How LLMs (ChatGPT) change the Future of Humanity
LLMs (large language models) like ChatGPT will take us a step closer to AGI.
With the recent amazing breakthroughs in machine learning and AI, I was wondering to myself what the future could look like. I want to look back in the future at this article to see how right or wrong I was.
My experiences with ML and AI
The Era Before ChatGPT and Midjourney
I remember when I started to learn ML, that was around two years ago. That was a nice experience, I remember that I did an image classification and bible generator in Hebrew, it even pass the Turing test! It definitely an interesting topic.
I saw ML as a limited field. You can’t just slap ML to any problem and expect it to be solved.
Then GPT3 and Midjourney got Released
And I was shocked. I would never thought it could be possible that computers will be able to create art or that I can have an interesting conversation with them. What is even crazier is the rate they got improved.
I remember when you could spot a mid-journey art just by the look, but now with version 4 of mid-journey, it can draw hands! A task that it was very poor at. And don’t even let me start talking about chatGPT.
It is funny to think that AI takes control of the creative fields first, in opposition to all the sci-fi movies about AI.
The New Revolution
First, computers were invented. Then later comes the internet which generates a lot of data and because of the improvement of computational power ML started to get used.
Regular software can do a specific task but you must provide the algorithm first, then the software can do only this specific task. ML can do a task without any algorithm being defined before.
But the model you trained can do only the task that it trains about. If you train an ML model that detects dogs in an image it would not be able to recognize dog sounds or even cats in images.
What is special about ChatGPT is the variety of tasks that it looks like it can do. It can find bugs in software, write stories, detect if the text is positive or negative, and so on. GPT was only trained to continue the next word in a given text, but it was trained on so much data that it looks like it can do a lot of things and sometimes feel human.
The rate of adaption of ChatGPT is amazing:
Also, Google and Microsoft announced the use of AI in google docs and Office. Everyone could use AI to do less work.
Example of Usecase of LLMs
Think about a large codebase that you need to maintain and refactor.
One possible way to do this is to hire a couple of developers, which is expensive only to find good ones. you will need to pay them and also they need to sleep, eat, and talk about their weekend.
On the other hand, think about kind of Github co-pilot that has the context of all tens of thousands of lines of code, that can remember all or most of the codebase.
This AI can scan all the files, write documentation for itself, and for you, can find code smells and bugs, debug, write code, write tests, open PRs, and even code review itself and fix it. It will be able to use tools almost like a human, like a debugger, linter, compiler, calculator, StackOverflow, and so on. And will work nonstop.
Maybe it will be stuck sometime or make mistakes and will need human intervention, but it will learn each time it happened and be more fine-tuned to the task and to the code base.
Something like that can be already exist or released soon, GPT4 has a context of 26000 tokens! Bing can search the internet. LLMs can use external tools.
In the future, there will be only human intervention to the AI. You will not develop code anymore or be a lawyer, just check that the AI is doing its job right, and when it is not you train it to become better in replacing you.
Current Limitations of the LLMs
The current limitation of LLMs is only the data. The new models are almost trained on all the internet and we will get there soon. Of course, there are a lot of ways to improve, train on more quality data, train multiple times, and so on. But in the end, we don’t have infinite data to train the model.
Why ChatGPT is so smart?
ChatGPT feels much smarter than GPT3. It was because it was fine-tuned to be chat assistance. OpenAI used real humans to classify which answers are better to train ChatGPT.
Researchers at Stanford train the model “LLaMA” to become better be using GPT. basically, AI teaches other AI.
Overcome the Limitations of the LLMs
Think about a lot of models of AI that specialized in each field. Think about LLMs as a common sense and a way to understand and talk with humans. Then on top of that it can be fine tunned with data that is relevant to its task and all the human supervision that it learns from to be great at its task.
Then all the AI can teach others. Image a big circle table where all the AIs are sharing their ideas and knowledge and experience with all the other AIs. They can teach others. Image AI which specialized in math teaches some concepts to an AI that is specialized in python code.
The AIs can have thousands of years of experience and learning in a couple of hours. Then possibly the AIs can become smarter just by themselves.
The Short Path to AGI
Then if the AIs will continue to be smarter to the point they can invent and research new things they will be able to think about a better way to create an AI, like humans create an AI that is smarter than them the AIs will do that also, then the new AI will create smarter AI and so on.
Conclusion
That is maybe sound like science fiction but humanity was never so close create general AI. What a time to be alive!
