While reading about how GPTs work, I stumbled upon an article that explained it in a simple language. Out of the many parts that stood out, here’s one – "To try to explain the meaning and define the words to a computer is a dead end, the task is way too complex (people have been trying for decades). How can you even represent the meaning of a word? Well, there is one thing that computers understand perfectly: numbers. What if we represented the meaning of words as numbers along several axes?" It made me realize that learning about how advanced AI models work is to unpack and understand how we function. Since AI models are built by studying biological brains, it should not be surprising or come as a revelation. But it’s still an important point though because nearly not as many people are interested in trying to understand how the human brain works in a fundamental way as they are interested in learning how AI models work. I hadn’t really thought about how we form meaning, or how human languages evolved, but now, trying to understand how these models work is helping me understand how I, or we all, work.
If meaning were to be represented with numbers, the article goes on to explain, then it can be modeled not as one thing, but a combination of a ton a ton of them in different dimensions, along different axes. What it means then for something to be subjective is that different people have different axes, both in number and form, with different weights for each one, which defines the meaning they arrive at. Let’s take an example – If someone asked me what parents mean to me, I might subconsciously think of the following ways to describe them –
1. Those who deeply care about me
2. Those who knows me well
3. Those who I can learn from
4. Those I can look up to, to build a moral compass
5. Those that I’m biologically related to
6. Those that I’m extremely similar to, by way of inherited genes
7. Those who speak the same language as I do and share the same house as I do
8. Those that quite often bug me about things and don’t let me be
9. Those who meddle in my private matters
10. Those that I’ll have to take care of in their old age
11. Those that have provided me financial and emotional support
12. Those that have made me who I am
... and the list could go on. You can think of each one as a point on a different axis. Now if I were to score each one of them, using a number between 0 and 1, one can arrive at a decent estimation of what parents mean to me, and can, with reasonable confidence, tell whether or not my general sentiment about them is positive or negative. But other people could have completely different dimensions and scores.
To know oneself then is to know the axes and the weights we carry on them. A deeper understanding might even illuminate how we come up with them in the first place. One of the best ways to understand our own models is through reflection, introspection. It's hard work, but the things we have strong reactions for are a good starting point. The things we're triggered by should give us a hint. They tell us that they hold heavy weights in certain axes. So it's helpful to reflect on things that trigger you, evoke strong reactions, either positive or negative. The more you do it, the better you get at finding patterns, and the better your chance of surfacing the associated dimension.
One other point that stood out to me from the article – although the models that are being produced now are fairly advanced and huge in the number of parameters and the knowledge sources they have, our brains have a few orders of magnitude more still, and the brain is a chemical, and hence an analog system, rather than a digital one, which means there are more nuances than bits could ever capture.
Thinking about it in terms of axes and weights also gives a model on how we can acquire new axes, i.e open up our worlds. The more restricted our sources are, the fewer the axes we have. Learning increases the number of axes available to us, and hence adds more "meaning" in our thinking about the world. The more nuanced our thinking, the more likely it is for us to interpret a situation from different angles. The models that are popular today are trained on the Internet data, which means that they are going to inherit the biases and preferences of the population that’s on the Internet, rather than the broader human population. While one might argue that that’s okay, because over time more of the human population will be online, it’s still not because the Internet has a tendency to attract extreme viewpoints, both on the good and bad side, so the neutral, more nuanced sources lose out eventually. You can think of this problem as akin to the loud voices making all the decisions in a meeting. They might be wrong, but it doesn’t stop them from having a bigger influence if deliberate measures aren’t taken to counter-balance them. One other implication of using the Internet as the source of training the models is that they are less useful for personalized tasks, because they don't know you. The future, as some articles argue, and I agree, is all about customizing the AI to suit your needs. Feeding it all the things you've read, heard, watched is going to be hugely beneficial. The more you align the model's sources with what you consume, the more the model is going to act like you, think like you, reason like you, which brings me to another interesting point – GPTs are reasoning engines, more than knowledge engines.
That’s a useful distinction, and helps understand what it’s good for. Current search engines, for example, are knowledge engines. They can regurgitate the information fed to them in a manner that makes it easier to search from a vast data source. But since GPTs are reasoning engines, they are going to have a hard time saying that they don’t know something. If the knowledge sources are limited, they’re going to bullshit their way through. It's not different from our brains in that respect. Our brains are really good storytelling and reasoning machines, which is evident from the fact that we can rationalize our actions even when they don’t make sense. To borrow Jonathan Haidt’s metaphor of elephant and the rider, the elephant, which is the primitive brain, compels us to do something, and the rider, the rational brain, builds a story that justifies that desire, or action.
Overall, although I hold a pessimistic view about the use of AI for good, the silver lining I see of the technology is that in a long-winded way it helps us understand what being human means. It helps us appreciate how the brain works, understand its quirks, limitations of the AI models, and how we can change, for good, the parts of our brain that’s not serving us well in today’s world.