Large Language Model chatbots are a specific type of chatbot that is designed to carry an ongoing, evolving conversation with the same person.

       In contrast to simpler rule-based chatbots, LLM chatbots try to process large chunks of information, maintain conversation history and context, and infer from prior conversation context in order to offer more comprehensive responses to open-ended dialogues. Rule-based chatbots, on the other hand, work based on specific keywords or commands only, or give you speech menus based on decision-trees (think the phone menus you get when you call your bank), offering only pre-defined, fixed answers.

       LLM chatbots differ in complexity, quality, strengths, and weaknesses. Individual LLM chatbots are often consciously tailored to be more suited to specific types of conversations and uses.

       Like many other LLM applications, most LLM chatbots are based on deep learning models that are pre-trained on vast amounts of data. The model then tries to identify patterns, and predict reasonably probable associations between one piece of information and another. While predictive text looks at dictionaries and sometimes the user’s input history to guess one word or phrase at a time, LLMs are meant to generate large spans of complex text, mimicking reasoning-based human language.

       Unlike standalone LLM applications like ChatGPT, LLM chatbots specifically try to maintain and draw on conversation history. Chatbots are meant to be more personable and to act as if they have identifiable personalities, endearing them to users. However, despite being designed to have “distinct” personalities in mind, they rarely push back against user responses, unless specifically designed to (e.g. an LLM app programmed to respond with the phone number of a suicide helpline if the user seems to be threatening suicide).

       Cleverbot, the online chatterbot that took the Internet by storm in 1997, was not pre-trained on any data, and is not designed with the transformer architecture that modern machine-learning is based on today. Instead, it takes the conversations with humans it has had since it launched online, and regurgitates user input. For example, when Cleverbot manages to say, “How are you today?”, then whatever you respond to it, it adds to its own database of potential responses to whenever a user asks it, “How are you today?”

       While my website focuses primarily on one-on-one, AI-assisted roleplaying sessions, there are quite a few use-cases for chatbots these days. Some of them are:

• Comprehensive, lore-heavy roleplaying campaigns involving multiple AI characters
• Supportive, 24/7 accessible companionship
• Sexting (AI girlfriends)
• Ending the human race
• Editorial and research support for writers
• Pre-diagnosis healthcare intake
• Post-diagnosis treatment advice
• Secretarial support
• General customer service
• Fiction tailored to the user’s highly specific interests (minimal user input)

       LLM chatbot applications and platforms sometimes come with other features, in addition to text generation. These include:

• Image generation (e.g. AI selfies; generate an image of an environment based on the text description)
• Text-to-speech (voice memos)
• Analysis of user-submitted images
• Social / community features, such as user-submitted lorebooks
• Customisable, saved user personae for interaction with different chatbots
• Support for third-party add-ons

• Make your own professional-looking world guidebook at Natural Crit’s Homebrewery.

• Customise your roleplaying game experience with Silly Tavern front-end.

• Have your chatbot automatically generate images based on the current environment by inserting an immersive HTML prompt.   (Mirror)
• Fine-tune your TTS model on this 1000-hour NSFW TTS dataset in high-res 24kHz FLAC, by setfunctionenvironment.