Comment | General artificial intelligence needs to be demystified, and we should calmly look at the popularity of "ChatGPTs"

When new tools, new modalities, and new technologies emerge, we must learn “how to use them correctly” while sighing in despair. The essence of AI is to give traditional industries the confidence to implement their own AI, otherwise AI will be demonized.

Written by Lei Tao Edited by Zhao Tianyu

Lei Tao (Chairman of Tianyun Data, corporate mentor of postdoctoral workstation, Zhongguancun high-level talent, winner of the "Wu Wenjun Artificial Intelligence Science and Technology Award", has more than 30 years of experience in IT industry. He started to develop products for distributed data and machine learning in 2010, and led and participated in the planning and construction of many large-scale information projects with a scale of 100 million yuan, such as China Mobile, China Unicom, and ICBC headquarters.)

"Ten months of pregnancy, one day of birth". On March 16, Baidu's "Wenxin Yiyan", known as the Chinese version of ChatGPT, was officially released. Judging from the on-site demonstration, Wenxin Yiyan's accuracy, logic, and fluency in answering are gradually approaching human levels. After the press conference, the outside world also showed mixed reviews, and it became a hot topic of social concern for a while.

Since the beginning of this year, topics related to AI have been constantly gaining popularity: first, ChatGPT triggered a nationwide "chatting craze", and then GPT-4, known as the "most advanced artificial intelligence system", was released; domestic manufacturers have followed up with new products and are ready to go. Regardless of whether you pay attention to technology news or have used "ChatGPTs", you must have heard the following comments and judgments about general AI products.

▲On March 16, Baidu's "Wenxin Yiyan", known as the Chinese version of ChatGPT, was officially released

Extreme myth theory: General AI will have a disruptive impact on mankind. All cognitive-based jobs will be replaced by AI. General AI will become the Matrix in the movie The Matrix, a system used by robots to enslave humans.

There is a serious belittlement theory, a typical "sour grapes" mentality, which believes that general AI is still not worth mentioning at present, and ChatGPT is just an enhanced version of the search engine and cannot escape the fate of "artificial intelligence".

Overly pessimistic views: European and American AI is an elite category, and is "born to be a king". Coupled with the domestic GPU embargo, chip blockade, and "lithography machine bottleneck" and other objective realities, people believe that Chinese AI has lost before it is even born.

The popularity of ChatGPT is just like AlphaGo seven years ago. AlphaGo does not mean that AI can only play Go, and ChatGPT does not mean that AI can only chat. Their real goals are both in the engineering field. The reason why the market is polarized is that there is no cognitive system of its own, and a lack of deep understanding of the technology industry. What you see and hear can only be regarded as the tip of the iceberg of the industry. As a senior industry practitioner, it is necessary to disenchant "ChatGPTs" when they achieve the Renaissance.

First of all, the success of ChatGPT is due to the Transformer framework, but the Transformer framework was not developed by OpenAI. The framework originated from a scientific research paper in 2017, "Attention is all your need". Currently, there are more than 130,000 models of the Transformer open source framework on Hugging Face, and it is still increasing. The GPT-3 model on which ChatGPT is based was discovered in 2021. Google and Meta are in an open source state, so no matter which company does ChatGPT, the theoretical basis is the same. GPT-4, Wenxin Yiyan, or other domestic major products on the way "exist justifies".

Secondly, to "demystify" general AI, we must return to the essence of technical engineering. The big model is just a language model. When we get the big language model, we cannot make any logical judgments. In industry practice, whether it is industry or finance, in-depth logical judgments are required, and these logics are in our human knowledge structure. If we want to build these knowledge structures into knowledge graphs, the cost is very high. In the past, we may have used a bunch of inductive methods to state a bunch of logic, but now we can use connectionism.

For example, Sinopec needs to organize more than 4 million scientific and technological documents, which are used to support the company's logging and surveying related businesses. However, these contents are highly specialized, and the knowledge abstraction cannot be expressed in a simple chat form. This is when the role of general AI is reflected. First, thousands of documents are thrown to professionals for semantic annotation. What they care about and what relationships they have are labeled and annotated, and then migrated to a large model with 480 million text data volumes for definition.

For example, when we are working for securities companies on PNN fundraising prospectus element identification, treasury bond yield forecasts, and financial embellishment models, we need to label and extract the content. After the extraction, the content is put back into the structured knowledge graph to quickly build the knowledge graph.

▲ Nowadays, GPU boards are no longer the only choice for computing center infrastructure

Finally, the algorithm has been iterated, and the logic of the emerging underlying algorithm has changed the computing environment of the hardware. In actual tests, the speed of DNN training using commercial CPUs can be 15 times faster than that of top-level GPU training. GPUs can be replaced by CPUs under training methods that are not limited to matrix multiplication. Therefore, the GPU embargo may not be a problem. Breakthroughs at the software level and underlying logic can actually break many boundaries.

ChatGPT is a product of the information age, and its emergence will inevitably lead to an improvement in information efficiency, but there is no need for us to over-interpret it. Because information itself is diverse, once it is guided by your needs, it can only serve the goal, just like the wave-particle duality.

So you will find that a portal website with a wide variety of products is cruelly eliminated by a simple search box. Even more algorithms based on probability methods are beginning to be replaced by methods that determine the a priori and the a posteriori, such as Taobao's product selection. However, this will also bring another problem: the high probability of low-probability events, and finally, we weave ourselves in a cultural cocoon.

One of the books that influenced Steve Jobs is called "Zen and the Art of Motorcycle Maintenance". The core of the book is about the connection between two cross-borders, so we must balance rational thinking and emotional pursuit. When new tools, new modes, and new technologies appear, we must learn "how to use them correctly" while sighing in despair. The essence of AI is to give traditional industries the confidence to implement their own artificial intelligence, otherwise artificial intelligence will be demonized.

Produced by: Science Central Kitchen

Produced by: Beijing Science and Technology News | Beijing Science and Technology Media

Welcome to share to your circle of friends

Reprinted by public accounts, newspapers, etc.

Please send an email to [email protected]

Reproduction without authorization is prohibited, and offenders will be prosecuted