I'm talking about things like my CEO wanting me to build an AI sales assistant when a simple web form would actually be faster to develop and use.
What are you seeing?
When I inspected the code for the site I saw it was just hard coded in to the web page and figured someone demanded AI should be used on the app so the dev just faked it.
Where I'm familiar with AI in healthcare is tools to help radiologists and cardiologists detect cancer early. This is a support AI, not the AI doing all of the work.
I'm also hearing about the potential for generative AI to provide a good bedside manner in "discussing" a diagnosis with a patient. Not diagnosing, but having more patience and time than a doctor has to communicate with the patient on known entities.
We use ML for measuring sleep state and finding opportunities to stimulate the brain to improve the efficiency of deep sleep.
Where are these "worst possible ways" that I'm unfamiliar with?
Some doctor patient visits may fall under multiple codes, so billers will optimize for the highest paying code.
One instance I observed all the ED providers huddled around a screen watching a yt video for a pediatric patient, we are not a children’s hospital. They all watched then discussed the video before going back out to the pt.
I’m sure there should be better documentation available but so much of their jobs is judgement based already.
Medical specific trained LLMs are likely to be commonplace in near future for providers based on feedback I’ve gathered.
If you push 3-4 documents into a language model and then ask it questions, it creates some dangerous situations
So many people seem to want to be able to want to contort themselves with prompt "optimizations" to try to get models to do exactly what they want. They think if they can just throw a good-enough prompt into this black box it'll miraculously write entire applications for them or do whatever else they want.
I believe this is a misguided overcorrection. Instead, I think we should strive to use AI to _extend_ (and eventually even build) our non-AI tooling rather than replace it. GPT-4 is not a reliable magic box that will solve all your problems - but it is a powerful way to enhance your existing brain power _and_ other tools. We can likewise build tooling that will help us offload more work to AI (e.g., develop systems that analyze AI output and reflect on themselves). We can build systems that clarify and break instructions down into smaller steps (some of which may involve AI queries and others not). But to spend hours trying to tweak the perfect prompt to get just the right kind of usable output from GPT-4, just so you don't have to open a spreadsheet or write your own code? We're wasting time insetad of saving it.
Thank You!
That and how pointlessly rambly they are. Maybe good if I'm trying to fall asleep, but totally grinds my gears when I'm looking for information! (Pretty sure they do it that way to increase the number of ads shown to users over the course of the material)
Are we living in a badly written sci fi book?
The latest was for a chatbot that 100% needs to give exact information or it could be be a big problem form the company. The idea was shot down, but a traditional chatbot would have been a better idea.
Given the LLM hype and over a year in, I was expecting there would be all these "why didn't I think of that" uses by now.
I can't think of a single example. A chat bot no one is going to use seems to be the main use case.
In other words, look at methods that are in Murphy's volume 2, rather than 1, especially parts II, V, and VI: https://probml.github.io/pml-book
Don’t worry when this one fades a new one will come right along…
If AI is being used in places it shouldn't be, then presumably after a little (or lot) of money has been wasted, firms will figure out that it's a bad fit for the job. I don't see any good reason to rule anything out before there's been time to evaluate it's efficacy.