Existential Questions on AI
In querying or understanding artificial intelligence, my preference was to improve my understanding first, and ask questions later. I veto this approach because it is more interesting to state my thoughts, assumptions and questions as they are today and compare them to their evolution once I have parsed more information*.
*This is the ethos of my project Unified Memory; to process memory bit by bit.
When asking ChatGPT to generate, for example, a list of 500 books featuring Jesus as the main character, you encounter difficulties arising from token management and a lack of true contextual memory. Prompt quality will control how your list is segmented and perhaps stop you requesting such a high number in the first place, but cannot remove core limitations.
An answer you could get from ChatGPT is a list of 31 books segmented by some theme, perhaps movements in literary history. You can ask for a list of 100 and will most likely reduce the instance of repeats. You can ask ChatGPT to proceed with the original task of 500, but it will provide some kind of chunked response of 30-40 in each reply, if you ask it to proceed it will likely generate “fluff” which is to say, make up titles for books it has already included. Where one token is a unit of text (word or component thereof), each response is limited to approximately 4,000 tokens, hence response chunking (you don’t get the full list in one go). This is bad if you want a distinct list of 500 books about Jesus, because ChatGPT can’t cross-check between past and future responses, including what I call ‘synonymous repeats’ where it will simply make up titles for works already cited, presenting them as though they are real, distinct entries. To explain further: something like ‘Jesus of Nazareth’ may be presented again as ‘Son of God from Holy City’, where only ‘Jesus of Nazareth’ is a real title.
Ozempic famously works by mimicking GLP-1, the hormone responsible for communicating with your brain about insulin management, blood sugar regulation, rate of digestion and satiety. Through insulin and satiety management, it works for diabetics (maybe even PCOS sufferers), obese people and less consistently, drug addicts. Isolating a channel of communication with our supercomputer allows us to manipulate it for specific results. I see this as narrow focus enabling depth.
Optical Computer Recognition (OCR) enables data entry via mechanical text recognition resulting in applications like digitizing print books, entering your credit card numbers by scanning them with your camera, testing CAPTCHAs by making them not OCR-able. I see this too as narrow focus enabling depth. Computer Vision in the 1960s focused on replacing visual processing instead of attempting to create artificial general intelligence (AGI). Maybe too wide a focus is not the problem, given language learning models (LLMs) use Natural Language Processing (NLP) in combination with text-to-image generation.
The existential line of questioning here is why was AGI ever a stated goal?
*This is the ethos of my project Unified Memory; to process memory bit by bit.
500 books about Jesus
When asking ChatGPT to generate, for example, a list of 500 books featuring Jesus as the main character, you encounter difficulties arising from token management and a lack of true contextual memory. Prompt quality will control how your list is segmented and perhaps stop you requesting such a high number in the first place, but cannot remove core limitations.
An answer you could get from ChatGPT is a list of 31 books segmented by some theme, perhaps movements in literary history. You can ask for a list of 100 and will most likely reduce the instance of repeats. You can ask ChatGPT to proceed with the original task of 500, but it will provide some kind of chunked response of 30-40 in each reply, if you ask it to proceed it will likely generate “fluff” which is to say, make up titles for books it has already included. Where one token is a unit of text (word or component thereof), each response is limited to approximately 4,000 tokens, hence response chunking (you don’t get the full list in one go). This is bad if you want a distinct list of 500 books about Jesus, because ChatGPT can’t cross-check between past and future responses, including what I call ‘synonymous repeats’ where it will simply make up titles for works already cited, presenting them as though they are real, distinct entries. To explain further: something like ‘Jesus of Nazareth’ may be presented again as ‘Son of God from Holy City’, where only ‘Jesus of Nazareth’ is a real title.
Ozempic & OCR
Ozempic famously works by mimicking GLP-1, the hormone responsible for communicating with your brain about insulin management, blood sugar regulation, rate of digestion and satiety. Through insulin and satiety management, it works for diabetics (maybe even PCOS sufferers), obese people and less consistently, drug addicts. Isolating a channel of communication with our supercomputer allows us to manipulate it for specific results. I see this as narrow focus enabling depth.
Optical Computer Recognition (OCR) enables data entry via mechanical text recognition resulting in applications like digitizing print books, entering your credit card numbers by scanning them with your camera, testing CAPTCHAs by making them not OCR-able. I see this too as narrow focus enabling depth. Computer Vision in the 1960s focused on replacing visual processing instead of attempting to create artificial general intelligence (AGI). Maybe too wide a focus is not the problem, given language learning models (LLMs) use Natural Language Processing (NLP) in combination with text-to-image generation.
The existential line of questioning here is why was AGI ever a stated goal?
Dream big?
As a layperson I question AGI as a vast, badly-defined goal. So is space exploration. Is a vast goal an inherently foolish pursuit? I don’t know. As an artist I know you must explore and invest significant time, accepting waste as part of the process, in order to produce the artefact that finds its purpose. It’s like drafting a text backwards; you draft it then realize it’s what you want to say, instead of trying to say what you want to say. In honesty, I consider this the superior approach to creativity.
It’s unsurprising that OpenAI began to monetize LLMs instead of truly pursuing AGI. ChatGPT’s training uses Azure which claims carbon neutrality since 2012. Does this mean nobody’s hurt by the impact of training, building and using ChatGPT? I’m not naive enough to pretend OpenAI understands the full scale of the environmental and human cost of ChatGPT, let alone that they’re able to be transparent about it.
Maybe I’m a cynic. I care about AI more than well, using the Internet because I have a fuzzy feeling that ‘things are probably balanced’ when using Google Search or hosting my website. In 2022, Mater’s website was redesigned with ASCII decoration to limit environmental impact, but I did nothing to remove the gigantic assets on my personal website then.
I’m not naive enough to think my website has more of an impact than my using ChatGPT, I’m not so learned as to say why specifically. Is it possible, ethical, reasonable to build things we aren’t sure we need against the backdrop of climate disaster? I don’t think the answer is a knee-jerk no, because I see human inventions save human lives. I don’t think the answer is a simple yeah don’t worry about it, because I see my birthplace Lahore covered with smog visible from space, with AQI 1,000. I don’t know why specifically that is happening, just that my country of ethnic origin appears to be suffering extreme weather, extreme conditions more frequently this last decade.
Grimes – My Name Is Dark (Algorithm Mix)
The Role of The Artist?
Perhaps due to ignorance I don’t see LLMs or AI-based applications as more than sophisticated algorithms, so I wonder why we deviate from the term Machine Learning to Artificial Intelligence. Human intelligence is millions of years of evolution and living your lifetime (training the model) resulting in the ability to create something entirely new that did not exist before (art, experiences, new life). I see ChatGPT responses as collages made with pieces of material cut so small as to mimic the new. I see the hypocrisy as an artist in pretending anything is original anyway. Is the distinction that it’s original when it’s human made? Is human involvement enough to term the result original?
I have existential questions about the path taken to the goal. Why AGI and not the replacement of a single component of cognition? Is that not what LLMs are? What is the net difference made in quality of life to humanity by having LLMs so widely available? When you have an AI-generated Coca Cola ad the stakes are an influenced purchase with a limited monetary cost and insulin spike, perhaps over time obesity and appetite dysregulation, the devaluing of human artists. You may have upsides, like adjustments in the kinds of jobs human beings do, perhaps human time allocated to tasks more important than advertisement. Do you buy the latter?
If you have an AI model allocating treatment to patients with limited resources, would you trust their decisions? How do doctors make decisions about allocating life-saving care if they have competing concerns? Do they look at the probability of life spans, the probability of allocating care to the patient who will have highest quality of life? What’s the difference?
OpenAI, Meta, DeepMind and other entities with the ability to conduct large-scale research are subject to money-based society. They are constellations of actors with personal aims, more often than not conflicting with the aims of those external to their organization, in conflict even within themselves as we all are. Nothing would ever get built if we took every opportunity to reinvent the wheel. Nothing should ever get built at too high the cost, but can we measure it, and who decides? And what do I do about this?