Analysis of the impact of ChatGPT on the development of social robot technology

Analysis of the impact of ChatGPT on the development of social robot technology

Abstract As a typical representative of generative AI, the new pre-trained basic model ChatGPT has become a "phenomenal" hot topic of discussion. This paper comprehensively reviews ChatGPT and its impact and inspiration on the development of social robot technology. First, the concept, main features and development history of ChatGPT are introduced; secondly, the current development trend of artificial intelligence technology reflected by the pre-trained basic model represented by ChatGPT is discussed; finally, the role of the recent rapid development of generative AI in promoting the research and application of social robots is analyzed. The review shows that the strong demand of social robots for the ability to generate content autonomously makes them closely follow the development progress of generative AI. In the future, they will continue to develop in personalized content generation, identification and control of generated content, and traceability of "emergent" capabilities.

1 Introduction

ChatGPT, launched by the US company OpenAI, has almost become another "milestone" node and "disruptive" innovation in the field of global artificial intelligence since AlphaGo in 2016, and is expected to usher in a new round of accelerated development of technology-driven industrial transformation. As a landmark achievement in the transition from "dedicated artificial intelligence" to "general artificial intelligence" [1], ChatGPT has been included in the watch list of the US Defense Information Systems Agency (DISA) in January 2023, which deserves high attention. On February 24, 2023, based on in-depth thinking and medium- and long-term development planning for the large-scale pre-trained basic model represented by ChatGPT, OpenAI released the "Planning for AGI and beyond" (as shown in Figure 1), believing that general artificial intelligence may become a powerful "amplifier" of human intelligence and creativity.

Figure 1 OpenAI releases the "General Artificial Intelligence Development Plan"

Fig.1 Planning for AGI and beyond from OpenAI

2 Conceptual Characteristics and Development History of ChatGPT

2.1 Conceptual Connotation

ChatGPT is a brand-new pre-training foundation model [2-3], which is mainly aimed at dialogue tasks. Based on the traditional pre-training foundation model of the "pre-training + downstream task fine-tuning" mode [4], it introduces advanced strategies such as reinforcement learning based on human feedback, and autonomously learns "world knowledge" from massive data. It has successfully achieved a qualitative leap in complex natural language understanding, "personalized" human-like language generation, multi-round human-computer capabilities, and has reached an unprecedented level of accuracy in understanding human thinking. It has become the fastest growing consumer application in history (accumulating 100 million users in 2 months).

2.2 Main features

The pre-trained basic models represented by ChatGPT have become the mainstream direction and the highest level of current artificial intelligence development. They are expected to become the strongest "booster" for accelerating and upgrading the research and development and application of artificial intelligence at this stage and in the next 3 to 5 years. The key to its success mainly comes from the following three aspects: First, it has a huge base model as support. The GPT series models based on ChatGPT are all base models with a huge parameter scale. It has undergone three major iterations, and the number of parameters has increased from 117 million to 175 billion [5]. The industry estimates that the number of parameters of ChatGPT is nearly 200 billion; second, it has high-quality real data as supply. OpenAI has made great investments in data accumulation, cleaning, and annotation for several years. The volume and quality of pre-training data are far ahead. The industry speculates that the scale of ChatGPT's training data is as high as 800 billion words; third, it has powerful computing resources as a guarantee. The computing resources that drive ChatGPT are very expensive. The cloud service provider it relies on, Microsoft Azure, provides about 10,000 (graphics) and more than 285,000 (CPUs).

2.3 Development History

Since 2018, OpenAI has released four major versions, namely GPT-1, GPT-2, GPT-3, and InstructGPT [6]. Among them, the hallmark of GPT-1 is unsupervised learning, the hallmark of GPT-2 is multi-task learning, and the hallmark of GPT-3 is massive parameters (as shown in Figure 2). After GPT-3, all models of OpenAI are no longer open source. The existing analysis and research on ChatGPT is mainly based on technical speculation of historical versions. OpenAI transformed from a non-profit to a profit-making company in March 2019, and successively received seven rounds of financing from Microsoft and other institutions. Among them, Microsoft's total investment amounted to billions of dollars. Adequate funds are an important condition for the success of ChatGPT. In 2022 alone, OpenAI's R&D investment exceeded US$544 million.

Figure 2 The development of ChatGPT and its historical versions

Fig.2 The development of ChatGPT and its historical version

3 Analysis of the significance of ChatGPT to the development of artificial intelligence technology

3.1ChatGPT Technology Positioning and Application Analysis

ChatGPT is an evolution of technological development, not a cross-generational disruption, but it will cause a new round of revolution at the application level, and its influence is no less than the birth of the Internet.

(1) Technical analysis

ChatGPT is a victory of the technical route. Since 2018, the BERT model of Google and the GPT series model of OpenAI have become the two mainstream directions of pre-training basic models. Among them, the training method of BERT is to let the model make two-way guesses based on the context to achieve self-supervised learning of unlabeled text, which was better than GPT for a long time; GPT is continuously iteratively optimized by one-way prediction, parameter upgrade, repeated training and generated content feedback, and is now significantly better than BERT. At present, the performance of ChatGPT is still based on massive data training and statistical analysis. The technology has not yet deviated from the category of "weak artificial intelligence". Although its performance in some text applications (such as content generation) and repeated code construction has exceeded most artificial intelligence, it still has not solved the technical problems such as complete "world knowledge" and deep cognitive reasoning. Therefore, although ChatGPT presents "intelligence", it does not have "intelligence".

The pre-training basic model technology represented by ChatGPT will develop in the following directions: First, multimodal fusion processing, effectively processing multimodal data and cross-modal information such as images, language, and voice; second, the ability to interact with the physical real world, exploring embodied intelligence, physical common sense, abstract world models, perceptible four-dimensional physical space, etc.; third, the timeliness of information is significantly increased, not only can it draw nutrients from historical data, but also can dynamically obtain and instantly analyze real-time information; fourth, the ability of careful logical thinking, mathematical solution, long-term thinking, and complex task processing is further improved.

(2) Application research

ChatGPT technology has broad application prospects, and it is urgent to actively seize the high ground. Taking its application in the military field as an example, the US Department of Defense's "Third Offset Strategy" points out that "the rapid development of artificial intelligence, together with technology, autonomy, big data, and enhanced cooperation with the community, will define the next generation of warfare." Deploying ChatGPT, an interactive artificial intelligence chatbot, on a reinforced edge computer is expected to play a key role in a multi-domain combat environment, providing soldiers with the real-time information, situational awareness results, and decision-making assistance they need to effectively coordinate actions in all areas of the battlefield. Therefore, it has strong application potential in massive information retrieval, deep acquisition and processing, multimodal intelligence processing and cognition, combat planning and scenario decision-making, and network confrontation. However, at present, such pre-trained basic models have no way to guarantee that the answers are stable, reliable, and correct, which poses a great risk to military applications and requires further research to solve.

3.2 Analysis of the main development trends of artificial intelligence technology reflected by ChatGPT

The extensive and in-depth successful application of pre-trained basic models has profoundly explained the main development trends of generative artificial intelligence [7] and also pointed out the direction for the development of artificial intelligence.

(1) Models, computing power, and data remain the basic factors that influence AI capabilities

The key driving force for ChatGPT's success mainly comes from computing power, data, and algorithms, which fully demonstrates that computing power, data, and algorithms are still the decisive factors that determine the height of artificial intelligence capabilities. Large-scale and advanced architecture models, powerful computing resources, and huge amounts of high-quality data resources determine the height that the new generation of artificial intelligence can reach with pre-trained basic models as the "wind vane". Domestic research institutes and enterprises have a certain foundation in large language models, corpus data sets, and domestic computing power, but there is still a gap compared to ChatGPT. Taking data as an example, there is no shortage of data in China, but there is still no effective mechanism for cleaning and labeling massive heterogeneous data. It is estimated that only 10% of the data in the tens of billions of data may be "usable", and data quality has to some extent restricted the development of my country's domestic large models [8-9].