How AI Works: Predicting Patterns
AI, or artificial intelligence, is a form of software that mimics human thinking. While it is not the same as human intelligence, it can be useful for completing tasks by approximating human thought processes. The field of AI, however, raises many questions about the nature of intelligence and whether it can be artificially created. The concepts behind AI models have been around for decades, but recent advances have made it possible to apply these concepts on a larger scale.
AI models, also known as machine learning models, share a common structure: predicting the most likely next step in a pattern. These models do not possess knowledge but excel at detecting and continuing patterns. This concept can be illustrated by imagining an octopus sitting on a telegraph wire used by two humans to communicate. Despite lacking understanding of language or humanity, the octopus can build a statistical model based on the dots and dashes it detects. Over time, it learns various patterns and can even carry on a conversation convincingly.
Large language models (LLMs) are similar to the octopus metaphor. These models, such as ChatGPT, map out language by encoding patterns found in billions of written articles, books, and transcripts. When given a prompt, the AI locates the pattern that best matches it and predicts the next word in that pattern. Given the structured nature of language and the vast amount of information the AI has processed, it can produce impressive results.
The Capabilities and Limitations of AI
The capabilities of AI are still being explored as this large-scale implementation of technology is relatively new. LLMs have proven adept at quickly generating low-value written work, such as draft blog posts or filler content. They are also useful for low-level coding tasks that junior developers often spend countless hours duplicating. Additionally, LLMs excel at sorting and summarizing large amounts of unorganized data, making them valuable for tasks like summarizing meetings or research papers.
In scientific fields, AI models map out and find patterns in vast amounts of data, whether it is language or astronomical observations. While AI does not make discoveries itself, researchers have used AI models to accelerate their own discoveries by identifying rare molecules or faint cosmic signals. Furthermore, AI models can engage users in conversation on any topic and provide quick responses, making them surprisingly engaging conversationalists.
However, it is essential to remember that AI is only completing patterns and does not possess true knowledge or thinking abilities. Even technical literature refers to the computational process as “inference.” It is crucial not to mistake AI’s impersonation of human mannerisms and emotions for genuine intelligence.
The Challenges and Risks of AI
One significant challenge with language models is their inability to admit when they do not know something. Similar to the octopus metaphor, when an AI encounters something unfamiliar, it guesses based on the general area of the language map where the pattern led. This can result in generic, odd, or inappropriate responses. AI models may even invent people, places, or events that fit the pattern of an intelligent response, which are referred to as hallucinations. The troubling aspect is that these hallucinations are not distinguished from factual information, making it challenging to detect when AI generates false information.
Preventing hallucinations in AI models is currently impractical. This is why many applications involving AI models require human involvement to review or fact-check the results. By incorporating a human-in-the-loop system, the speed and versatility of AI can be utilized while mitigating its tendency to fabricate information.
Another challenge with AI is bias, which stems from the training data used. AI models require vast amounts of data to train effectively, but this data can contain objectionable or biased content. For example, if training data includes neo-Nazi propaganda alongside legitimate sources, the AI may treat both as equally important. This bias extends to images as well, where a lack of diverse representation can lead to skewed results. AI models cannot prevent biased training data, and efforts to address this issue are ongoing.
The Process of Generating Images with AI
AI-powered image generation is made possible by language models. Platforms like Midjourney and DALL-E use language understanding to associate words and phrases with image contents. The model analyzes numerous images, creating a map of imagery. This map is connected to the language map through a middle layer that links corresponding patterns of words and imagery.
When given a phrase like “a black dog in a forest,” the model interprets the phrase and finds the corresponding statistical representation on the image map. The image is then produced using techniques like diffusion, which gradually removes noise from a blank image to create an image that matches the given description.
The improved language understanding in AI models has enhanced image generation capabilities. Previously, image models required reference photos in their training data to understand specific requests. However, with improved language models, concepts like color and objects can be understood independently, allowing the model to connect them in its “latent space.” This reduces the need for guesswork and produces more realistic images.
While AI’s image creation capabilities are impressive, it is important to recognize that AI is merely completing patterns and does not possess true intelligence.
The Concept of Artificial General Intelligence (AGI)
Artificial general intelligence (AGI) refers to software that surpasses human capabilities in any task, including self-improvement. AGI is a concept that remains highly theoretical, and there are varying opinions on its feasibility. Some experts believe it may not be possible or may require resources beyond our current capabilities.
Although we have created sophisticated machine learning models for specific tasks, it does not imply that AGI is within reach. Predicting the nature or time horizon of AGI is comparable to trying to predict the future of interstellar travel. The current focus should be on addressing the challenges and potential risks associated with AI rather than speculating about AGI.
While the debate about AI’s potential existential threat continues, it is crucial to address the immediate problems caused by poorly implemented AI tools. The path of AI innovation remains uncertain, and whether it leads to superintelligence or reaches a limit is still unknown.