First statement: This article is not generated by GPT-4! The GPT-4 released by OpenAI is popular all over the Internet. It is foreseeable that the next 5-10 years will be dominated by large models. How many essential breakthroughs has it brought? How many jobs will be replaced? Is NLP still worth learning and if so how? This article takes you to analyze the above questions!
No.1
GPT-4 out of circle
If you haven’t heard of RoBERTa and T5 in the field of NLP large-scale models before, then ChatGPT and GPT-4, which are completely out of the circle this time, seem very "sudden". So close, of course, domain experts who are familiar with NLP know that they all belong to the Transformer architecture, but based on different parts.
Let's first take a look at the capabilities demonstrated by an example picture in the official GPT-4 paper (Tech Report). The user uploads a picture and asks where is the picture funny? GPT-4's answer already nailed the humor in this picture: "Putting a large, outdated VGA port into a small, modern smartphone charging port is ridiculous," which certainly shows that GPT-4 already has Multimodal information input and integration capabilities.
And one day after the release of GPT-4, Baidu released Wenxin Yiyan. It is worth mentioning that Wenxin Yiyan has multi-modal output capabilities, that is, images can be obtained through text descriptions. The following figure is an example of text generated images. (ps: Baidu understands programmers)
At the same time, Google Bard was released in a "low-key" manner, and the ChatGPT project of Alibaba Dharma Academy also started internal testing. When various manufacturers started to follow the investment mode, OpenAI released another blockbuster on March 24, announcing that ChatGPT can inherit Third-party plug-ins, which means that ChatGPT can be used for shopping, air tickets, ordering meals and other areas of basic necessities such as food, housing and transportation. This commercialization scenario finally allows the capital behind OpenAI to initially taste the sweetness of investment (after all, from GPT-1 to 3 generations only ran out A demo...), in addition, programmers who write GPT-4 may also face the situation that "the AI trained by themselves is better at writing code questions than themselves".
Does that mean that GPT is about to completely replace conventional documents, accountants, content creators, painters and even programmers in various fields?
No.2
Who will GPT-4 replace?
From the field of accounting or statistics, ChatGPT can be used as an Excel operator to produce the functions that users want to achieve (as for why the English Prompt is used because the English effect is better, but Chinese answers can be required at the end).
Another example is the programmer mentioned above, please note that he is better at "writing code questions" than himself. It cannot be used directly and requires manual verification, but even this is already very powerful, which means that the underlying coding of ordinary programmers can basically be replaced by GPT, and only further Debug or Update is needed on this basis.
Therefore, whether it is ChatGPT or the more powerful GPT-4 positioning is more like an auxiliary AI, which can help people better complete professional work in vertical fields. Bottom work.
No.3
Will GPT-4 end NLP?
Furthermore, after the release of GPT-4, has the work in the NLP field been completed? This reminds me of the famous British physicist William Thomson who said when reviewing the great achievements of physics, "The building of physics has been completed, and all that remains is some decoration work." However, it is these so-called insignificant decorations The work unfolded into two dark clouds, the first dark cloud developed the theory of relativity, and the other dark cloud developed quantum mechanics.
At this time, we may be like William Thomson at that time. In fact, Mushen also said at the recent "group meeting" of station B that the development of Transformer is still in its early stage, and the current prompt paradigm is obviously not the ultimate AI form. It did not pass the Turing test, and just yesterday OpenAI officially announced that GPT-5 would be released at the end of the year. Stop training AI systems stronger than GPT-4 for at least 6 months", including Turing Award winner Yoshua Bengio, Stability AI CEO Emad Mostaque, Apple co-founder Steve Wozniak, New York University professor Marcus , Musk, and Yuval Noah Harari, the author of "A Brief History of Mankind" and so on.
Recently, Yann Lecun, one of the three giants of AI, even claimed that within 5 years from now, no one with a normal mind will use the autoregressive model, and the autoregressive model mentioned by the boss is the underlying layer that the popular GPT family model relies on. learning paradigm.
Because in LeCun's view, such models have no future (Auto-Regressive LLMs are doomed). Although they have performed very well, there are many problems that are difficult to solve, such as factual errors, logical errors, inconsistencies, limited reasoning, and easy generation of harmful content. Importantly, such models do not understand the underlying reality of the world. So it remains to be seen whether GPT-5, which will be released at the end of the year, will solve this problem.
To sum up, occupations cannot be replaced, and field research still has a long way to go. Compared with NLP or AI fields, GPT-4 is undoubtedly a technological change, which will promote the development of artificial intelligence in a more "smart" direction, so , students who have not yet learned NLP or students who are already in the pit, hurry up and get on the bus!
No.4
NLP learning route
The first is the bottom-level learning route of NLP. As the threshold for NLP use gradually decreases, the cost of NLP "SOTA" (that is, models that produce better models than the industry's best models) is getting higher and higher. Achievements in one's own direction and the ability to form one's own core competitiveness are all things worth continuing to think about.
No.5
NLP theoretical framework
Secondly, enter the huge territory of NLP, starting from word2vec, then to Transformer and then to BERT, in the middle you will feel the charm of the intersection of natural language and computer language, and thanks to the paving of the predecessors, we stand in torch or keras With 10 lines of code on your shoulders, you can reproduce these studies and quickly see the results of your own model. Of course, this is not enough, we need to further adjust our model to meet the needs, which requires us to study the following theoretical knowledge and practical skills in specific fields in depth, so that we can finally become experts in the field of NLP!
No.6
NLP dissertation route
Finally, of course, a map of classic and soon-to-be classic cutting-edge papers in the NLP field must be attached. The number of papers in it may not account for 1% of the NLP-related papers published in recent years, but it is definitely the most worthwhile selection. Papers can obtain the desired information in the shortest possible time. And in the reading of the paper, you can read the abstract, introduction and conclusion first. If you are sure that it is the paper you want to read, you can read it further. This reading order is the most efficient after trial and error. All the papers in the picture have been uploaded to the official account, you only need to reply [NLP papers] to get a PDF collection of all papers!
In general, Natural Language Processing (NLP) is a very broad and deep subject, including many aspects, such as machine learning, computer science, mathematics, statistics, psychology, etc., which requires a lot of time and effort And with a lot of reading and practice, you can learn the real thing. Everything is difficult at the beginning. At the beginning, you can use some B station tutorials or online tutorials to quickly cut in and collect points of interest. Slowly learn step by step, and you will find more natural language. Mystery, I finally discovered that text, images, and videos are all the same thing, and maybe I will go to a higher and farther stage to show my talents. I wish you all success in learning and making money every day!
It's not easy to be original, and watching collections is a great motivation for sharing! If you have any questions, please leave a message directly~
Directly reply [NLP mind map] in the official account to get the high-definition PDF download link
History recommendation
Original super detailed Python learning route + mind map
Data Analysis of "High-Frequency Face-to-face Classics"
"High-frequency face-to-face classics" machine learning
In-depth learning of "High-frequency face-to-face classics"
Selenium crawls Jingdong products
Scrapy framework crawls QQ music
Data Analysis and Mining
Data Structures and Algorithms
Deep Learning and Big Data Framework
Welcome to pay attention, thank you for "watching", it's rare to follow the fate~
