How I built the world’s best chatbot with Steve Worswick

Picture of Kane Simms
Kane Simms

More ways to listen and watch

About this episode

We speak to the creator of the world’s best chatbot about how to design Loebner prize-winning conversational experiences.

Steve Worswick is the creator of Mitsuku, the general conversation chatbot that has won the Loebner prize for the last two year’s straight.

13 years in the making, Mitsuku convinced a panel of judges that it’s the most humanlike bot over the course of a 20 minute conversation, two years in a row, to be crowned the world’s best chatbot and conversational agent.

It’s featured in the Wall Street Journal, BBC, The Guardian and Wired. And, unlike most chatbots that focus on serving a specific set of use cases, Mitsuku is a general conversational agent. That means you can speak to it about anything.

This week’s Flash Briefing question is from Brielle Nickoloff of Witlingo: What would an open source voice assistant look like? Send us your thoughts and you could feature on the VUX World Flash Briefing this week!

What about voice?

Although Mitsuku is a text-based chatbot, this episode looks at how to take Steve’s 13 years of experience in creating conversational experiences and apply that to the voice first space.

In this episode

This episode is all about how to design and create a world-leading general conversational experience.

We get into detail about how Mitsuku is built (hint: it doesn’t use natural language processing or machine learning like most other conversational AI) and how Natural Language Processing-based conversational agents don’t quite hit the mark.

Steve tells us about Mitsuku’s rule-based supervised learning and how that’s leading to better experiences.

Despite Mitsuku passing the Turing test, Steve tells us why the Turing test is redundant.

We discuss user behaviour and how people treat a general conversational agent, from counselling to romance, bullying to marriage and money worries, and how to be sensitive on those topics.

We hear how varied responses can increase engagement. So much so that one person has spent 9 hours talking to Mitsuku!

We find out how to deal with pronoun resolution and how to refer back to what was said earlier in the conversation.

We uncover how brands are using Mitsuku as part of their conversational experiences, handing off to her when a user strays away from the use cases that their bot can handle.

We chat about how Alexa fairs against Mitsuku as is shown in this video:

And hear where Siri would have finished if it was entered in to the Loebner prize competition.

Perhaps one of the most valuable lessons in this episode is the importance of persisting. Creating a conversational agent, a true conversational experience, takes time. It’s not a quick fix that you cobble together with a quick Alexa Skill. It takes years of development, iteration and constant improvement. But, if you stick with it, you might end up with the next best conversational agent.

Our guest

Steve Worswick started out in IT support and built Mitsuku as a passion project on the side. 13 years of hard work and 3 Loebner prizes later, he’s now working at the world’s largest chatbot agency and provider, PandoraBots.

Where to listen

Links

Contact Pandorabots
Check out Mitsuku on Pandorabots
Check out Steve’s talk at the Chatbots and Voice Assistants London event

Share

Weekly newsletter

The latest in AI-powered customer experience. Make better strategic decisions with the help of our weekly newsletter.

Host

Kane Simms

A strategic AI advisor who, for the past decade, has helped business leaders and product owners transform customer experience using conversational and generative AI.

Share

✓   Link copied

Weekly newsletter

The latest in AI-powered customer experience. Make better strategic decisions with the help of our weekly newsletter.

Top articles and podcasts

Related content

Cobus Greyling: why intents, graphs and conversation design have come full circle
Cobus Greyling on where enterprise AI is really heading: data, security, forward-deployed engineers and conversation design.
From answers to actions: building better government AI
Pete Kowalczyk on knowledge graphs, goal detection and moving AI from answering questions to taking action.
What’s really breaking your AI agents?
Discover where conversational AI projects actually break: contradictory source data, manual process and weak testing.
How Simplyhealth scaled voice AI to resolve 30% of calls
Paul Atkins, Head of Business Optimisation at Simplyhealth, breaks down how they scaled voice AI to resolve almost a third of calls, and what it took to get there.

Register now: Why your contact centre playbook is obsolete