2023/04/20 | 4mins

Sam Altman, CEO of OpenAI and father of ChatGPT, appeared on the podcast of artificial intelligence researcher Lex Friedman and talked about GPT-4 and AGI (Artificial General Intelligence). In this content, we summarize Sam Altman's daring to imagine the future that the development of AI will bring.

The Future of AI Seen with GPT-4 and ChatGPT

Sam Altman says there will come a day in the not-too-distant future when people will think of GPT-4 as a very early artificial intelligence . Although GPT-4 still has limitations in performing many tasks perfectly, if it develops gradually in the future, it will become something that plays an important role in our lives. It was the same when computers were first introduced into the world. Compared to the past, today's computer performance has advanced dramatically, and it is as if it is difficult to imagine everyday life without computers.

Then, if a Wikipedia page dealing with the history of artificial intelligence 50 years from now lists a turning point in AI development at this point, what GPT model can be recorded in history? To this question, Sam Altman chose 'ChatGPT' . He says the development of AI is exponential, so it would be more appropriate to view it as a continuous curve rather than pinpointing any specific event as a turning point. However, he added that the reason why ChatGPT can be selected among the models released so far is because he thought “usability” was important. Of course, the basic model is important, but RLHF (Reinforcement Learning with Human Feedback) and the interface are both important factors.

ChatGPT and RLHF

ChatGPT is a pre-trained model using large-scale text data, and can be used for various natural language processing tasks with a conversational interface. ChatGPT understands and analyzes the sentences you type and predicts the next word based on that. When the process of predicting the next word is gathered in this way, the sentence is completed and sent to the user as an answer. Therefore, it specializes in expressing and using the language.

ChatGPT, which operates on this principle, has improved its performance by utilizing 'RLHF (Reinforcement Learning from Human Feedback)' .

RLHF is human feedback-based reinforcement learning, which improves the accuracy, relevance, and fairness of large-scale language models such as language GPT-4 through human feedback . For example, when a deep learning model is trained with a large text dataset and the model shows two outputs based on this, a person gives feedback on which output is better. Open AI is said to be putting a lot of effort into gathering pre-training data from various sources to create better models.

RLHF also played a positive role in enabling ChatGPT to understand users' questions and have natural conversations. Sam Altman noted that while building RLHF models like ChatGPT is just the beginning, reinforcement learning makes it possible to make models much stronger with less data than before. In addition, through this process, even after a small amount of deep learning model training, it is possible to identify the characteristics of a fully trained deep learning model in advance, just as a one-year-old can know what kind of performance a one-year-old child will perform on the SAT test in the future. This is possible because we investigate and analyze various aspects of the model in detail.

GPT-4 AND AGI

THERE IS A SAYING THAT “GPT-4 IS THE MOST COMPLEX PIECE OF SOFTWARE MANKIND HAS EVER CREATED.” IN FACT, GPT-3 AND GPT-3.5 ARE OPERATING WITH 175 BILLION PARAMETERS (PARAMETERS), AND THE NUMBER OF PARAMETERS FOR GPT-4 IS NOT DISCLOSED, BUT IT SEEMS TO BE BASED ON A MUCH LARGER AMOUNT THAN THIS.

However, Sam Altman says that in a few decades, models with hundreds of billions of parameters will be of a size that anyone can easily create in everyday life . Of course, it is still very complicated to adjust the size of this parameter because GPT is equivalent to compressing all of the human beings on the Internet through learning. How far can we make AI human-like by training it from this massive amount of online data? Artificial intelligence with general human intelligence that can successfully perform any intellectual task that humans can do is called 'Artificial general intelligence (AGI)'.

Sam Altman said many things will be needed to make AGI going forward , but the 'Large Language Model' (LLM) will play a part. The paradigm of GPT must be expanded to make new scientific discoveries possible. What will happen if GPT develops further and permeates deeply into our society? AI has already entered our daily lives, such as programming using GPT, but in the future, AI will be the most useful tool to further amplify our capabilities and automate many tasks.

This will greatly improve the quality of our lives, but on the other hand, there are concerns about AGI. Eliezer Yudkowsky, an American artificial intelligence researcher, says that if AI has superintelligence, AI alignment, which adjusts AI to the designer's purpose or preference, will be almost impossible. In response, Sam Altman said it is important to acknowledge the possibility that these concerns may arise, and that new technologies should be researched and developed to address potential threats.

Due to the rapid development of AI, the term 'AI takeoff' has also appeared. The current AI Safety has not yet been sufficiently updated, so we also emphasized the importance of improving people's understanding of technology, improving it, trying it out, and preparing for AI alignment.

ChatGPT, which brought the AI craze around the world, took just 5 days to amass 1 million users. This is a record for the fastest growing online service in history. How will our daily lives change in the future due to AI that develops day by day? We are now entering a phase where a specific definition of AGI becomes important. As Sam Altman mentioned in this talk, it seems necessary to think about building a safe AI model development method and guidelines so that we can create better AI models and at the same time prepare for situations that may arise due to rapid development.

←Back to Blog