Chatbots are becoming increasingly sophisticated, but what techniques do they use to create conversations that seem so human-like?
Important Points
- Rule-based chatbots operate based on predefined conditions and keywords for generating responses, lacking the capacity to adapt to context or learn from prior interactions.
- On the other hand, AI chatbots, like ChatGPT, employ extensive language models trained on extensive datasets to emulate human-like conversations and grasp conversational context.
- Innovations in AI chatbots involve integrating artificial general intelligence (AGI) and physical embodiments, such as humanoid robots, demonstrating the potential for more interactive and engaging interactions with humans.
Chatbots have long been an intriguing and practical online tool. The emergence of AI-driven language models, like GPT-4 and the ChatGPT chatbot it underpins, has brought a fresh dimension to human-bot-human interactions. However, the question remains: How do AI chatbots recreate conversations that resemble human-like interactions? How can computers engage in simulated conversations with people?
What Exactly Are Chatbots, and How Do They Operate?
Before advanced AI chatbots like ChatGPT, Claude, and Google Bard, there existed more basic chatbots referred to as rule-based chatbots or decision-tree chatbots.
Rule-based chatbots lack the ability to adapt to situations, grasp context, or simulate human logic. Instead, they rely on a predefined set of rules, patterns, and dialogue trees crafted by developers.
When presented with a prompt, rule-based chatbots follow predetermined conditions, often utilizing keywords as crucial indicators. These chatbots scan user inputs for specific words to understand the query since they cannot comprehend context. Their responses rely on clues like keywords to generate useful answers.
Businesses frequently employ rule-based chatbots as intermediaries between customers and human representatives. For instance, when contacting an energy or cell service provider, you might encounter a chatbot that first prompts you to describe your issue. Similarly, a chatbot might appear on a website to address inquiries.
However, rule-based chatbots struggle with complex, multi-layered questions. They excel at responding to short, straightforward queries, such as “Update my account information.” Questions containing numerous variables typically exceed the capabilities of rule-based chatbots, either due to their limited ability to interpret natural language or their constrained knowledge database.
Rule-based chatbots cannot enhance their performance autonomously and require manual intervention by developers. This limitation stems from their inability to learn from past interactions.
AI chatbots also adhere to rules. For instance, ChatGPT is programmed not to use profanity or offer illegal advice. Nonetheless, AI chatbots operate and engage in ways that far surpass the capabilities of rule-based chatbots.
How AI Chatbots Work
AI chatbots like ChatGPT were not the first of their kind. Prior to ChatGPT’s rise to prominence, there were less sophisticated chatbots that also utilized AI for interacting with human users.
One notable example is Eviebot, which made its debut in 2008. Evie employs AI to engage with users and functions as a learning AI chatbot. Evie has the ability to enhance her conversational skills by taking note of interactions with previous users. Interestingly, Evie shares the same AI system as Cleverbot, another chatbot that achieved widespread popularity in the late 2000s and early 2010s.
But this chatbot is a far cry from the modern versions we use today.
As evident in the screenshot displayed above, Evie doesn’t excel in delivering accurate responses to questions or maintaining a consistent conversational history. In just a matter of seconds, the chatbot initially identified itself as Eliza but then switched to Adam in the subsequent reply.
Furthermore, Evie is not a reliable source of information. For instance, when we inquired about the size of the sun, Evie responded with a humorous remark, saying, “Bigger than my future.” While such responses can be amusing, Evie struggles to provide users with factual information, regardless of how common the queries may be. If you’re seeking a chatbot experience that leans more toward the fun and unconventional side, Evie might be an appealing choice for you.
Sites like Cleverbot and Evie undoubtedly offer entertainment value, but they are not well-suited for practical or informational purposes. It wasn’t until late 2022 that the world began to witness the incredible utility of AI chatbots.
How Are Conversations Simulated by Chatbots?
The question remains: how do AI chatbots like ChatGPT simulate accurate conversations with humans? How can they appear almost indistinguishable from a regular person sitting at a keyboard?
In November 2022, OpenAI released a publicly accessible version of its GPT-3.5 large language model, known as ChatGPT. This marked a significant milestone in AI chatbot technology, as ChatGPT demonstrated the ability to engage in highly human-like conversations. While we have a detailed article delving into ChatGPT’s workings, there are some key points to highlight here.
Firstly, the “GPT” component in the tool’s name stands for “Generative Pre-trained Transformer,” which refers to a type of large language model (LLM). These terms have been widely discussed in 2023, but what do they actually mean?
An LLM is an AI learning model used by major AI chatbots today. It operates using a sophisticated AI algorithm that leverages deep learning. LLMs are trained with extensive datasets, providing them with a vast repository of knowledge to solve problems and respond to inquiries. For example, ChatGPT-4 underwent training with an estimated 1 to 1.7 trillion parameters and terabytes of data (although the exact figures have not been disclosed by OpenAI).
A GPT is a specialized type of LLM featuring a neural network with deep learning capabilities. GPTs are pre-trained models that are exposed to vast databases of information for learning purposes. In the case of ChatGPT, this training data includes text extracted from books, journals, articles, and various other sources. However, even with access to this extensive data, how does ChatGPT manage to engage in conversations with people in such a human-like manner?
During the development of ChatGPT, a method called “reinforcement learning from human feedback” (RLHF) was employed for training. This training approach utilizes reinforcement techniques to shape ChatGPT into the desired chatbot. Through a system of rewards and feedback, ChatGPT can discern which responses are considered useful or “good” and which are not. This methodology also enables ChatGPT to better comprehend the context of a conversation, allowing it to provide more effective responses to prompts.
ChatGPT’s proficiency in natural language processing is another crucial factor in its ability to interact with users. This involves recognizing specific language patterns and sentiments. During its training, the algorithm was exposed to examples of human conversations to gain a deeper understanding of how humans communicate. It can even take note of conversational cues, such as greetings and farewells, to gauge the progression of the conversation.
How Are AI Chatbots Progressing?
OpenAI has shared limited details about GPT-5, the next version of its Large Language Model (LLM). What’s particularly intriguing about GPT-5 is its rumored incorporation of Artificial General Intelligence (AGI) into its algorithm. AGI, which theoretically simulates human cognition, has the potential to be a game-changing development.
While ChatGPT made a significant impact and continues to do so, the world of AI chatbots is not limited to OpenAI. Companies worldwide are actively working on improving their AI chatbots to engage in human-like conversations, and some are even taking chatbots to a physical level.
For instance, there’s Desdemona, a humanoid robot developed by Hanson Robotics and SingularityNET. Desdemona is known as the “sister” of the well-known robot Sophia, who has garnered attention for her human-like appearance and demeanor. Unlike Sophia, Desdemona specializes in music and is part of a band featuring human musicians. The robot uses AI algorithms to sing along with popular songs, drawing from a library of pre-existing music. Desdemona has even performed live alongside her human bandmates.
Desdemona possesses the capability to engage in conversations and communicate with people. In 2022, she was featured in an interview conducted by the YouTube creator Discover Crypto. During the interview, Ben Goertzel, the creator of her AI algorithm, also addressed questions related to AI and its future.
While Desdemona’s humorous reference to keeping humans in aquariums might be unsettling to some, her ability to provide spontaneous responses to unscripted queries highlights the potential of AI to interact with humans in a friendly and conversational manner.