SCCMPod-577 CCE: Exploring AI Chatbots in the PICU

So first of all, as I said, this is not the area that I have dived into like you have.

So can you explain large language models for those who may not understand what that actually is?

Yeah.

And I mean, I think the way you framed it is perfect.

And I will say, despite being very bullish about the technology, I also personally think this is an interesting time for healthcare where we often, as a field, have been slow to adopt a new technology.

But for whatever reason, with generative AI, I mean, it has just kind of entered into our daily practice, whether it's through ambient scribes or open evidence, for example, claims more than a million visits per day of clinicians asking questions to these tools.

And to your point, there are real risks with them.

And we're at a funny point where we don't have a lot of data, but we do have a lot of use.

So understanding the fundamentals of what is generative AI or a large language model, it really does go a long way.

So I'll try to be succinct, but A large language model at its base is, to me, it's the tool algorithm or technology that's driving some of the chatbots that we interact with, like ChatGPT.

And now, as you said, there's a number of other software programs that are sort of software wrapped around these large language models.

And from my perspective, the easiest way to think about it is the large language model itself is basically a big algorithm that read a huge portion of the internet.

So there's these big databases called Common Crawl or whatever, but essentially there's just a bunch of text from internet, books, articles, whatever.

the models sort of ingest them and understand how humans tend to communicate, how words tend to flow together.

And so baked into the DNA of the models is when you ask a question or you put in a request, the way it's generating the answer is, you know, you could say probabilistic, where really one word at a time, it's basically predicting, and they're called tokens, we can just think of them as words, but it's predicting the most useful next string of text based on everything that it's been exposed to plus whatever you've given it.

And so the models today are much more sophisticated.

So you'll see them, you know, Googling things or for lack of a better term, but using internet websites to get information, getting information from specific databases or whatever it is.

But it's important to note that when the actual output is coming back to you, it's still generated in this sort of predictive way.

And so, you know, what you hit on is one of the biggest things that people talk about, quote unquote, hallucination.

The models are so good at sounding fluent and providing convincing responses.

But they're sometimes not as good at being calibrated or expressing a lack of confidence in whatever reply they're giving.

And so to me, you said it exactly right, that verifying output is incredibly important in their current form.