ChatGPT and the Human Race
This past week, ChatGPT was released and has been gradually permeating into the twitter feeds and discord channels of various in-groups across the tech world. On multiple occasions, anecdotes have been highlighted to me that indicate that it is capable of solving essentially any common programming question, and is able to scale to even graduate level CS algorithm questions. It can also draw out of large body of text accurate summaries that gracefully express the core thesis. This follows the popularity of image generation models, DALL-E and the like.
It’s really too early to tell what we are looking at here. But it is clear that a fundamental revolution of technology is happening. It will have far-reaching implications into every aspect of every industry. Nothing will be left untouched by, what I will dub, the Deep Learning Revolution.
The Deep Learning Revolution, to my understanding, will present a two core axioms to the methodologies that humans take to solve problems. I will list them, and then articulate my reasonings behind each. The first axiom is that, eventually, every problem that has been previously solved by humans will not need to be solved again. The second axiom is that the differentiator for things that are “good” and “bad” will change to highlight only the portions of the human experience that these neural networks cannot reproduce. And importantly, the portions of human experience that these models cannot replicate is slowly shrinking. Only time will tell if an island of human ability will remain untouched by the Deep Learning Revolution.
The first axiom can be approached by understanding how these models “learn”. The model can be conceptualized as a massive ensemble of linear regression problems. Newer and more advanced deep learning models add nuance and complexity to allow the model to have greater capacity for learning, but we can imagine that essentially we are asking the model to “train” a hyper-parameter W such that Wx = Y. Our x is the questions we ask our model to train it, and Y is the answers we deem “correct” to these questions. Once this has been trained, this hyper-parameter now can act as a tool to answer new questions. The higher-level mathematics is not too core for our understanding, because the point that must be expressed is this: The model is only capable of predicting accurately on a new question when it’s x it was trained on has some understanding of the question.
Out of this realization, we can see one important thing become clear. Deep Learning Models are highly effective at solving problems they have seen before. As the interface between the total corpus of all human code ever written and the ability to access a model trained on this code grows more ergonomic, engineers no longer have to toil over trivial solutions to any problem. Any problem that has ever been solved before will be suggested to the engineer or mathematician or artist before they have a chance to bother how they would approach it. At some point, there will be a tool that can be given a prompt for a software application and will return the application to the user. The time requirement for conceptualization and construction of massive software systems will shrink. Ultimately, in my opinion, this will greatly accelerate the ability for the human race to progress technologically. Teams can become leaner as any trivial code can be repurposed instantly.
The second axiom ties to the first axiom. There will become two types of end-users after the Deep Learning Revolution. There will be those that will solve problems that have not been solved before, and those that will re-arrange the corpus of human knowledge in creative ways which benefit society. The second group leverages creativity, empathy, and understanding of the human experience. This is all they can offer, and all that is needed. As the first group solves more problems, the second group has more tools in their arsenal to arrange modern technology to improve the lives of other humans. It is improbable that the first group runs out of problems to solve, we hope.
I believe that much of the phenomena I have described has already begun to happen. The Deep Learning Revolution will change how and what humans learn, and what we deem necessary to be a successful and contributing member of society. Right now, I believe my view may be myopic in the sense that I can think about how this will affect the world of engineering and tech, but not it’s far-reaching effects into the corners of our world. I hope I can improve this understanding, and I hope that ultimately this is a net-positive on our world.
Hey, welcome to the site! I’d recommend reading through a bunch of the most popular posts; you’ll be more able to write posts that are upvoted after gaining more understanding of what folks are looking for in a post on lesswrong.