At the end of 2022, a new AI chatbot named ChatGPT was launched by OpenAI, followed by the release of its next-generation model, GPT-4, in the subsequent months.
This chatbot can not only query various materials like a search engine but also create poetry and scripts, write essays, and code programs. It can fluently understand human questions and commands, and even comprehend images.
The emergence of ChatGPT has propelled a new wave of AI, sparking global discussions among the public.
Debates on whether it possesses self-awareness, whether it will break ethical rules, and whether it will replace humans are rampant. Meanwhile, ChatGPT has pushed the popularity of large models to its peak, marking a turning point in AI development, with countries competing to establish their AI industries. A profound transformation in information technology is unfolding.
Many people worry that ChatGPT will replace them; is this concern necessary?
Viewpoint One:
Looking back at history, the Industrial Revolution replaced a large number of workers with machines. Out of fear of the unknown, workers at that time smashed machines, which is akin to the panic brought by the emergence of ChatGPT today. Unlike previous instances where machines replaced humans in laborious tasks, this time, people fear being completely replaced by machines.
However, in reality, the replacement of human intellectual activities by ChatGPT may only be the tip of the iceberg. In the past, we were bound to basic labor, leading to inefficiency, and human creativity could not be fully unleashed. The emergence of AI like ChatGPT is not intended to replace humans, but to liberate them.
Viewpoint Two:
Humans are not omniscient and omnipotent. Our understanding of the world is largely achieved through tools, such as telescopes, microscopes, and large scientific facilities. The relationship between humans and tools is not one of replacement. The emergence of ChatGPT provides humans with a more powerful tool to help them leverage their advantages. This will not fundamentally change in the future.
As ChatGPT evolves, will it develop “self-awareness” and “intelligence”?
Viewpoint One:
Elon Musk once proposed a viewpoint that humans, as “carbon-based life,” only contribute to generating a program to serve as a precursor for future “silicon-based life.” However, I strongly oppose this view because the fundamental elements behind ChatGPT come from data, followed by reinforcement learning based on human feedback.
Data contains human ideologies, and adjustments are ultimately based on human feedback; therefore, the fundamental source of ChatGPT is still humans. Before breakthroughs in understanding the mechanisms of the human brain, AI cannot possess “awareness,” nor will it experience “self-awareness.”
Viewpoint Two:
The mechanism of “intelligence” can be likened to flight. In the past, when humans wanted to learn to fly, it was similar to how they now seek to understand intelligence. Previously, people observed how birds flew and mimicked them through biomimetics to create wings. Alternatively, they understood the Euler equation through aerodynamics to manufacture airplanes. Airplanes are entirely different from birds, yet they can fly faster and higher.
Today, intelligence is similar. We can understand the mechanisms of the human brain as akin to bird flight. Do we need to fully replicate the human brain to create intelligent machines? Perhaps not. From an engineering and technological perspective, we might find equations and mechanisms about intelligence itself, leading to the creation of intelligent machines that could fly faster than birds. This possibility is what concerns us and is the goal we continuously explore.
The answer to this question has eluded humanity for thousands of years, but it does not hinder the continuous advancement of AI. The emergence of ChatGPT has expanded the boundaries of human intellect, just as the steam engine made the physical capabilities of a 50 kg person nearly indistinguishable from a 100 kg person.
In the future, individuals with an IQ of 80 may also become indistinguishable from those with an IQ of 180. This is also the direction we aim to develop, step by step expanding the range of tasks we can accomplish with AI.
Viewpoint Three:
“Intelligence” is not exclusive to humans. When discussing AI abroad, they often prefer the term “machine intelligence.” However, human intelligence and machine intelligence may represent two distinct types of intelligence.
During human evolution, intelligence is not only individual but also collective. That is to say, human evolution is not merely the result of individual evolution but rather collective evolution. The emergence of language is a result of human collaboration. Language reflects humans’ structured response to the objective world and also reflects their subjective imagination.
However, all knowledge ChatGPT has about the objective world is input by humans; otherwise, it cannot perceive the objective world. The breakthroughs ChatGPT has made at the level of language do not imply a complete mastery of human intelligence.
What technical gaps must China overcome to achieve its expected results in the field of AI?
Answer:
Taking ChatGPT as an example, objectively speaking, it is not that the algorithms are particularly advanced, but it has truly achieved a three-step process: first, pre-training; second, supervised fine-tuning (SFT); and third, reinforcement learning based on human feedback. When we aim to create a new AI that matches ChatGPT, it means our goal is to develop an AI that possesses the following characteristics:
First, it should be a general-purpose AI capable of solving most tasks, including text generation, numerical calculation, Q&A, etc.
Second, it should have multi-turn retrieval capabilities.
Once the goals are clear:
First, solve algorithm-related issues. Although there are numerous papers in the field of AI, replicating a product requires a lot of experience not covered in the papers.
Second, address computing power issues. The performance of domestic chips and how to achieve parallel architecture for computing acceleration are challenges we face.
Third, address data issues. How to promote systematic design, how humans can intervene and label data, how to shorten response times, and how to reduce costs all require systematic solutions.


Content sourced from “Towards Self-Reliance and Strength – China Science and Technology Hall Forum, Volume 2”

Editor for this issue: Zhou Ting
Some images sourced from the internet