First, an announcement about an update
I have already embedded ChatGPT into EAC's chat rooms. In your daily conversations, whether you need to know how to use a certain word or need a host for a game, you can make use of it.
A few stories about me and ChatGPT
First of all, EAC was developed in collaboration with ChatGPT.
A month ago, I knew almost nothing about Next.js, audio encoding/decoding, WebRTC, and the like. When I began development, I estimated it would take at least two months to develop and test. Along the way, I might even need to ask friends for help—asking them to draw some design mockups or help me solve technical issues.
However, on the very first day, after I used GPT-4 to generate a reasonably laid-out page complete with icons, I realized that this wasn't actually that difficult.
In fact, it even changed my development habits and workflow.
First, let's talk about how I used to approach a small project. For example, if I wanted to build a speech recognition tool right now without knowing anything about the field, what would I do?
- Search engines help me establish some basic concepts. For instance, many blogs and encyclopedias mention ASR, so I'd first have to figure out what ASR is and how the concept relates to what I want to build.
- Divide into layers: While establishing concepts, gradually figure out what content might be relevant to what I want to do and what isn't, which things I can just use open-source tools for, and which features require training a model from scratch.
- Build a minimum viable demo: This step usually takes the most time and is the hardest. Going from 0 to 1 requires a lot of time to figure out those technical terms and the principles behind them.
- Start official work, gradually expanding from core functions to other features.
As a mature engineer, endless training teaches you how to go from a verifiable idea to final implementation.
The advent of GPT undoubtedly simplified many of the above steps. I can jump straight to the minimum viable demo in one step, and then turn around to ask questions continuously and learn the preceding background knowledge.
In other words, the execution order can become 3 -> 1 -> 2 -> 4.
Right off the bat, I have GPT help me write an executable demo. Then, I continuously interact with this minimally verifiable code snippet, which allows me to interact with core features much faster and more directly. My brain doesn't need to process a lot of complex, incomprehensible concepts; instead, it deals with vivid, modifiable content.
What does this mean? It means
GPT provides tools for first-principles thinking
It's like a differential gear: a teacher can demonstrate it repeatedly in class with PPTs and videos and it's still hard to grasp, but if you let kids hold one and play with it for a few hours, they can understand it right away.
GPT can quickly lower the barrier to entry for many things.
A lifelong learner will notice that a huge part of the barrier in some industries lies in concepts and jargon. When we don't understand these buzzwords and their underlying principles, a beginner feels like hitting a wall at every turn.
But GPT allows you to directly manipulate functionality in a short period of time, setting aside intermediate conceptual issues for the moment.
For instance, I recently wanted to study the proof of the SVP problem on lattices. But my math foundation isn't that strong; I know very little about lattices and groups, and I don't even recognize some of the mathematical symbols. Yet that didn't stop me from directly asking GPT questions I was interested in and having it prove them to me using some practice problems. This greatly reduced my learning curve and accelerated my learning process.
Here is how ChatGPT and I developed EAC:
I am relatively familiar with Rust and Python, but my knowledge of things like HTML dated back to years ago when I learned PHP in college. So when developing EAC, the first question was what tools to use and how to go about it.
For GPT, getting an answer might take just one second, provided you describe your problem in sufficient detail. At the time, its recommendation was Next.js plus Tailwind CSS for the frontend, and Rust could be used for the backend.
I've heard about the drama and hype in the frontend ecosystem and know that they follow different trends every year, but I only wanted the simplest and most stable way to update my code.
So I asked GPT to write the same page using different frontend frameworks, allowing me to see what different code styles looked like, what kind of syntactic sugar they had, and what the experience of writing CSS in different ways felt like.
In the end, my choice of Tailwind CSS really was related to GPT. Because Tailwind writes CSS directly into the JSX file, whenever I adjusted the page layout, if it were SCSS or CSS, I'd have to copy and paste twice, and GPT would also need to generate two chunks of code. But with Tailwind, I only needed to copy and paste once.
The above was just the basic stage of using GPT; later, I entered the deep-usage stage.
Ever since I found that the code generated by GPT was actually usable, I've barely written code by hand. My daily development process basically looks like this:
- Describe what kind of page I want
- Copy the code generated by GPT into WebStorm
- Check the result, then request further modifications
It has even developed to the point where, to this day, I still haven't figured out how flex layout really works or how to vertically center an element.
Every time, I just tell GPT, "Please center my content," and then directly copy the modified code it gives me.
So you realize that sometimes the knowledge we acquire is just junk knowledge—content we might not have needed to understand in the first place, yet it undeniably holds you back. And GPT is the salvation from all of that.
A few extra thoughts
I once worked as an NLP engineer in China for a period of time, and the sci-fi novel "Dumpling" was written during that phase. The AGI described in "Dumpling" emerged from all natural language being trained into ultimate general intelligence through simple reflexes amidst infinite computational resources.
I think every engineer who has worked in natural language processing has had this dream: processing all the corpora in the world so that the accumulated knowledge of human civilization can be expressed through computation. Yet every rational engineer knows that such a vector space could be as complex as the universe, making it cost-prohibitive. Arthur C. Clarke also wrote a similar story about the last five minutes of the universe; in his imagination, after humanity went extinct, the computer calculating everything still hadn't produced a result.
Yet today, LLMs seem to already possess a form of language understanding, coupled with nearly all knowledge. Every individual's capabilities are amplified by them, and anyone can do things that were previously beyond their reach. Why not give it a try right now?