What is the future of generative AI beyond chat interfaces?
Chat is just the beginning. The future of generative AI lies in integrated, multi-model systems built for complex, real-world tasks that go well beyond conversation.
Conversational AI — A meteoric rise!
It feels like just overnight, chat-based models like ChatGPT have become synonymous with user-friendly interactions. The simplicity and intuitiveness of chat interfaces have propelled their popularity, catering to non-technical users and providing a comfortable way to interact with AI systems. This has been followed by rapid advances in large language model (LLM) capabilities, enabling many companies to build question-answer systems on top of their proprietary data, in some cases utilizing tools for a variety of tasks.
As a result, consumers have gotten used to interacting with these systems for customer service and personal use. As we delve deeper into the capabilities (and limitations) of chat-based AI, it becomes evident that the future of Generative AI (GenAI) extends beyond the confines of casual conversation and that business value will scale as we push beyond the chat interface.
Limitations of chat for extended conversations
The prominence of chat-based AI models like ChatGPT reflects a shift towards accessible and intuitive user experiences. The natural feel of the chat interface has undeniably contributed to their popularity.
In addition to the naturalness afforded by the chat interface, the ability of LLMs to follow instructions enables them to complete a wide range of natural language and reasoning tasks with impressive accuracy. For example, a user can ask an LLM-powered chatbot to recommend a menu for a party they are hosting, or they could instruct it to rank a selection of recipes based on their difficulty. That said, asking questions and giving instructions are not the only ways we want to utilize language technology. Some tasks, like summarization which requires looking at very large documents, do better outside the confines of the chat model. Others, like classification, are simpler and don’t require a large, general-purpose model.
LLMs are powerful, but many tasks can be carried out just as accurately with smaller models that offer greater efficiency. Tasks requiring a balance between efficiency and generality, such as synthesizing and summarizing documents, showcase scenarios where both small and large models find their utility.
The rise of GenAI has marked a shift in language technology from merely processing, structuring and enriching data, to generating content. While this generation reflects existing data, it often does so in the aggregate and has the potential to omit key pieces of information, draw on bias present in a dataset, or generate new data points of its own. These types of errors can be extremely damaging, and compliance-driven fields will be slower to roll out Generative AI products that use LLMs.
Beyond chat: The role of LLMs
LLMs have emerged as the dominant force powering a chat-centric user experience. These models, with their exponential growth in parameters, have a historical context rooted in their capacity for complex language processing. However, the argument surfaces that LLMs may be overkill for certain tasks, and the chat interface may not be the optimal solution for every use case. For example, many natural language processing tasks involve ingesting or writing very large blocks of text or even code. Doing so in a conversational setting is unnatural and we can imagine many interfaces that are more amenable to a copywriter, programmer or analyst’s workflow.
LLMs have also shown promise as pieces in larger systems that incorporate other forms of input. While many of the most popular LLMs focus on textual input and output, we have begun to see these systems integrated with models that process or produce voice, images or videos. As we see GenAI become more and more prevalent, we can expect it to transcend a single modality and instead be able to interact with information in all different forms.
The future: System integration and efficiency
Looking ahead, the future of GenAI lies in the integration of systems. Rather than relying solely on chat interfaces or large models, many traditional machine learning (ML) systems were composed of multiple ML components that worked in concert to provide a given experience. This approach is still relevant today, particularly because it emphasizes the importance of consistency and accuracy in generating text, which is crucial for business applications.
At Kensho, it’s important to us that we ensure accuracy for our customers and build transparency and explainability into the products we bring to market. We understand that AI is only as good as the data it leverages, so it is our top priority to ensure the integrity and utility of the data that our customers rely on. To accomplish this, Kensho and S&P Global teams are working to make our high-quality financial data LLM-ready, which means it’s accessible for model training and retrieval augmented generation (RAG). This enables us and our customers to develop trustworthy AI systems that make explicit reference to the data sources they utilize.
The demand for both accuracy and compliance in business applications necessitates a thoughtful consideration of the tools at our disposal. By creating integrated systems that efficiently handle a spectrum of tasks, the focus shifts from the limitations of individual models to a holistic approach that ensures consistency and correctness across diverse applications. For this reason, Kensho has developed many purpose-built models including Scribe to transcribe audio, Classify to identify and tag concepts in documents, NERD to recognize and annotate text with named entity labels, and Extract to transform PDFs into a machine-readable format. These products can be integrated with LLMs to provide high quality data for question-answering systems in a chat interface. At the same time, these modular components can support any range of generative AI solutions beyond chat.
While chat-based AI models have dominated the landscape, the future of GenAI appears to extend beyond casual conversation. By acknowledging the limitations of chat for extended and diverse tasks and recognizing the suitability of both small and large models in specific scenarios, the trajectory of AI points towards integrated systems. The emphasis on system integration, efficiency, and accuracy in generating text for business applications paves the way for a future where GenAI evolves into a versatile and indispensable tool in our technological arsenal.