Shulou(Shulou.com)11/24 Report--

In ancient Greek mythology, a god named Hermes acted as a messenger between human beings and gods, wearing flying shoes with wings and walking between gods and human beings.

According to the Epic of Homer, "among the gods, Hermes likes to guide mortals most." This sentence is used to describe the relationship between OpenAI and AI, although it is not far away.

Last week, ChatGPT created by OpenAI was in the limelight and became the number one hot topic in the field of AI at home and abroad. About the dialogue ability of ChatGPT, you may have experienced it through many articles. To sum up, it is omnipotent and capable of writing code for programmers, working out solutions for business people and writing stories for writers. AI threat theories such as "Google is dead" and "the XX profession is going to be replaced by AI" that have been lost for a long time have begun to emerge in large numbers.

With regard to the magic of ChatGPT, I am also a little tired after watching too much. Calm down and think about it:

Why are NLP applications such as AIGC, Q & An and dialogue more likely to cause a sensation and inspire people's hopes for general artificial intelligence?

Why is the same pre-training model, compared with BERT, GPT3 and other predecessors, ChatGPT's dialogue ability has produced a qualitative leap?

Why do you also have to fight NLP when doing AI,OpenAI, iterating from GPT1 to ChatGPT?

CEO and co-founder Sam Altman of OpenAI once said, "Trust the exponential,Flat looking backwards,vertical looking forwards", believing in the power of the index, looking back in parallel and looking forward vertically. The emergence of ChatGPT represents that AI seems to have reached the key point of an exponential leap. But the take-off of ChatGPT is not achieved overnight.

From GPT to ChatGPT, it just represents the actual course of OpenAI in the field of large model, from which we can see that OpenAI has explored its own way in the competition of large model of AI, just like Hermes, it has become the emissary of leading the progress of AI technology.

If OpenAI is Hermes who conveys the advance of AI, ChatGPT is the pair of gold shoes with wings. While we should pay attention to how magical the ChatGPT shoes are, it is even more necessary to understand the mystery of the big model path chosen by OpenAI.

Today, Chinese science and technology enterprises and research institutions are actively investing in the layout of large models, so it is better to ask, from the evolution of this series of models of GPT, we might as well look at the strategic thinking and development of OpenAI about AI and large models.

From GPT-1 to ChatGPT, OpenAI, the evolutionary footprint of the supergod model, wrote on the blog that ChatGPT was born with fine-tuning of the models in the GPT3.5 series.

As the name implies, GPT- 3.5 is the fourth in a series of NLP models designed by OpenAI, followed by GPT- 1, GPT- 2, and GPT- 3.

Before the emergence of GPT, the NLP model was mainly trained based on a large number of labeled data for specific tasks. This leads to some restrictions:

Large-scale and high-quality labeling data are not easy to obtain.

The model is limited to the training received, and the generalization ability is insufficient.

Unable to perform out-of-the-box tasks, which limits the landing application of the model.

In order to overcome these problems, OpenAI embarked on the road of pre-training and training large models. From GPT1 to ChatGPT, it is a process that the pre-training model is getting bigger and bigger and the effect is getting stronger and stronger. Of course, the way to implement OpenAI is not just to "work miracles".

The first generation: from supervised to unsupervised GPT-1. In 2018, OpenAI launched the first generation of generative pre-training model GPT-1. Previously, NLP tasks required supervised learning through large-scale data sets and costly data tagging work. The key feature of GPT-1 is semi-supervised learning. First, with the pre-training of unsupervised learning, we spent a month on 8 GPU to enhance the language ability of the AI system from a large amount of unlabeled data and gain a lot of knowledge, and then made supervised fine-tuning and integrated with large data sets to improve the performance of the system in NLP tasks.

The effect of GPT-1 is obvious, and only a few fine-tuning is needed to enhance the power of the NLP model and reduce the need for resources and data. At the same time, GPT-1 also has obvious problems: first, data limitations, GPT-1 is trained in books and texts on the Internet, and the understanding of the world is not complete and accurate; second, generalization is still insufficient, and performance will decline in some tasks.

The second generation: bigger, higher and stronger GPT-2. GPT-2, launched in 2019, is not fundamentally different from GPT-1 (note this), with the same architecture, using a larger dataset WebText, with about 40 GB of text data, 8 million documents, and adding more parameters to the model (up to an astonishing 1.5 billion parameters) to improve the accuracy of the model, which can be said to be an enhanced or bloated version of GPT-1.

The emergence of GPT-2 further proves the value of unsupervised learning and the widespread success of pre-training models in downstream NLP tasks, which have begun to meet the requirements of the Turing test. Studies have shown that the text generated by GPT-2 is almost as convincing as the real article in the New York Times (83%).

(GPT-2 performance) the third generation: GPT-3 with leapfrog progress. In 2020, this iteration of GPT-3 took a major leap forward and became a very different species from GPT-2.

First, GPT-3 is unprecedentedly large, with more than 175 billion parameters, 117 times that of GPT-2; second, GPT-3 does not need to fine-tune, it can identify the hidden meaning in the data, and use the knowledge acquired by previous training to perform downstream tasks. This means that GPT-3 can understand and provide good performance even for examples that have never been touched before. Therefore, GPT-3 also shows a high degree of stability and practicality in commercial applications, through API access on the cloud to achieve commercialization. This ability to get into the laboratory and the workshop makes GPT-3 one of the most amazing models in the field of AI in 2020.

Of course, GPT-3 is not perfect. As co-founder Sam Altman said, the level of GPT-3 is still in its early stages, and sometimes it makes very stupid mistakes, and we are still a long way from the real world of artificial intelligence. In addition, many of the basic models of GPT-3 API are very large, requiring a lot of expertise and high-performance machines, which makes it difficult for small and medium-sized enterprises or individual developers to use.

The fourth generation: ChatGPT generated based on understanding. Finally, in 2022, there was another subversive iteration on the road of OpenAI's pre-training language model, resulting in another directional change in the technical route: reasoning and generation based on manual tagging data + reinforcement learning.

As mentioned earlier, the initial emergence of the pre-training model is to reduce the dependence of supervised learning on high-quality tagged data. On the other hand, ChatGPT began to rely on a large amount of manually tagged data on the basis of the GPT-3.5 large-scale language model (it is said that OpenAI found 40 PhDs to mark the data), how can it return to the "old way" of supervised learning?

The reason is that although GPT 3.5 is very strong, it cannot understand the meaning of human instructions (such as writing a blog post, changing a piece of code), and cannot judge the input, so it is very difficult to give high-quality output answers. So OpenAI uses professional tagging personnel (said to be 40 PhDs) to write entries, give high-quality answers to corresponding instructions / questions, and adjust the parameters of GPT-3.5 based on these data, so that GPT-3.5 has the ability to understand human instructions.

On the basis of manually tagging the training data, reinforcement learning is used to enhance the ability of the pre-training model. Reinforcement learning, the simple understanding is to do the right reward, do the wrong punishment, and constantly update the parameters according to the score of the system, so as to produce higher and higher quality answers. So in the past few days, many people have found in the interaction that ChatGPT will admit mistakes and revise their responses, precisely because of its ability to strengthen learning and rethink from human feedback.

Because ChatGPT has the ability to understand, it is regarded as the path to general artificial intelligence AGI.

Of course, ChatGPT is not a perfect evolution. OpenAI's website makes it clear that ChatGPT "may occasionally generate incorrect information" and that "knowledge of the world and events beyond 2021 is limited". Some of the more difficult knowledge, such as "what did A Dream of Red Mansions say", ChatGPT will talk nonsense seriously.

From the evolution and iteration of the GPT model, we can see that OpenAI is constantly moving towards the goal of natural language understanding, with a larger model, more advanced architecture, and finally found a path for general artificial intelligence.

The vertical evolution from GPT-1 to ChatGPT will see OpenAI's unique understanding and technical context of the large model-to improve the NLP index through model pre-training to reach strong artificial intelligence. What is so special about the field of NLP that it is worth OpenAI to be so persistent?

In the previous article, it is not difficult to see that OpenAI is committed to the text generation model because it has been done for a long time and invested enough, so it is very strategic for OpenAI to do better.

In contrast, the pre-training model launched by GPT-1 in the same year, and Google's BERT, but the latter's influence has significantly diminished after a period of popularity; while the field of NLP question and answer has always been led by Meta, Meta AI's OPT model and GPT-3 have reached the same number of parameters, but the effect is not as good as OpenAI. Among the contestants at the same time, OpenAI obviously paid more attention to the language model.

On the one hand, the input of resources, whether it is a growing model, requires a huge amount of computing resources. The high-quality tagged data required by ChatGPT depends on professionals at the doctoral level, which obviously consumes more manpower and financial resources than distributing the data tagging task to the crowdsourced platform.

On the other hand, technology input, large-scale pre-training, enhanced learning and other technologies are used to improve the understanding and reasoning ability of NLP dialogue system in the open general field. NLP is cognitive intelligence, and knowledge dependence must be solved in order to improve, and knowledge is very discrete and difficult to express. It is very technical challenge to solve the problems such as lack of marked data and common sense knowledge. Years ago, IBM's Frederick Jelinek said, "every time I fire a linguist, the performance of the speech recognition system improves a little bit." There is a strong sense of "if you can't solve the problem, just solve the problem." So it can also be said that OpenAI chose a more difficult path to solve the really difficult problem.

In addition, focusing on the NLP area also means that OpenAI will bear hidden opportunity costs.

This year, AIGC (AI generated content) has made great progress in both the capital market and the application market. Compared with the generation tasks such as AI painting, audio and video generation, and protein structure prediction solved by AlphaFold2, NLP tasks express concepts directly in terms of words and symbols. Such models complete commercial services through "API + cloud services", whether in terms of cloud resource consumption or interface call service charges. The income obtained is also far less than that of image, audio and video or scientific calculation. If you use the same energy to make ten or eight Dalle models, you will certainly earn more.

Technology blogger Wang Yonggang shared a story on his blog, saying that he communicated with the two co-founders of OpenAI and found that they didn't even know what AIGC meant!

At this point, it may be concluded that OpenAI, as a company that aims to "achieve secure general artificial intelligence (AGI)", is focused on improving the indicators of NLP tasks by pre-training large models, regardless of investment or commercial return, so as to approach the vision of AGI.

Why can OpenAI walk out of this trendy road of big model differentiation?

On the one hand, there is something special about NLP.

NLP is not magic, but the result is sometimes almost magic. General artificial intelligence must have cognitive intelligence, which is also the key bottleneck restricting the greater breakthrough and wider application of artificial intelligence at present, and NLP is the core of cognitive intelligence. Geoffrey Hinton and Yann LeCun have expressed a similar view that the next big progress in deep learning should be to let neural networks really understand the content of documents.

In other words, when AI can understand natural language, AGI may be implemented.

In addition, the operation mode of OpenAI also plays a key role.

Breakthrough innovation needs a lot of investment in the early stage, and the development of large models requires a lot of infrastructure investment, while the dialogue system of ChatGPT is difficult to balance the R & D costs on a large scale in the short term. Therefore, OpenAI is a non-profit research organization with no urgent commercial pressure, so it can focus more on basic research in the field of NLP, which is difficult for commercial AI companies to achieve.

In 2011, the natural language leader Kenneth Kenneth Church published a long article, "the pendulum swings too far" (A Pendulum Swung Too Far), in which he mentioned that our generation of scholars have caught up with the golden age of empiricism, picking down the low-branch fruits that are readily available and leaving the next generation with "hard bones".

Deep learning is a new peak of empiricism, and the fruits of low branches in this field will always be picked out one day. In recent years, a large number of AI scientists have warned that deep learning faces many limitations, and it is difficult to solve some complex tasks with deep learning alone. It may not take too long for basic breakthroughs to become an important support for the AI industry.

The evolution of GPT also shows that the breakthrough of AI needs to be realized step by step from small to large. Today, every AI enterprise and research institution is building a larger model. Compared with CV computer vision, digital human, meta-universe and other AI applications, NLP is much dimmer. And if a swarm to pick easy fruit, it will eventually restrict the pace of AI deep into the industry.

The emergence of ChatGPT reminds us that only by gnawing away the hard bones of the basic field can we really bring qualitative change to AI.

This article comes from the official account of Wechat: brain polar body (ID:unity007), author: Tibetan fox

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.