses were enough to act as a mirror for human thoughts and emotions, reflecting what looks like understanding, intelligence or cognition.
In much the same way, people might interpret ChatGPT's responses as having a certain "depth," gumption or insight, or to be representative of vast study, knowledge or wisdom. But it's vital to remember that, much like the lamenting reed flute in Rumi's poem, the depth and nuances we perceive are more a reflection of our own human condition than of the AI's capabilities. We are, in essence, the ones breathing life and meaning into its words - symbolic representations of mathematical probability, and reading into it.
The "unknowns" in AI and neural networks, while they present challenges, are not intrinsically negative. In fact, they are a testament to the complexity and sophistication of these technologies, and they open up exciting avenues for research and discovery.
The endeavor to recreate neural structures with computers is indeed rooted in a desire to understand the workings of the human brain, one of the most complex and mysterious systems in existence. We are far from fully understanding how human neural networks work, and the same is true for their artificial counterparts. This parallel between human and artificial neural networks underscores the fact that "unknowability" is not unique to AI; it is a fundamental aspect of our quest to understand complex systems, whether they are biological or artificial.
Moreover, the "unknowability" of AI can be seen as a reflection of the richness and diversity of the patterns that these models can capture. The fact that we can't fully predict or understand what a neural network will learn from a given dataset is, in a way, a testament to the model's ability to learn from data in a flexible, adaptable manner, much like a human would.
This is not to say that we should blindly accept the "black box" nature of AI. Efforts to improve the interpretability and transparency of these models are crucial, particularly when they are used in high-stakes applications. However, we should also recognize that a certain degree of "unknowability" is inevitable when dealing with complex systems, and that this can be a source of creativity and discovery rather than a cause for concern.
In fact, embracing the "unknowns" of AI can lead to new insights and innovations. It can spur us to develop new methods for understanding these models, to explore novel applications, and to engage with important ethical and societal questions. It can also foster a sense of humility and caution, reminding us of the limits of our understanding and the need for responsible, thoughtful use of AI.
In digital humanities, these inscrutable aspects of AI can be particularly fruitful. They can inspire new research questions, drive interdisciplinary collaboration, and enrich our understanding of digital culture. They can also inform the development of educational programs and resources, helping to equip individuals with the critical thinking skills they need to navigate an increasingly AI-augmented world.
Creativity and the unknown share a profound and intricate relationship. At its core, creativity is the act of venturing into the unknown, of exploring uncharted territories and pushing the boundaries of what is known and familiar. It involves generating new ideas, concepts, or perspectives that deviate from established norms and conventions. In this sense, the unknown serves as both the canvas and the muse for creativity.
The unknown fuels creativity by providing a space for imagination and innovation. It invites us to question, to wonder, to dream, and to experiment. It challenges us to look beyond the status quo and to envision new possibilities. The unknown, therefore, is not just a void of uncertainty, but a realm of potential and promise; engaging with it can stimulate creative thinking. It requires us to be open-minded, flexible, and adaptable. It demands that we tolerate ambiguity, embrace uncertainty, and accept the risk of failure. These are all key traits of creative individuals and are essential for creative problem-solving and innovation.
The unknown itself is a crucial identity in the creative process. Creativity often involves a phase of incubation, where ideas are allowed to simmer below the level of conscious awareness. This phase is shrouded in mystery, as it is not fully understood how this process leads to the emergence of creative insights. Yet, it is this dance with the unknown that often leads to the most surprising and innovative ideas.
In the context of AI, the unknown takes on additional layers of significance. AI technologies like ChatGPT can generate outputs that are unpredictable and surprising, even to their developers. This ability to venture into the unknown, to generate ideas that deviate from human expectations, can be a powerful tool for creativity. It can lead to novel and unexpected ideas, narratives, or perspectives that might not have been conceived by human minds alone.
However, this also underscores the importance of a critical and reflective approach to AI-augmented creativity. While AI can help with the venture into the unknown, it does not possess an understanding or awareness of the implications of its outputs. Therefore, the role of the human in the creative process remains crucial, not only in guiding the AI, but also in interpreting, contextualizing, and critically evaluating its outputs.
The Role of Digital Humanities and Directions for the Future
The field of Digital Humanities (DH) and organizations like the ETCL can play a crucial role in addressing these inscrutables. They can contribute to the understanding of these challenges, develop new methods for addressing them, and promote awareness and education about the capabilities and limitations of LLMs.
Digital Humanities, with its interdisciplinary approach, can bring together insights from computer science, linguistics, philosophy, and other fields to shed light into this terra incognita. It can explore the ethical, social, and cultural implications of these challenges, and contribute to the development of responsible and inclusive AI practices.
Interdisciplinary Research
Active DH organizations like the ETCL, with its focus on technology and culture, can serve as a hub for this interdisciplinary research. They can foster collaborations between researchers, practitioners, and policymakers, and facilitate the exchange of ideas and knowledge. They can also engage with the public, disseminating research findings and raising awareness about the opportunities and challenges of LLMs
- Computer scientists and linguists could work together to understand the intricacies of LLMs' language generation capabilities, focusing on how these models understand and generate language.
- Philosophers, ethicists, and AI researchers could dive deeper into the ethical implications of LLMs, exploring questions around consent, safety, and the distribution of knowledge and power.
- Scholars from the humanities could investigate the impact of LLMs on creative writing and literature, examining how these models are changing the way we create and consume literary works - how hybrid human-computer creativity could enter various polysystems, what the creation process would be for the resulting products and artifacts, and which effects could be expected from their functions within local communities and the global society.
Facilitating Knowledge Exchange
Digital Humanities entities can serve as vibrant hubs for the exchange of ideas and knowledge about LLMs. They can host seminars, talks, and discussions that bring together researchers, practitioners, policymakers, and other stakeholders. They can also create online platforms and forums that enable continuous dialogue and collaboration among these diverse actors.
Promoting Educational and Interpretive Projects
DH Research entities can promote projects aimed at making the inscrutable aspects of LLMs more understandable. They can support the development of educational resources that demystify the workings of these models. They can also foster projects that aim to create tools and techniques for interpreting and explaining the behavior of LLMs. Such projects can contribute significantly to enhancing the transparency and interpretability of these models:
- Technical workshops that provide hands-on experience with LLMs, teaching participants how to use these models for various tasks.
- Ethical training sessions that explore the ethical considerations of using LLMs, discussing issues like consent, safety, and the distribution of knowledge and power.
- Creative writing workshops that explore the use of LLMs in creative writing, providing participants with the opportunity to experiment with these tools and reflect on their implications for the creative process - these could end up in also demonstrating the shortcomings and areas where research in NLP is still needed, such as summarization techniques, or the hard problem of text consolidation.
Raising Public Awareness
Entities engaged in Digital Humanities can play a crucial role in raising public awareness about what is not readily understood about LLMs. They can organize public engagement activities, create educational content for the media, and leverage social media platforms to reach a wider audience. They can also engage with policymakers and advocate for policies that promote transparency, accountability, and responsible use of LLMs.
- Public lectures and panel discussions that bring together experts from different fields to discuss the implications of LLMs.
- Educational content for the media that explains the workings of LLMs in an accessible way, helping the wider public understand these complex models.
- Advocacy work with policymakers to promote policies that ensure the responsible and inclusive use of LLMs.
Advocating for Ethical AI Practices
Research entities can take a leading role in advocating for ethical AI practices. They can engage with tech companies, regulators, and policymakers to promote practices that ensure the responsible and inclusive use of LLMs. They can also contribute to the development of guidelines and standards for ethical AI practices. A range of activities to promote openness, transparency and accountability in the use of LLMs could include:
- Research projects that aim to make the workings of LLMs more transparent, developing tools and techniques for interpreting and explaining these models.
- Advocacy work to promote open access to LLMs and their training data, helping to democratize access to these powerful tools.
- Contributions to the development of open standards for ethical AI practices, helping to shape the norms and practices around the use of LLMs.
Digital Humanities and research entities can make significant contributions to understanding and addressing the unknowns and unknowables of LLMs. Through their efforts, they can help steer the development and use of these models in a direction that is responsible, effective, and beneficial to society.