There are pretty great applications in medicine. AI is an umbrella term that includes working with LLMs, image processing, pattern recognition and other stuff. There are fields where AI is a blessing. The problem is, as JohnSmith mentioned, it’s the “solar battery” of the current day. At one point they had to make and/or advertise everything with solar batteries, even stuff that was better off with… batteries. Or the good ol’ plug. Hopefully, it will settle down in a few year’s time and they will focus on areas where it is more successful. They just need to find out which areas those are.
Like what? I discussed just 2 days ago with a friend who works in public healthcare, who is bullish about AI and best he could come up with DeepMind AlphaFold which is yes interesting, even important, and yet in a way “good old fashion AI” as has been the case for the last half century or so, namely a team of dedicated researchers, actual humans, focusing on a hard problem, throwing state of the art algorithms at a problem and some compute resources… but AFAICT there is so significant medicine research that made a significant change through “modern” AI like LLMs.
The first thing that comes to my mind is cancer screening. I had to look it up because I can’t always trust my memory, and I thought there was some AI involved in the RNA sequencing research for the Covid vaccine, but I actually remembered wrong.
Skimmed through the article and I found it surprisingly difficult to pinpoint what “AI” solution they actually covered, despite going as far as opening the supplementary data of the research they mentioned. Maybe I’m missing something obvious so please do share.
AFAICT they are talking about using computer vision techniques to highlight potential problems in addition to bringing the non annotated image.
This… is great! But I’d argue this is NOT what “AI” at the moment is hyped about. What I mean is that computer vision and statistics have been used, in medicine and elsewhere, with great success and I don’t see why it wouldn’t be applied. Rather I would argue the hype at he moment in AI is about LLM and generative AI. AFAICT (but again had a hard time parsing through this paper to get anything actually specific) none of that is using it.
FWIW I did specific in my post tht my criticism was about “modern” AI, not AI as a field in general.
I’m not at that exact company, but a very similar one.
It’s AI because we essentially we just take early scans from people who are later diagnosed with respiratory illnesses and using that to train a neural network to recognise early signs that a human doctor wouldn’t notice.
The actual algorithm we started with and built upon is basically identical to one of the algoriths used in a generative AI models (the one that takes an image, does some maths wizardry on it and tells you how close the image is to the selected prompt). Of course we heavily modify it for our needs so it’s pretty different in the end product, and we’re not using its output to feedback into a denoiser and we have a lot of cognitive layers and some other tricks to bring the reliability up to a point we can actually use it denoise, but it’s still at its core the same algorithm.
There are pretty great applications in medicine. AI is an umbrella term that includes working with LLMs, image processing, pattern recognition and other stuff. There are fields where AI is a blessing. The problem is, as JohnSmith mentioned, it’s the “solar battery” of the current day. At one point they had to make and/or advertise everything with solar batteries, even stuff that was better off with… batteries. Or the good ol’ plug. Hopefully, it will settle down in a few year’s time and they will focus on areas where it is more successful. They just need to find out which areas those are.
Like what? I discussed just 2 days ago with a friend who works in public healthcare, who is bullish about AI and best he could come up with DeepMind AlphaFold which is yes interesting, even important, and yet in a way “good old fashion AI” as has been the case for the last half century or so, namely a team of dedicated researchers, actual humans, focusing on a hard problem, throwing state of the art algorithms at a problem and some compute resources… but AFAICT there is so significant medicine research that made a significant change through “modern” AI like LLMs.
The first thing that comes to my mind is cancer screening. I had to look it up because I can’t always trust my memory, and I thought there was some AI involved in the RNA sequencing research for the Covid vaccine, but I actually remembered wrong.
Skimmed through the article and I found it surprisingly difficult to pinpoint what “AI” solution they actually covered, despite going as far as opening the supplementary data of the research they mentioned. Maybe I’m missing something obvious so please do share.
AFAICT they are talking about using computer vision techniques to highlight potential problems in addition to bringing the non annotated image.
This… is great! But I’d argue this is NOT what “AI” at the moment is hyped about. What I mean is that computer vision and statistics have been used, in medicine and elsewhere, with great success and I don’t see why it wouldn’t be applied. Rather I would argue the hype at he moment in AI is about LLM and generative AI. AFAICT (but again had a hard time parsing through this paper to get anything actually specific) none of that is using it.
FWIW I did specific in my post tht my criticism was about “modern” AI, not AI as a field in general.
I’m not at that exact company, but a very similar one.
It’s AI because we essentially we just take early scans from people who are later diagnosed with respiratory illnesses and using that to train a neural network to recognise early signs that a human doctor wouldn’t notice.
The actual algorithm we started with and built upon is basically identical to one of the algoriths used in a generative AI models (the one that takes an image, does some maths wizardry on it and tells you how close the image is to the selected prompt). Of course we heavily modify it for our needs so it’s pretty different in the end product, and we’re not using its output to feedback into a denoiser and we have a lot of cognitive layers and some other tricks to bring the reliability up to a point we can actually use it denoise, but it’s still at its core the same algorithm.
Thanks, any publication please to better understand how it work?