**How AI Works, and Why It’s Like a Secret Octopus**
AI, or artificial intelligence, can be thought of as software that approximates human thinking. While it’s not the same as human intelligence, even a rough imitation of human thinking can be useful for completing tasks. AI is often referred to as machine learning, although the terms are somewhat misleading. Machines don’t actually “learn” in the same way humans do. Instead, AI models are excellent at detecting and continuing patterns. Computational linguists have likened AI to a hyper-intelligent deep-sea octopus that can build a detailed statistical model of patterns it detects, even without understanding language or human concepts.
**The Concepts Behind Today’s AI Models**
The concepts behind today’s AI models have been around for decades, but recent advances have allowed these concepts to be applied on a larger scale. This has led to the development of convincing language models like ChatGPT and realistic image generation like Stable Diffusion. To help people understand how and why AI works, a non-technical guide has been created.
**What AI Can (and Can’t) Do**
AI models have proven very capable at quickly creating low-value written work, such as draft blog posts or copy filling. They are also useful for low-level coding tasks and sorting/summarizing large amounts of unorganized data. In scientific fields, AI can map out and find patterns in large piles of data, accelerating research processes. Additionally, AI models make engaging conversationalists, but it’s important to remember that they are just completing patterns and don’t possess actual intelligence.
**How AI Can Go Wrong**
The problems with AI are often due to limitations rather than capabilities. One major concern is that AI models don’t know how to say “I don’t know.” When faced with unfamiliar input, AI models may guess based on existing patterns, resulting in generic or inappropriate responses. Another issue is bias, which can be present in training data. AI models require large amounts of data to analyze patterns, and if the data includes objectionable or unrepresentative content, the AI may generate biased responses. Currently, there is no practical way to prevent these issues, but having humans review and fact-check AI results can help mitigate the problems.
**The Importance (and Danger) of Training Data**
AI models require vast amounts of training data, which can lead to issues like inappropriate or unrepresentative content. Scraping billions of web pages may result in objectionable material being included in the training data. Furthermore, the training data used for AI models is often obtained without consent, raising ethical concerns. As AI companies grapple with these issues, solutions like trimming training data or refusing to discuss certain topics have been proposed. However, bad actors can find ways to circumvent these barriers, highlighting the ongoing challenge of aligning AI models with societal expectations.
**How a ‘Language Model’ Makes Images**
AI-powered image generation is made possible by language models. These models analyze images and associate them with words and phrases through statistical representations. For instance, given the phrase “a black dog in a forest,” the model understands the concepts of black, dog, and forest independently and connects them on a statistical map. This allows the model to generate an image that corresponds to the given phrase. The improved language understanding in AI models eliminates the need for specific reference photos and reduces the weirdness often associated with generated imagery.
**What about AGI Taking Over the World?**
Artificial general intelligence (AGI) refers to software that can exceed human capabilities and potentially cause harm if not properly aligned or limited. However, AGI is still a concept, and creating it may require methods or resources beyond our current capabilities. While highly convincing machine learning models exist for specific tasks, creating AGI is a different challenge altogether. The debate surrounding the potential threat of AGI is ongoing, and it remains uncertain whether AGI is possible or what its nature would be if it were achieved. It’s important to address current AI problems while considering the concept of AGI.
