ChatGPT Might be the Killer App Operators Are Waiting for

GPT-3 is a natural language processing (NLP) model developed by OpenAI. ChatGPT is a fine-tuned version of GPT-3, boasting 175 billion parameters and a training set of more than 10 trillion words. Within just a week of becoming available for public testing, the new chatbot passed one million registered users. But is ChatGPT really just a chatbot? It is my view that ChatGPT utilizes original insights to satisfy deep-seated needs for answers and companionship. It does not require extraordinarily strong networks, but it does bring significantly increased uplink traffic for operators, creating 13.5 GB in dataflow of usage (DOU) per month, which is comparable to YouTube. ChatGPT's incredible hunger for data means a huge demand for computing and networks, and it may be the next killer app that we have been looking for.

ChatGPT is an NLP-based chatbot that, in the short term, is likely to replace Google as part of the next generation of search technologies. By leveraging AI video recognition (already a mature technology), ChatGPT is able to quickly understand individual users and serve as a personal AI assistant, which will help widen operators' uplink data pipes. In the future, when multi-modal AI applications evolve from image to video, ChatGPT will be used to quickly produce videos. This will mean a further explosion in the data traffic going through pipes.

I. Early chatbots "spoke" in smart ways but were uncreative and unproductive

In 1966, the first chatbot, ELIZA, was born at MIT, and was designed to appease patients in clinical treatment by imitating a psychologist. This development actually had very little to do with intelligence. Technically, it followed established rules to give pre-programmed replies according to keywords. The replies were skillfully designed to make conversations look meaningful enough to be worth continuing. Two ground rules were set:

• Identify keywords and translate them into established statements so that the chatbot appears to understand the semantics of what's being said (while it actually does not). For example, when I type "I want to ___", it replies "What does this mean to you if you ___".
• Syntactically change personal pronouns such as switching "my" to "your".

If the word "mother" was included in the input, ELIZA would reply, "Speak more about your family." If the user said, "I'm depressed much of the time," ELIZA would reply, "I am sorry to hear that you are depressed." It was a clever way of imitating real conversation, but cannot be regarded as artificial intelligence.

Any technology that uses rules to allow a robot to understand human language can be regarded as NLP. This is not particularly difficult when using computers, which gave rise to the natural evolution of ELIZA to ALICE (Artificial Linguistic Internet Computer Entity). ALICE was more powerful than ELIZA, and had much more versatile applications than simply appeasing patients. It offered thousands of possible responses and could also store conversation history (with limited learning ability). But regardless, ALICE's answers were still manually prepared, and so was the algorithm that determined how it responded (such as being proactive or passive). ALICE may be regarded as narrow AI, but, like ELIZA, it was merely imitating human dialogue and essentially playing the role of a non-player character (NPC) in gaming.

II. Initial success of deep learning-based chatbots in the business market

To act like a human, a robot needs creativity and logical inference capabilities. Creative, strong AI is the direction in which chatbots are evolving. This will be characterized by the ability to create something beyond what is manually preset. Over the past two decades, a variety of AI chatbots based on deep learning, mainly neural network algorithms like recurrent neural network and hierarchical recurrent neural network, have been developed. These chatbots (or programs) were first applied in the business market.

Today, over a hundred AI chatbot platforms are available that can be used for application development to help businesses build powerful, efficient, and scalable chatbots. Examples of such platforms include Meta's MobileMonkey, which integrates Facebook messenger, and Google's DialogFlow chatbot development framework. Chatbots can help businesses provide more personalized experiences and exclusive services to customers, which is crucial for enhancing overall service quality and efficiency. AI chatbots have become a disruptor in the way products and services are delivered globally in some business-oriented markets, including e-commerce.

III. ChatGPT: A potential AI assistant for the public

ChatGPT and its subsequent variants will be presented with a much larger market — the consumer market. ChatGPT's current popularity is comparable to that enjoyed by Meta's metaverse strategy a year ago. ChatGPT has displayed impressive performance in two key areas:

The first is proactive reasoning. When I spoke to ChatGPT about cats and dogs, it was able to reason that I was unlikely to have a cat named Puppy. The conversation went as follows:

Q: I have a dog whose name is Mew. I also have a cat whose name is Puppy. So, who is Mew?

A: Mew is the name of your dog. Dogs are domesticated mammals, often kept as pets for their companionship and ability to perform tasks. The name "Puppy" is often used for a young dog, so it is likely that your cat's name is not actually Puppy.

The second is creativity. At my request, ChatGPT was easily able to write a program to search for short videos about the 2022 FIFA World Cup final.

So, what can ChatGPT do? Given its current performance, employing it as a search engine may be a good use case. ChatGPT at least has the capabilities to replace and outperform traditional search engines.

How does ChatGPT differ from a Google search?

First, ChatGPT has a shorter information search chain. ChatGPT provides direct answers in simple sentences. This is a contrast to what Google does, where long lists of hyperlinks are provided that often look like gibberish and require us to find answers within the large amount of information.

Second, ChatGPT interacts semantically with users rather than through data (or narrow-sense information). ChatGPT is friendlier and better at understanding users and thus more likely to be accepted by the general public, giving it the potential to create huge value in the consumer market.

Third, ChatGPT can explore users' deep-seated needs through interactions. When users are struggling to accurately describe their question, ChatGPT can explore their real needs through authentically interactive methods and provide the desired answers.

Finally, ChatGPT has incredible creativity in areas like providing new ideas, developing business plans, reasoning, and making judgments. This means it not only searches for answers, but "produces" answers.

AI Generated Content (AIGC) may be yet another application scenario, where ChatGPT can create new content and free us from monotonous content creation work. However, the real killer application may be an AI assistant for the general public that can meet needs for knowledge and companionship. Behind all of this exists a gigantic market roughly the size of the entire operator industry.

IV. ChatGPT doubles uplink traffic as AI starts watching videos

Why do we need an AI assistant that can seamlessly communicate, provide services 24/7, never complain, and record our lives while constantly learning and evolving? To a certain extent, an AI assistant like this reflects human nature and certain needs that have gone unfulfilled due to technical and cost limitations. ChatGPT may meet these needs and be of great value to individual users, telecom service providers, operators, equipment suppliers, and cloud service providers alike. ChatGPT may achieve this in a number of ways.

First, ChatGPT provides individual users with answers, security, and companionship. With access to the appropriate network and cloud services, anyone can have a lifelong personal assistant who perceives and understands the world through a camera lens, a capability that may well become more important than our phones, wallets, and keys. Without the need for AR/VR or 6G, such an AI assistant can instantly make the unknown accessible to us through the application of regular cameras and connectivity. The AI assistant can constantly scan any objects we see and we can ask it questions of any kind: What species is the bird that just flew by? Where can I buy the coat that guy is wearing? What's the song I just heard? Are there saturated fats in this cup of milk tea? We may not have asked these questions before, not because we weren't curious, but because constraints repressed our curiosity. When a real-time, friendly, and customized method for getting answers and suggestions exists, no one will refuse it, as this is exactly what we need.

Second, while making the world friendlier and more accessible, ChatGPT can facilitate new growth for both telecom service providers and operators.

• The computing power of the personal assistant must be cloud-based, ensuring applications do not require powerful devices or high bandwidth like AR/VR applications do. However, the personal assistant will require connectivity for video (at 720p–1080p) anytime and anywhere, so that it can identify any objects within sight. Assuming each user uses the application one hour per day, at a bandwidth of 1 Mbit/s, the monthly DOU will be 13.5 GB, which is approximately the same as that of a heavy YouTube user.

• Current ChatGPT usage costs single-digit cents per session. Assuming that future services cost just one cent per session, with two inquiry sessions being made per minute for one hour each day, the service will be a golden egg that generates an average revenue per user (ARPU) of US$36 per month .

Third, ChatGPT expands the market potential for equipment suppliers and cloud service providers. The AI assistant places higher requirements on networks and cloud computing, and the extra 13.5 GB of DOU will undoubtedly boost network construction. GPT-3 also represents one of the world's largest neural networks for natural speech applications. To produce highly realistic and diverse output, OpenAI fed GPT-3 over 45 TB of data. OpenAI uses the Azure public cloud to provide the computing power required by GPT-3. More than 10,000 GPUs are used for training alone, with GPU power consumption exceeding 2,400 kW. If the number of users being served expands to over 1 billion, image and video recognition and the GPT engine may require millions of processors.

V. ChatGPT: Potentially the next super consumer application of operators

Three capabilities are required for the AI assistant to become the next super consumer application: Provide truly inspiring content, be user-friendly, and present results in real time. ChatGPT does not yet possess these capabilities, but it is close.

In terms of content, ChatGPT can provide original content, information, and suggestions. Second-hand content is boring, and original insights are more important, as they serve as the basis for a personalized and exclusive experience.

In terms of friendliness, it is not just about an easy-to-use interface. "Friendly" refers to how well the chatbot understands users and their semantics, as well as how it responds based on this understanding. Popular, commercially-available smart speakers are already a major overhaul over traditional speakers, all thanks to a few command input innovations. A personal assistant must have analytical and reasoning capabilities to infer based on user semantics. If questions are unclear, it should be able to guide users through a clarification process. ChatGPT already possesses this capability.

In terms of presenting results in real time, there should be mere seconds between a person's query and receiving the answer, with the optimal time being 0 to 2 seconds. This is the normal response time in human interactions, which is also about the same as the average loading time when opening a webpage. This far exceeds the millisecond-level time required for real-time 3D rendering, but is much less than the minute-level process of using services like Google to find information. The response time of ChatGPT in text already meets this requirement.

ChatGPT is expanding the consumer base of video traffic from human users to machines. It absolutely has the potential to become the next killer application for the telecom industry.