Conversation-based dementia therapy – a problem only generative AI can solve

Three years ago, not even one of the largest companies in the world could solve this problem. Now a small startup can. The reason is generative AI.

There are currently no drugs that cure dementia, but there are scientifically proven conversation-based treatments that can slow the progression of the disease and even improve cognitive abilities.

The problem is that these conversations need to take place with a trained conversation partner, preferably several times a week. In the U.S., eleven percent of everyone over the age of 65 has dementia, and a total of twenty-two percent have mild cognitive impairment. In other words, we’re talking about many millions of people. That makes it practically impossible to help them, as it would be astronomically expensive to train all the necessary conversation partners and pay for all those sessions.

But what if we could train an AI to be that conversation partner – one that can carry on a natural human conversation while also being trained in accordance with the scientific conversational methods?

Meet Sunny.

Sunny is exactly such an AI, built by a startup called NewDays AI.

Sunny’s creator is Babak Parviz. Together with Daniel Kelly, he co-founded NewDays. When I was recently in Seattle, I took the opportunity to have coffee with Babak.

My relationship with Parviz, in fact, has been marked by a series of coffee meetings, or fika, as we call them in Sweden. The first time we met was a few years ago at a café in Seattle. Back then, he was Vice President at Amazon, leading their Grand Challenges – Amazon’s equivalent of Google X. A sort of freewheeling innovation lab pushing the frontiers of technology. Babak had come there straight from Google X, where he led the development of Google Glass and Google Lenses. 

A while after our first coffee, we started Warp News, and I asked him if he might consider investing in us. He did, and since then, he’s been both a sounding board and someone who has helped me understand new technologies – especially around AI.

This time, we met at a French bakery in University Village. He orders a sandwich with some kind of jam on it – I think it’s his lunch. I’m still stuffed after breakfast at Denny’s, so I just go for a cappuccino.

He tells me how they built Sunny. It was definitely not a matter of just dropping a clever prompt into ChatGPT. Sure, AI tools like ChatGPT are now really good at human-like conversations, both in text and speech, but they still have limitations. When conversations go beyond ten minutes, the AI itself starts to become a bit “demented” – it can forget what was said at the beginning. It also needs to remember things between conversations. If you said a month ago that your brother lives in Kentucky, but now you say he’s dead – is that a memory failure or did the brother actually die?

The solution is a combination of multiple language models and agents working together. It hasn’t been easy, he says.

My guess is that the agents extract information from the ongoing conversation, transfer it to other language models where it’s processed and stored in a way that allows checking and recalling during future conversations. That way, Sunny can keep track and also evaluate how the discussions and exercises are progressing.

Babak and I are having coffee in Norrtälje.

What’s so fascinating is that this simply wouldn’t have been possible just three years ago, not even at a giant company. Parviz testifies to this. He tells how Amazon tried to build various health-related solutions and manage huge amounts of text and data, like medical records. It required massive resources, and many times the solutions did not work well.

The solution is called generative AI. Before the Transformer breakthrough at Google in 2017 – and before OpenAI used it to create ChatGPT – AI was simply too poor at understanding text.

Now, a small startup with dramatically fewer resources can build services that no company in the world could have built three years ago. To make this kind of system work properly still requires high competence, but it’s doable.

The kinds of solutions NewDays AI is developing – solving the problem of short-term memory in AI – are what will create the next wave of innovation. That same kind of solution would, for example, be perfect for an AI acting as a private tutor for a student. There, too, a long memory and the same kind of analytical capability are needed.

Continue reading the article on Warp News.