Development of large-scale models is limited and requires the creation of explainable AI theories

The limitation of GPT-4 is that it can only interact with the digital world, and we ultimately need to interact with the physical world. For this reason, the emergence of robots is particularly important, as it represents the realization of embodied intelligence. Zhang Bo pointed out that it is not necessary to develop humanoid robots, nor does it require overly complex hardware. He advocates conducting reinforcement learning research based on certain hardware. Such an approach allows fine-tuning without changing the original meaning.

The whole world is surprised by the powerful capabilities and potential of large models, but cannot explain the reason and can only attribute it to the "emergence" phenomenon. For the healthy development of the artificial intelligence industry, scientific research, technological innovation and industrial development must be integrated. To develop the third generation of artificial intelligence, explainable and robust theories and methods must be established, otherwise AI technology will never be convincing.

There is a ceiling for large language models.

Although the road to general artificial intelligence is still difficult, large language models have opened up a broad path for the AI ​​industry. At the Zhipu AI 2024 Annual Technology Open Day, Academician Zhang Bo said that large models provide opportunities for the development of general hardware and software.

The traditional AI paradigm uses specific algorithms and rules to complete specific tasks. The generative AI paradigm is based on a general model called the basic model. It is trained in an open domain (open domain) through extensive text data training. It can generate high-quality text, images, and other content similar to human levels, and can be adapted to a wide range of downstream tasks through fine-tuning and other methods. Generative AI takes one step towards general AI, the second step is AI agents, and the third step is embodied intelligence. Zhang Bo said that GPT-4 can only deal with the digital world. We must eventually deal with the physical world, which requires robots, that is, embodied intelligence. The proposal of embodied intelligence helps to construct a complete intelligent agent, allowing the intelligent agent to both perceive and think. "You don't have to make a humanoid robot, because in many cases you just need your hands or feet, and you don't need to make the hardware very complicated." He advocates conducting reinforcement learning research based on certain hardware.

Generative AI large models have three major capabilities and one major shortcoming. The first is its powerful generation capabilities, which amaze people with its ability to generate coherent text with a variety of context and past conversations. Secondly, it has strong migration capabilities, that is, through training and fine-tuning of agent tasks, it can be applied to downstream tasks of interest. The third is powerful interaction capabilities, including human-computer interaction, interaction between multiple agents, and interaction with the environment, allowing AI to demonstrate intelligence levels comparable to humans in various fields. However, these large models also suffer from a drawback: illusion. Sometimes they generate made-up or nonsensical answers that seem reasonable.

Artificial intelligence helps promote economic growth. Industries such as construction, maintenance, and installation are difficult to be automated and intelligent, but white-collar jobs such as administrative management may be replaced by AI. AI can improve the quality and efficiency of most human jobs, but there are still only a few jobs that are completely replaced by AI. The reason why AI cannot yet replace most jobs is because there are still insurmountable ceilings for large models. Zhang Bo said that all the work of the large model is external prompts rather than proactive, and uses probabilistic predictions to complete tasks under external prompts, while human work is driven by internal intentions. The language generated by large language models and human language generation are only similar in behavior, but the internal mechanisms are fundamentally different. Large language models have ceilings such as unawareness, uncontrollable quality, untrustworthiness, and unrobustness. Different prompt words are given. Large models will output different answers. Large models can also produce hallucinations. “No matter how big the model is, the shortcoming of hallucinations always exists.”

He proposed that to develop the third generation of artificial intelligence, we must establish explainable and robust artificial intelligence theories and methods, develop safe, controllable, credible, reliable and scalable AI technology, and promote innovative applications of AI. and industrialization. If an explainable and robust artificial intelligence theory cannot be established, AI technology will be unreliable and will never be credible. "So far, this theory has not been established, which is why the development of artificial intelligence is slow and tortuous. The reason why the theory cannot be established is because it is subject to three specific limitations. In the past, specific models could only be used to solve specific tasks in specific fields. , how is it possible to establish a general theory? The emergence of large models provides the possibility of establishing this theory."

Zhang Bo said that large models provide us with the opportunity to develop general hardware and software. Artificial intelligence is entering a stage of steady development and has a huge impact on all walks of life. We must seize the opportunity to develop the artificial intelligence industry. But there are still a lot of uncertainties because AI is unpredictable and uncontrollable. The whole world is surprised by the powerful generation, migration and interaction capabilities of large models, but they cannot explain it and can only attribute it to "emergence". Therefore, for the healthy development of the artificial intelligence industry, scientific research, technological innovation, and industrial development must be combined.

The above is the detailed content of Development of large-scale models is limited and requires the creation of explainable AI theories. For more information, please follow other related articles on the PHP Chinese website!