【by Enkaryon Ang, Dec. 2023】
For some artificial intelligence researchers and technicians, one exciting aspect of ChatGPT is not only its demonstration of novel understanding capabilities, but also its potential role in building more general artificial intelligence abilities across contexts with language components, which may even be what some people call artificial general intelligence (AGI). ChatGPT has had huge impact in a short time and sparked many discussions. ChatGPT has made text workers nervous, especially because the effect of automatically generating articles has aroused tension in the world. But what kind of new technological situation does it reflect? And do we have to be anxious all the time?
The advancement of algorithms has enabled people to seemingly possess another sense. The recent invention of Paragraphica seems to illustrate the relationship between algorithms and our senses.[1] Paragraphica is a special camera that does not capture real images but uses artificial intelligence to create “photos” based on location data. It gives you a text description based on where you are, the time, the weather, and the surrounding scenery, letting you feel the atmosphere of the moment.
The designer of this camera is Bjørn Karmann from the Netherlands, who wants to explore how artificial intelligence understands our world and how we interact with images, social media, and algorithms. “The camera allows us to experience the surrounding environment in a way that goes beyond visual perception.”
The possibility of such a thing is due to the emergence of a “prompt” algorithm and giant database. First of all, prompts can guide the behavior of AI, allowing AI to perform specific tasks, such as answering questions, generating text, drawing images, etc. Different prompts can produce different results. To this day, the close fit between many effective databases and algorithms makes Paragraphica possible.
To some extent, Karmann’s camera mimics the perception of star-nosed moles, which do not use light but use tentacles to “see” things. The camera lens also has a similar tentacle design. Although the algorithm is not necessarily close to perfect, the emergence of Paragraphica reminds us of a daily habit – the process of deciding the itinerary based on the weather forecast. At first, the weather forecast was inaccurate, but with the advancement of mathematical simulation and the rise of accuracy, it became a part of people’s decision-making. From this perspective, Paragraphica is not particularly surprising, and it can even be said that we got this surprise from the weather forecast. From Karmann’s example, this also implies the indispensability of predictability, algorithm and database size.
Machine learning is rather than “intelligence”. It is the principle of neural networks based on random learning (a variant of probability theory) and statistical models, which focuses on the probability of correlation between A and B; neural networks can identify correlation, but cannot determine causality. Using a large language database assumes that if we can cover all the characteristics of language, it will provide us with the weather forecast of language.
This kind of relationship reminds me of the novel “Onmyoji” by Japanese writer Yumemakura Baku, the second chapter of the first part, “The Woman of Gardenia.” For example, the “prompt” related activities that have emerged in the past two years are like the existence of speech spirit in the original novel. In the novel, this speech spirit comes from the word “likeness” (如 ju) in Heart Sutra (心經 Xin jing). “The same also applies to feelings, perceptions, volitions and consciousness” (受想行識亦復如是 Shou xiang hang shi yi fu ru shi). This is also the shock that ChatGPT or Midjournal brought to contemporary society. From this word “likeness” (如). Shuowen Jiezi (說文解字) said 如 is the same as the word “follow” (隨). One says if also, same also. This ambiguous feature of ChatGPT brings its market-shocking power.
However, as linguists such as Chomsky point out that this is not a real language.[2] Chomsky believes that large language models (LLMs) can only simulate the surface structure of real language but cannot capture the deep structure and meaning of language. The creativity and universality of language, because it can only learn from limited data, is difficult to obtain directly. A recent German novel, “Berlin, Miami” published by Hannes Bajohr with ChatGPT also pointed out similar problems.[3] The plot of this novel does not progress and is full of strange jumps. Some of his language sounds like it was translated from a foreign language by someone who has never spoken German, using a dictionary, grammatically correct, but without a sense of word choice or idioms. Instead, we learned from the ancient classics that this is nothing more than a haunting of similarity and even added our own delusions.
Regarding recent media thinking, Dutch theorist Florian Cramer is worth referring to. Cramer opposes the general usage of digital. According to the more precise technical definition of the word “digital,” digital information does not necessarily have to be encoded as zeros and ones, nor does it require any type of computing device (whether electronic or non-electronic) to process. On the contrary, “digital” more broadly refers to […] divided into (a) clearly countable units (b) any type of information derived from a limited symbol library.[4]
Here, “any type of information derived from a limited symbol library” deserves our re-evaluation, as advanced algorithm technology intervenes in life. Or, we are actually stuck in the problem of infinite pursuit of databases today. This is about our unfinished pursuit of authenticity. In the eyes of engineers, this kind of “creating reality” is called “illusion” in artificial intelligence terms. Sometimes this is desirable, and it is the intended use of the program, such as in AI for generating images, such as Dall-E. But in other cases, it leads to ChatGPT madly looking for connections, especially when people know very little about a topic, finding two facts is enough to establish a causal connection. And the cause of this illusion is the reality that the LLM is not big enough when they claim it is.
On June 15, 2023, Hito Steyerl published “Mean Images” in New Left Review.[5] This article discusses how these synthetic images are both a threat and how they act as the statistical average of their constituent data sets, hindering real creativity. To some extent, we return to Cramer’s basic question about digital; either way we are trapped in the constraint of a limited symbol library, which is also about our relationship with accurate prediction. More radically, in the unfinished symbol library restriction, reality is the final result of labor. “The hidden layers of neural networks also hide the reality of human labour, as well as the absurdity of the tasks performed,” Hito Steyerl said.
The speech spirit in Yumemakura Baku’s novel “Onmyoji” is ultimately a dream. In the era of ChatGPT, “similarity” is a trap. Whether in the invalid approximation of database size or in the participation of labor, it is the outcome of an automatic relationship. This “similarity” implies the change of virtual in this era – a new era of participation, the emergence of prompts. And you are the only one speech spirit, self-help.
AUTHOR
Enkaryon Ang, Independent Researcher, Poet, and Critic, Taiwan.
NOTES
[1] Sejal Sharma. “This lensless AI camera uses only textual prompts to ‘take’ a photo.” Interesting Engineering. June 05, 2023.
[2] Noam Chomsky. “The False Promise of ChatGPT,” The New York Times. March 8, 2023.
[3] Harald Staun. “Das kommt dabei heraus, wenn KI einen Roman schreiben soll.” Frankfurter Allgemeine Zeitung. November 5, 2023.
[4] Florian Cramer. “Digital Code and Literary Text.” Journal für Kunst und Kultur digitaler Medien. October 22, 2001.
[5] Hito Steyerl, “Mean Images.” New Left Review 140/141, March–June, 2023.