We trained up ChatGPT on their website and we continue to add to our hundreds long list of questions and answers document. We’re training all the time, as new people use it, and fresh questions are asked.

A Naughty Five Year Old

About a month into our training, we started asking the Bot questions, and quickly discovered the articulate 5 year old lurking somewhere inside of our Bot. I say 5 year old in reference to a TED talk given by Sir Ken Robinson, titled ‘Do Schools Kill Creativity?’. Somewhere in his talk he speaks of a marvelous characteristic of children to try things without the fear of failure. He described children’s behaviour in this regard as, ‘if they don’t know, they’ll have a go’. He then goes on to describe how we lose this characteristic as we grow older.

Well, our Bot ‘has a go, when it doesn’t know’. Our first insight into this was asking it for directions from Sandton City to Lanser’s on Main? It answered articulately and confidently that you needed to turn right on Rivonia Road into William Nicol Drive. And then finished the directions pretty well. What we realised is that William Nicol Drive and Rivonia Road never intersect, so you can’t perform that turn. There’s a road missing.

We appreciated that ‘it had a go’, but having a go can be disastrous in some situations. We’ll keep on watching, tweaking and training, learning the best way to get our Bot up to speed.

The post Occasionally our Bot behaves like a Naughty 5 Year Old appeared first on Calidascope.

One of our clients is a restaurant in Bryanston. Lanser’s on Main. It’s rated the best Bistro in the area on TripAdvisor, and Lanser and his team have shown enough interest in our sandpit experiments to be the perfect test case for what ‘could be’?

Last week we finished training ChatGPT on everything Lanser’s on Main. It wasn’t as easy as handing over his website URL for ChatGPT to wander through and gather information (although that was part of it). The heavy lifting was building a document for it to read and learn from, on every angle we could come up with. In the end, it consisted of hundreds of possible questions and answers people might ask (and we’re not finished yet). Having built ‘automated bots’ for websites for the past few years, you’ll not imagine the questions people will ask on a given subject, and more importantly, how they’ll ask them. In the old days you had to get pretty close to predicting. Today, there’s a lot more ‘play’ allowed for a great answer to be delivered.

Ozee is Given Life

Today we’re turning the Bot for Lanser’s on Main on, and letting it play with his guests and future guests via the website (bottom left corner). It’s name is Ozee, and it’ll greet you with:

Hi my name is Ozee. I think I’m a really intelligent Bot. Ask me something about Lanser’s on Main

Ozee is only trained on Lanser’s on Main, and in theory Ozee isn’t supposed to tell you who the President of South Africa is, or even take a guess at the weekends Lottery numbers. However, if you want to know the opening hours of Lanser’s on Main, whether there are vegetarian options, or what’s in the Steak Tartare, Ozee is going to shine (in theory).

We’ve only allowed Ozee to answer 4  of your questions at a time (within a particular time period). We’re doing this to watch Ozee for now, and also because each answer has a cost, and we don’t want a lonely human, with 2 bottles of wine in them, playing with Ozee for hours on end. Well, not just yet.

A Hybrid Model Makes Sense

The reason we embarked on this project is because a hybrid model makes sense to us. Websites are still useful for many things, but if a Bot can intelligently answer questions about some of the information you’re not sure where to find, why wouldn’t you ask it? It’ll save you tons of time if Ozee can do the heavy lifting and answer your questions about Lanser’s on Main, while you work out what date you’d like to make a booking.

The future is always arriving. Let’s see how Ozee does?

The post We Built a Chatbot for a Client, and it’s Pretty Cool appeared first on Calidascope.