Does AI understand the world?
Do you understand the world?
ChatGPT is the fastest growing software application in the history of the world. Some are saying ChatGPT is the most complex set of software engineering ever made. I’m not sure about that claim. The Internet and the Web are an amazing set of technical protocols that have enabled ChatGPT. The elegant simplicity of the Web is transparent to most people. Likewise, the complexity behind ChatGPT is unknown to most. Like the Internet, ChatGPT just works.
The core technology behind ChatGPT is in its name: GPT (Generative Pretrained Transformer). Less known are the letters LLM for Large Language Model. For most people, the letters GPT and LLM might signify something important but will remain largely no more understood than the letters http or HTML that people see and use everyday.
You know you’re in a hardcore AI bubble if you know the meaning of LLM and GPT. When I casually talk with most of my faculty colleagues at the university, they’ve all heard of ChatGPT but barely recognize (if, at all) LLM and GPT.
LLMs build a representation (i.e., a model) of the world by processing a massive amount of textual information (i.e., language). LLMs are enabled through a computing architecture known as neural networks. (For this post, let’s set aside the lengthy technical discussion of deep learning in neural networks to another time.)
Future forms of AI, later this century, may build their mechanisms on foundations other than LLMs. But for the next decade, probably the next twenty years, LLMs are the foundational aspects of what AI knows about our world.
There are many limiting factors today, like the text only nature of these models. Multimodal models incorporating audio and video are in development. Another factor contributing significantly to what AI knows about the world through LLMs is human feedback. Much of the value that makes ChatGPT appear so compelling is enabled through the human input that has enhanced the responses and behavior of the system.
Let’s step back and think about LLMs and text. This is a discussion not only for computer scientists and software engineers. Students and scholars of the humanities have vast insight that’s needed in shaping the future of AI. Humans through the last few hundred years have largely came to understand the world through reading text.
How do we, as humans, understand the world around us?
Or, we can remove ourselves, and ask, “How do animals understand the world?” I should clarify that I mean “non-human animals”.
Animals do not have access to Wikipedia. Animals are not using the Internet. Animals are not walking into the library and doing research. Animals are not reading books or watching movies and documentaries.
Animals watch each other. Young animals learn by observing their parents. Based on their sensory perception, animals develop an understanding of their world.
Humans also learn by observing, though the dominant method for humans to gain knowledge is through language. Concepts expressed through words are communicated through speech and written down as text. Long before computers, we systematized knowledge in books. An educated person was synonymous with a well-read person. A person who read a lot books is considered intelligent. A person who reads deeply on a subject has the knowledge to analyze a problem and identify solutions. Analytical and problem-solving skills increase with experience gained from observing best practices and the behavior of more experienced people in unusual situations in the specified domain of work. Knowledgeable people can recommend what to do next based on the patterns that they have seen or their prior experience and their depth of reading the literature in that field.
Our actions are based on recognizing patterns.
Often, we have flaws in our thought processes caused by biases, skewed assumptions, and errors in what we have learned.
Of course, how the brain works is tremendously complex, and I’m skimming over extensive areas of study in neuroscience.
The takeaways for consideration:
Our choices in making decisions or taking action are based on our perceived probability that we are making the correct decision based on our experiences and what we know.
Our knowledge is how we understand the world and actions we take next are statistical calculations based on our experiences.
How does AI differ?
As humans, we get a lot wrong. Human expertise is expected to get things right, though the experts are not always correct in everything within their area of expertise.
With machines, we aspire for a higher probability of accuracy than from our fellow humans. For critical matters of health and engineering, for example, we have certifications and checkpoints that attempt to strengthen the quality of work done by those people. With machines, there will exist (for certain tasks) regulations and standards that certify specific machine applications as meeting a specified level of accuracy. A new industry and career opportunities are emerging to ensure AI responsibility.
On the day I’m making this post, April 11, 2023, an agency of the US government, the National Telecommunications and Information Administration issued an RFC (Request for Comments) about AI accountability. That RFC is worth a read. I’ll be taking a longer look at it, and looking forward to reading the public comments on proposals for the evolving policy issues.