Okay, let's
continue to make progress. In 2010, a field in machine learning began to shine,
called Artificial Neural Network. You see, our human brain actually relies on
over 100 billion neural sources through meshed connections to judge and convey
information. Although each neural source is simple, they can be combined to
judge very complex information, so this artificial neural network is actually
trying to simulate this form of the human brain. After inputting information,
it will go through several hidden time nodes to make judgments, just like
neurons, and then give you output results. In fact, the idea of this neural
network has been around for a long time, even tracing back to the 1960s. But it
needs two things to support it: a large amount of data and powerful computing
power, which were not available before. So, this neural network was just a
paper tiger. In the 2010s, it was the age of the Internet. Data must have been
available, and computing power has also been continuously improved
exponentially, which allowed this neural network to start to be applied. People
found that this pattern is really suitable for solving problems where people
can instinctively know, such as when you see a face, you can quickly know who
it is. Previously, it was extremely difficult to let a computer determine who
this person is, but with this neural network, machine learning can slowly
explore the rules. Now its application is already very wide, not only face
recognition, voice recognition, automatic driving, but also the AlphaGo that
defeated Ke Jie in Go a few years ago was trained by this. Therefore, this
neural network can display its talents in all those fields we just mentioned.
What about the textual field? Its development hasn't been smooth, why? Because
this machine learning usually uses a kind of recurrent neural network, called
RNN, to process text. Its main way is to look at each word in order, so it
cannot do a lot of learning at the same time. Moreover, this sentence cannot be
too long, otherwise, when you are studying later, you will have forgotten the
beginning. Until 2017, Google published a paper proposing a new learning
framework called Transformer. The specific mechanism is more complicated, and
it is certainly not something that Lin can understand, but the result is that
it allows the machine to learn a large amount of text simultaneously. Like the
original words, you had to learn them one by one, just like connecting
circuits. Now you can learn them at the same time, just like parallel
connections. Suddenly, the training speed and efficiency have greatly improved.
With this transformer, machines in textual learning are like opening up the
Governor Vessel and the Conception Vessel.
Many natural language processing models are actually built upon the framework of transformers, including the T in Google's BERT and the T in ChatGPT. As we have seen major breakthroughs in technology, all that is left is for us to invest the necessary resources, including people and money, to bring ChatGPT to the forefront. In 2015, Peter Thiel and some other high-profile investors, including Elon Musk, funded a non-profit organization called OpenAI to conduct research in AI. OpenAI's research findings, including patents, are all public because it is a non-profit organization that exists solely to promote the development of AI technology.
OpenAI has been a great success, and in 2017 they began researching and learning from Google's transformer model. In 2018, they released the first generation of their Generative Pre-trained Transformer (GPT) model, followed by the second generation in November 2019, which had an increased training data volume. GPT was different from previous language learning models in that it did not need human supervision or labeling. Instead, by feeding it large amounts of data, the machine could learn on its own. The key to machine learning is having a good model and large amounts of parameters, which require significant computing power and funding. Open AI has great confidence in this model, but it will take significant investment in both aspects to make further strides.
In 2019, due to financial pressure, OpenAI became a for-profit organization with a revenue cap. This means that any investor's return on investment cannot exceed 100 times their original investment. However, Microsoft immediately invested $1 billion in OpenAI, resulting in a win-win for both parties. OpenAI gained the funding they needed, as well as access to Microsoft's fifth-largest supercomputer in the world, which greatly improved their training efficiency. Microsoft, in turn, got access to OpenAI's technology and team. However, OpenAI's research findings will no longer be made public.
OpenAI's GPT-3 model, which was released in 2022, has greatly improved the effectiveness of intelligent conversation. However, it still has a weakness: it can provide excellent answers at times but not so well at other times. One reason for this is because it lacks a good feedback mechanism during training. For example, when learning to play chess, the machine knows that winning is good and losing is bad. But with chatting, it is difficult to judge what is good or bad. To address this issue, OpenAI added a human feedback mechanism during training, called reinforcement learning. With the addition of this mechanism, the efficiency and effectiveness of ChatGPT were greatly enhanced. In March 2022, OpenAI released GPT-3.5, followed by ChatGPT in November of the same year. Although ChatGPT may have some limitations, it is able to chat about almost any topic with a convincing level of language expression. Thanks to these enhancements, ChatGPT is expected to easily pass the Turing test after half a century's worth of development—an impressive achievement.
Article write in 4-26,2023
