June 2023 Summaries

What is a multidisciplinary conversational AI team (CAI)? A multidisciplinary CAI dream team includes conversation designers, NLU designers, response designers, developers, product managers, product owners, and customer support people, all part owners of the platform, allowing for diverse perspectives, better coverage, and fighting against product bloat. This approach enables smarter decisions, greater innovation, faster communication, and clearer decision-making by balancing user needs with business objectives, ultimately leading to more creative teams that can drive meaningful outcomes.

Voiceflow has released a new feature called Instant Assistant that allows users to build custom voice assistants from scratch or integrate it into their existing workflow. The feature offers several benefits, including safe experimentation with large language models (LLMs), accelerated workflows, supercharged customer conversations, and reduced operating expenses. Users can choose from various LLM options, such as GPT-3, GPT-3.5, GPT-4, Claude, and Claude Instant. To get started, users can follow three steps to build an instant assistant from scratch or integrate it into their existing workflow using a knowledge base data source. The feature is available on Voiceflow's free-forever plan, making it accessible to businesses of all sizes.

The four foundational layers of conversational assistants are world knowledge, company-specific knowledge bases, dialog managers with goals and actions, and saved customer conversation context, which work together to provide a more effective and human-like experience for users. LLMs can fill the gap in general world knowledge but need to be filtered through the company's context to prevent misinformation. The combination of these layers enables assistants to capture user goals and context, providing a better interaction that is dynamic and satisfying, rather than rigid or dead-end. However, building one's own LLM may not be realistic for most companies due to the large amount of data required and the complexity of training it, making pre-existing LLM technology a viable option.

When using Claude, a conversational AI builder in Voiceflow's dashboard, expect a fast and concise experience that performs well when given organized information. The tool is designed as a methodical model, pairing well with the Knowledge Base feature or structured lists. This is just the beginning of multi-LLM vendor support in Voiceflow, with users encouraged to share their suggestions for additional LLM model vendors on LinkedIn.

A product manager needs to push their team to roll out features that can and will break the existing bot, as this is the only way to keep the team creative and the customers delighted. By doing so, teams learn from what breaks and what works, leading to a far better end experience. Good product managers also communicate wins to the larger team, spreading the word widely about quick wins across teams. To visualize the conversational AI journey, product managers should create a timeline that maps out capabilities or experiences within the assistant, using frameworks such as Brian's horizons framework to plot out informational, transactional, and conversational horizons. By doing so, everyone on the team is aligned with the user goal in mind, making it easy for teams to make decisions about new requests. Ultimately, product managers should think of conversational AI experiences as products rather than just features, pushing the boundaries of experiences and carrying out their role efficiently.

ChatGPT is a faster alternative to GPT-4, ideal for cases requiring quicker user responses, chatty interactions, or cost-effective solutions for longer responses. It's suitable for chatbots, knowledge bases, entity extraction, and intent classification, but may not be as effective for complex tasks or those requiring structured formatting. In contrast, GPT-4 excels in more challenging tasks, such as exams, factual questions, and logic problems, while also providing better performance on tasks involving vision and images. However, when it comes to tasks like classification, code generation, or summarizing large chunks of text, GPT-4's capabilities are often surpassed by other models, like StableDiffusion or DocumentAI. Ultimately, the choice between ChatGPT and GPT-4 depends on the specific requirements and goals of the project.