Artificial Intelligence: When the Real Thing is in Short Supply

The term Artificial Intelligence (AI) has been around for decades. Back in the day, if anything was automated, it was considered “AI”. From early computerized task automation and robotics to autopilot and cruise control, AI was anything an engineer or programmer wanted it to be. 

In 2023 a new level of AI was introduced in the form of an LLM (Large Language Model).  As the acronym may imply, an LLM has three major characteristics: It’s Large, it relates to a Language and incorporates learning Modeling. 

The Large in LLM takes advantage of recent advances in computerized storage in the petabyte (1,000 terabytes) range. With interconnected arrays allowing related storage in the exabytes (1,000 petabytes) range.

The Language function incorporates vast amounts of text data, such as books, articles, and web pages, in order to learn (store) the patterns and structures of a given language.

The Model brings the enormous storage and the language functions together through programming. Large Language Programs and their parameters, however, require an equally large processing capability.  Only recently has this capability been enhanced through neural network architectures and will soon benefit further from quantum computing technologies.

The programming and the large relational database are combined with billions of parameters that can generate human-like text that is difficult to distinguish from text written by humans. This has led to a “self-learning” Q&A type of AI that has only been possible in the last couple of years.

A ‘Perfect’ LLM would be one that is only populated with facts. But if an AI is tasked with ‘talking’ to a human, the language model parameters must be capable of understanding human abstract thought – in the context of a conversation. A conversation among humans is rarely, if ever, based solely on fact – but rather the lack thereof.

So, the current ChatGPT LLM, which is taking society by storm, had its initial language model populated from years of content of an online chat service (Reddit). As it matures, the inputs and reactions from actual human conversation are meant to strengthen its interactive reliability and applicability. 

Reddit was not the only source of ‘information’ however. When I asked ChaptGPT: “Where do you get your facts? Is there a primary source like Wikipedia?” It answered thus:

“As an AI language model, I don't have the ability to browse the web or search for information like humans do. Instead, I have been pre-trained on a large corpus of text data, which includes a wide range of sources such as books, articles, websites, and other textual materials.

"The training data is carefully curated to ensure accuracy, diversity, and relevance, and it comes from a variety of sources, including but not limited to Wikipedia. However, it's important to note that while my training data is designed to be as accurate and reliable as possible, there may still be errors or inaccuracies in some of the information I provide.

"To help ensure the accuracy of my responses, I rely on a combination of statistical patterns and machine learning algorithms, as well as various natural language processing techniques. Additionally, my responses are generated based on the specific wording and context of each question or prompt I receive, so the information I provide may be influenced by the wording or phrasing of the original question.”(1)

Herein lies the rub.

The basis that the AI model has “…been pre-trained on a large corpus of text data...” and “The training data is carefully curated” - means program parameters, curators and trainers are the primary basis of information and interaction. Machine learning takes up the slack – as much as allowed.

So, what if the LLM programmers, curators, and trainers are all Chinese communists? And what if the platform for interaction with the LLM is monitored and censored by the government? Or an Islamic state was to launch and train a similar AI? Or a Jewish state, Or Christian... yada yada…. Humans have considerable biases that they are required or indoctrinated to accept as fact and truth. The LLM they create will contain the same biases and likely the same misinformation.

Reminiscent of the early days of the internet - when all that could come out of a global social network was seen as good and wholesome – the AI revolution could very well create even greater walls between cultures, governments, and the people therein, than ever before.

The internet of today is almost a religion in and of itself. Take it away from the population and you will likely see chaos. AI could be the next tech-religion and another way for those with power to control those without.

Be careful of what you wish for. 

JWB

(1) ChaptGTP 3.5 – March 23, 2023 – Captured & copied from https://chat.openai.com/chat, April 7th, 2023.