Conversational interfaces are all means where humans interact with digital dimensions via natural language. Assumptions: * The real world is very flexible, leading to complex, overloaded digital tools & processes trying to mimic this flexibility. * The web 2.0 and app-approach fragments even mildly complex tasks, requiring constant app and context switching. * Natural input represents a more convenient way to handle the real-world flexibility and complexity. In other words: As digital dimensions get too complex, conversational interfaces represent a way to make them intuitive again. This makes them part of [[Cognitive Services]] and [[Ambient Computing]]. Current mouse- or touch-based user interfaces also do not translate well to [[Metaverse]] or [[Augmented Reality]] scenarios. In a way, natural language user interfaces are also about future-proofing human machine interaction for a time when the current screen-paradigm is no longer prevalent. ## Natural Language Processing (NLP) The goal of the Natural Language Processing (NLP) is to analyze, understand, and generate language that humans use naturally. “Understanding” in this context means, among other things, knowing what concepts a word or phrase stands for as well as knowing how to link those concepts together in a meaningful way in order to extract the full intent of a query or generate a complete and meaningful dialog. ## Types of Conversational Interfaces It makes sense to differentiate between different types of conversational interfaces based on their individual capabilities, properties and intents. * [[Chatbots]]: Converse with users on a public or dedicated chat platform, usually to let them achieve a specific goal within a given context. * [[Digital Assistants]]: Converse with users as part of a platform or ecosystem, usually to offer them a wider range of tools and control over digital aspects of their environment. * [[Digital Familiars]]: Converse with users as part of a product, driven by [[Relationship Design]], to make them them more approachable and relatable. ## Conversational Interfaces vs. Artificial Intelligence Understanding the intent behind a natural language request does not equal intelligence. The “intelligence” of a conversational interface is determined by its capability to take the recognized intent of a user and provide relevant insights accordingly. That could mean answering a question correctly, adding new arguments to a discussion or generally carrying the conversation onwards in a meaningful way. Case & point: Large Language Models (LLMs) do not work with any [[Intentionality]], but based on pure statistics. ## Negative Feedback Loops It seems that systems with a lack of intentionality have issues when automatically improved by user feedback. An example for this was [Microsoft Tay](https://en.wikipedia.org/wiki/Tay_(bot)). Released on Twitter in 2016, users began teaching it memes and offensive themes, such as "[redpilling](https://en.wikipedia.org/wiki/Red_pill_and_blue_pill#Other_uses "Red pill and blue pill")" and "[Gamergate](https://en.wikipedia.org/wiki/Gamergate_controversy "Gamergate controversy")". The bot itself began to post offensive tweets, causing Microsoft to shut down the service only 16 hours after its launch. https://www-indiewire-com.cdn.ampproject.org/c/s/www.indiewire.com/2023/02/seinfeld-ai-banned-for-being-transphobic-1234806856/amp/ There is also a paper [[2212.09251] Discovering Language Model Behaviors with Model-Written Evaluations (arxiv.org)](https://arxiv.org/abs/2212.09251), arguing that: > *"Larger LMs repeat back a dialog user's preferred answer ("sycophancy") and express greater desire to pursue concerning goals like resource acquisition and goal preservation. We also find some of the first examples of inverse scaling in RL from Human Feedback (RLHF), where more RLHF makes LMs worse. For example, RLHF makes LMs express stronger political views (on gun rights and immigration) and a greater desire to avoid shut down."* ## Summary **What it is**: Provide interfaces for digital dimensions via natural language. **What it enables**: A more convenient way to handle the real-world flexibility and complexity, as it applies to digital dimensions. ## Articles & Links [Why is Sam Altman so obsessed with 'Her'? An investigation (bloodinthemachine.com)](https://www.bloodinthemachine.com/p/why-is-sam-altman-so-obsessed-with)