The other day I went for a bike ride along the Monsal Trail with my old pal Dr Baz Broxtowe of Fradley University. Dr Baz heads a team developing a conversational AI system called Frad. It’s part of his research into machine learning applied to social nuances.
There have been problems with Frad though, as I found out after we pulled up for a drink of water and a short break.
You could easily do another couple of miles tubby.
The voice came from Dr Baz’s phone almost as soon was we climbed off our bikes by one of the trail picnic tables.
“What was that? Was it Frad?” I asked the obvious question.
“Yes it was and it’s one of our problems,“ Dr Baz replied as he unwrapped a bar of chocolate. We can’t get the tone of Frad’s responses right. We’re looking to develop a rounded personality but can’t find the ideal mean between dull know-it-all and acerbic.”
“And at the moment it’s tuned to acerbic?”
“That’s right, I have to say that Frad is pretty good at telling it as it is when tuned towards acerbic, but obviously we can’t go public with it however good the results. There would be uproar.”
“You mean it could be worse that calling you tubby?”
He is tubby.
“Shut up Frad. We’ve had a few other problems with it,” Dr Baz went on to explain. For example an overweight staff member asked Frad about diet and losing weight. It suggest she should buy three pizzas, stack one on top of the other and scoff the lot as a pizza sandwich. Then it said –
You know you want to, so be true to yourself.
“Oh - I bet that didn’t go down well.”
“It did not go well at all, we had to make modifications. The trouble is, we seem to have a fundamental problem with conversational AI. There is not much conversational space between dull know-it-all on the one hand and acerbic but interesting on the other. ”
“It’s a real headache,” added Dr Baz as we climbed back onto our bikes.
I assume we are not going back already. said his phone.
You could easily do another couple of miles tubby.
The voice came from Dr Baz’s phone almost as soon was we climbed off our bikes by one of the trail picnic tables.
“What was that? Was it Frad?” I asked the obvious question.
“Yes it was and it’s one of our problems,“ Dr Baz replied as he unwrapped a bar of chocolate. We can’t get the tone of Frad’s responses right. We’re looking to develop a rounded personality but can’t find the ideal mean between dull know-it-all and acerbic.”
“And at the moment it’s tuned to acerbic?”
“That’s right, I have to say that Frad is pretty good at telling it as it is when tuned towards acerbic, but obviously we can’t go public with it however good the results. There would be uproar.”
“You mean it could be worse that calling you tubby?”
He is tubby.
“Shut up Frad. We’ve had a few other problems with it,” Dr Baz went on to explain. For example an overweight staff member asked Frad about diet and losing weight. It suggest she should buy three pizzas, stack one on top of the other and scoff the lot as a pizza sandwich. Then it said –
You know you want to, so be true to yourself.
“Oh - I bet that didn’t go down well.”
“It did not go well at all, we had to make modifications. The trouble is, we seem to have a fundamental problem with conversational AI. There is not much conversational space between dull know-it-all on the one hand and acerbic but interesting on the other. ”
“It’s a real headache,” added Dr Baz as we climbed back onto our bikes.
I assume we are not going back already. said his phone.
3 comments:
And (serious point) since there is a great deal of left-inspired 'material' on the interwebs and a dearth of right-inspired 'material' the training for Large Language Model AI is inevitably going to prioritise a left-influenced output.
It would be interesting if an AI was trained on right-inspired 'material' and the two AIs output was compared. I suspect there are few businesses around willing to risk the opprobrium of creating a 'fascist' AI - even though you could argue that most AIs already achieve this.
Arguably "find[ing] the ideal mean between dull know-it-all and acerbic” very much depends on where you place the balance point and what your predispositions are.
"The trouble is, we seem to have a fundamental problem with conversational AI. There is not much conversational space between dull know-it-all on the one hand and acerbic but interesting on the other."
Convey my congratulations to Baz - he's cracked it! AI that is indistinguishable from human intelligence.
DJ - so many people are so good at taking things the wrong way that there must be all kinds of pitfalls with conversational AI unless it is a dull know-it-all. How it develops should be interesting, but apart from specialist applications, dull know-it-all may be all we should expect for a while.
Sam - I'll tell him. He may need a simple toggle switch to go from one to the other which could make it even less distinguishable from human intelligence
Post a Comment